America | Tech | Opinion | Culture | Charts

Two things. First, we’re excited to welcome Peter Levine back as a full-time GP. Read more about his return here.

Second, we get asked a lot: who does your charts? Today I’m really happy to get to unveil his identity: Moses Sternstein, known to many of you as the author of Random Walk. Before we get to charts today (they’re there, if you scroll down), here’s Moses introducing himself and sharing some insights from this year’s Sohn conference. -AD

First, a brief introduction, before the main event(s). Some of you may know me from my other publication, Random Walk, but if not, welcome to an a16z production of Random Walk. As Robert Hagstrom famously wrote (reflecting on the great Charlie Munger), “investing is that last liberal art.” It selects for curiosity, discovery, and pattern-matching–a “multi-disciplinary mindset” and a “latticework of mental models”--and offers the rare premium for being different, early and right.

That Mungerism has always been the message, and in Random Walk’s case, charts are most definitely the medium. By joining up with Erik, Alex, and the extremely talented a16z New Media team (and a16z more broadly), we’re going to shout it from the rooftops like never before–weekly charts for now, but much more to come.

Tales from Sohn

Now, onto the show, beginning with (ironically enough) a chart-free prelude: some brief reflections of my third consecutive tour of the Sohn Investment Conference.

Sohn is the rare conference where the presentations are the best part (and it raises a ton of money to fight cancer, and since cancer is very bad, that makes Sohn very good). Now in its 31st year, Sohn features professional investors (both established and up-and-coming) pitching their investment ideas, and otherwise musing on the general state of markets.

The conference is idea-dense–some presenters like to pack multiple pitches into a single 5 minute presi–but mostly it’s solid affirmation that very smart people have very different views of where the opportunities lie (although, sometimes, more so than others).

It’s also highly performative–investors share the ideas that reflect the views they want everyone else to know that they have. It’s the investor equivalent of being seen at the opera, and your outfit definitely matters. And because Sohn happens every year around the same time, it’s fun to look back on Sohns-past like rings on a tree, each one telling a story of the investor-zeitgeist at the time–and, of course, mapping that zeitgeist to what eventually transpired (because hindsight is nothing if not 20/20).

Basically everyone missed the AI boat last year

The first big observation about Sohn 2026 is really an observation about last year’s Sohn Conference.

As far as I’m aware, no one at Sohn ‘25 made any calls about the cycle’s biggest winner: memory (or even semis). And no one made any calls about the cycle’s biggest losers, either: software (and, relatedly, private credit, although there was one exception). Now, to be fair, Sohn 2024 (i.e. two years ago) had pitches for both TSMC and ASML, a conversation between Daniel Gross and Eric Steinberger (about agents), and plenty of bullish commentary on GPU demand, but Sohn ‘25 was mostly crickets on the AI trade.

In fact, at Sohn ‘25 there were at least two prominent software longs: Adobe and Ncino (neither of which was spared in the saaspocalypse). There were a few payments and marketplace plays, all of which had a bad year. And even while there was a decent amount of interest in industrial and infrastructure, almost none of it sat squarely in the AI infra buildout–other than pitches for a Nat Gas supplier, radiation detection, and a wiring installer, there was more interest in plane parts and Japanese cosmetics than AI, let alone silicon.

Subscribe now

That AI might break SaaS didn’t come up, at all, last year (again, that I can recall). And the only Private Credit short was limited to a specific BDC where valuations looked unreasonably stable (supposedly due to unique control positions the BDC holds, giving it a lot of sway over marks, but having nothing to do with software). “Software lending” wasn’t an area of concern whatsoever (and the BDC in question actually has/had relatively low software exposure).

In general, AI and its second-order effects (which have otherwise defined the market for the past year+), were very absent from Sohn 2025. That isn’t to say that Sohn ‘25 presenters ignored AI completely—or that there were no hits among the non-AI pitches—but they do appear to have completely missed the main AI boats (for better and for worse). And since those were the main everything boats this past year, then by the transitive property . . . well, you do the math.

What is one to make of that?

I suppose the most straightforward takeaway is that investing is hard, the future is unknowable, and the AI landscape (especially) is changing very rapidly. AI coding took a real leap forward in 2025, with the biggest leaps in the Fall (well after Sohn ‘25). It’s also possible that there was some “AI fatigue” at the time, and presenters wanted to focus on less-obvious ideas. Alternatively, perhaps yet other presenters wanted to keep their best AI ideas closer to the vest. Or maybe one ought to conclude that professional investors are bad at their jobs, or at least that the professional investors who appear on podiums are bad at their jobs?

Or perhaps some combination of all or none of the above. Be humble, make bold calls, live to fight another day.

No AI Bears at Sohn (but only one AI winner)

This year at Sohn, however, AI was very much front-and-center (and yet, somehow, not).

Lots of presenters offered some angle on how the surge in demand for power and silicon is going to transform previously sleepy (or otherwise down on their luck) companies and industries.

• There were some longs in smaller semi and powertrain players, in say autos. One pitch focused on a niche market for nuclear reactors previously limited to US Navy submarines that had the potential to ride the SMR tailwind.

• Can a crypto business provide on-prem power to data centers? At least one presenter said “yes it can.”

• Remember Nokia? They don’t even sell phones anymore—it’s a fiber optics business now (and data centers need those).

• Engineering, Procurement, and Construction (EPC) services is suddenly hot-hot-hot, and the hard part is finding an EPC business that isn’t already ripping (but will be soon)

• Memory and switchboards used to be as generic and commoditized as they come, and now, all of a sudden, having specialized expertise is generating historic pricing power for the most mundane parts of the chip stack.

• One presenter mused about how AI Capex was likely to transform Korea’s entire economy, given how much money is flowing to Korea’s memory-makers (and eventually their Korean employees).

On the short side, the question wasn’t whether AI would threaten software, so much as whether there were still unidentified victims, and/or perhaps whether some good companies (but really, their loans) had been thrown out with the bathwater.

But in all events, the consistent refrain of Sohn ‘26 is that AI is here, it’s growing, and it’s transformative. AI wasn’t quite the everything pitch, but it was close.

Certainly some of the discussion touched on whether all this demand for AI infra was sustainable, and at what point high prices would drive breakthroughs on efficiency. As Jonathan Ross (ex-Groq, now Nvidia) pointed out, up until recently, no one was really trying to squeeze more performance out of memory chips, but they sure are now. Others noted that Google is working hard on its TPUs and Amazon its Trainiums, both in an effort to be less Nvidia-dependent (even while reaffirming that Nvidia is still in a class of its own for what it does). Likewise, while it’s true that we’re likely to see innovation in previously commoditized products, it turns out that the same firms producing the commodities are the ones with all the technical know-how, operational scale, and client mandates to do the innovation (and keep up with the rapidly accelerating cycles).

Whatever doubts one might have, [Alex Sacerdote of Whale Rock] pointed out that the number of AI power-users is currently a small fraction of one percent of AI users—and yet, we’re already in a compute shortage. If “power-user” is the future state of what using AI looks like, then consider how early the compute buildout actually is (and how much farther it still has to go). In the meantime, investors are hiring AI “ninjas,” and the entire “old economy” is in catch-up mode.

You think running a compute-cluster is a commodity business? Sure, so is grocery, and building a world class grocery business is as easy as running 1,000 different grocery stores, with efficient pricing, inventory management, quality control, staffing, and site-selection. It’s so easy, you can count the number of successes on your fingers (more or less).

Bigger picture, everyone agreed that AI is a very big deal, and the underlying premise was very bullish AI—not whether, but how fast and how far (and how to grab a seat on the train). Sure, there will be some hills and valleys along the way, but when are there not? If there’s market froth, it’s in companies that try to ride the train by adding “AI” to their names–one short-seller used name changes as a screen–or rebranding a legacy software roll-up (loaded with related-party transactions) as an “AI powered” roll up. But no one came close to using the “b word,” (except to say, “this isn’t a bubble.”)

You’d have to dig deep to find a truly bearish comment, of any kind. Gavin Baker (hardly an AI bear) came closest by offering up the call that we’d go cold turkey on new terrestrial data centers, once orbital compute changed the game.

Could one detect a little apprehension? Sure. But, bearishness? Not even close.

So many AI winners, then why no AI winners?

All that AI bullishness, though, invites a different sort of question: with all those potential winners in the AI semi supply- and power-trains, why the paucity of good old fashioned businesses set to win with AI?

There were 1 or 2 exceptions, of course–a hardware company adding an AI software license built on decades of proprietary public safety data; a gaming company building an AI-driven gaming engine, with end-to-end commercial support for game devs. On the private side, Orlando Bravo (of the PE software giant, Thoma Bravo) had a whole session to himself where he talked about how AI was challenging software companies, yes, but also creating opportunities for software companies (especially cyber), where domain expertise provides an edge.

But, for the most part, as one panelist reflected, “it’s easy to spot the AI losers—it’s a lot harder to spot the winners.” (And, perhaps the winners will be diffuse and sort of boring, i.e. 200bps of margin improvement for a broad set of businesses, which would be a massive change, but lacking in star power).

What is one to make of that? Again, who really knows.

Maybe the Sohn ‘26 presenters will be just as “wrong” as the Sohn ‘25 presenters were (for better or, possibly, for worse). Maybe Sohn ‘27 will be the rip-roaring return of software companies or AI-driven biotech, or maybe Sohn ‘27 will be the year the value guys finally get their incremental margin-improvement and big wins for the slow n’ steady, or maybe, heaven forfend, at Sohn ‘27, the shorts will preside. Or maybe AI will undergo yet another step-function change, and the talk of Sohn ‘27 will be some other bottleneck that no one is talking about now, but will seem obvious in retrospect.

Truthfully, if anyone really knew the answer, they’d probably keep it to themselves, for now.

In all events, Sohn is a great conference (for a great cause), and one of the few conferences where actually listening to the presentations is the best part. It’s also something of an investor’s time capsule, or a freeze-frame on the capital markets zeitgeist, that’s informative, for sure, but mostly just humbling. Can’t wait for the next one.

Peak computer science

There’s a big change afoot in terms of what students are choosing to study.

According to data from Revelio, computer science (and IT) enrollment has become much less popular, of late:

Computer science enrollment is regressing close to 2018 levels, and as a share of students, it’s dropped by ~1/3.

Engineering (and finance), by contrast, are on the rise.

Why are students shifting their enrollment preferences? One strong possibility is that they’re responding to market signals (albeit on a bit of a lag):

More enrollment appears to be roughly correlated with changes in the “salary premium,” i.e. the compensation premium/discount for a given major, relative to their school’s median.

The compensation premium for engineering roles has increased (and so has enrollment), while the premium for computer science–while still relatively high–has compressed (and enrollment has followed). The one major exception appears to be biology and biomedical engineering, where the premium has declined, but the enrollment share has nonetheless increased.

Maybe biomed knows something the market doesn’t, or maybe they just do it for the love of the game.

Small Business Boom

As we’ve pointed out various times before, US companies are becoming more profitable than ever. Margins have been expanding for some time (especially, but not only, in tech), and the recent run on AI-related parts and services, is pushing expectations for profit-growth upwards in unprecedented ways.

In the bigger scheme of things, though, profitability has been a trend for a while, and leadership begins at the top—you might even say that leadership begins at the tippity-top:

Since 1960, the share of sales attributable to the top 1% of firms has grown from ~60% all the way to ~80% (and the top 20 firms are responsible for about half of all sales by themselves).

The biggest, most profitable companies have become even more central to the overall story. These are super-pareto distributions, if that were a thing.

At the same time, there’s a possibility that AI begins to change the story, at least somewhat. We’ve noted before that new business formation appears to have accelerated recently, and while these new businesses aren’t adding much headcount, they are spending a lot on software. In other words, there’s some evidence that AI is unlocking a rise in “solopreneurship.”

Recently, Ernie Tedeschi of Stripe Economics, made a similar observation, based on Stripe’s own data:

All categories of founder appear to have grown in Q1, but “non-AI, solo-founders” grew the most (and “AI, solo-founders” grew the second-most).

In other words, business formation is on the rise, but for the most part, it’s not the “high propensity,” hiring kind. In this case, “non-AI” simply means that AI is absent from the product description, so it’s possible (if not likely) that many of those non-AI businesses are using AI (at least in part) to make things happen. In Tedeschi’s words:

The rapid development of AI tools might be lowering both the barriers to starting a business and the returns to doing so for individual operators—contributing to a rise in AI solopreneurs: individuals using AI to establish sole proprietorships or side businesses without the need or intent to hire.

Obviously it’s still early, but as ARR/FTE has become the new en vogue metric to track operating leverage, it’s hard to imagine a better setup for ARR/FTE than operating a firm of one (assuming the solopreneur can hack the ARR part, too). We’re a big fan of outliers, but perhaps the AI-small business renaissance will begin to recapture some part of the profit-pool from the bigcos. Or maybe the pie will just keep getting larger, and that’s pretty good too.

Token maxxing

At Google’s recent I/O 2026, Google’s CEO shared some data on the explosion of token demand:

Google is now processing more than 3.2 quadrillion tokens per month–a 7x increase from a year ago.

The point being that token demand is growing a lot. More firms are using more AI, and the change is happening rapidly, and at massive scale.

It’s not just Google’s data that shows a surge in token consumption. According to Silicon Data (via Bloomberg), total token spending (i.e. dollars spent on tokens), has risen to new highs, and growth recently took the shape of a near-vertical wall.

Total token expenditure is now more than double what it was in just December of 2025.

So, more tokens consumed, and more dollars spent on tokens. That’s interesting so far as it goes, but the really interesting thing is the ongoing apparent relationship between token-pricing and demand: the cheaper that tokens get, the more demand appears to accelerate. In other words, this is a Jevons-like acceleration, whereby efficiency-gains expand the frontier of demand.

The “price of intelligence” continues to decline.

The CEO of Silicon Data made a similar point a few weeks back, even before the recent surge:

“The mechanism is textbook Jevons:

• Cost-per-token fell ~40% YTD (Epoch AI)

• Query volume and context usage scaled faster

• Net spend compounded

A divergence worth sitting with:

• Token expenditure (SDLLMTK): 1.27 → 1.76, +38% YTD

• B200 non-hyperscaler rental (SDB200RT): $4.30 → $5.16, +20% YTD

Token spend is compounding ~2x faster than GPU rental rates.”

The “textbook Jevons” relationship appears in other data, as well. According to Yipit Data, weekly token-pricing has been steadily declining (for the most part), while token consumption increased substantially:

This is almost Jevons+, insofar as simply holding token-costs mostly steady, has coincided with a surge in demand–possibly because the tokens themselves have more intelligence bang-for-the-buck.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Last Saturday morning, three teenagers stole a handgun from Central Texas Gunworks in Austin. In the next twenty-eight hours, they shot four people and stole five vehicles. The oldest teenager was seventeen.

Sunday afternoon, surveillance footage showed a white Kia sedan approach a victim, then drive off as he fell to the ground. Austin Police sent a BOLO, short for “be on the lookout,” to surrounding agencies, and in another town twelve miles east of Austin, the Manor Police Department located the car thirty-eight minutes later. A Flock Safety camera had picked up the plate as the car drove past, matched it against the BOLO, and flagged it. Two suspects were caught within the hour, and a search with nearly two hundred officers, helicopters, and K-9 units continued into the night before the third was found.

Austin had used Flock cameras since 2023, but ended the contract in June of 2025 over privacy concerns raised by local advocacy groups. At a press conference on Sunday announcing the arrests of the three teenagers, Austin Police Chief Lisa Davis was asked whether license plate reader technology like Flock cameras could have helped.

“Yes,” Davis said. “It could have helped.”

The $1,000 Prototype

Garrett Langley was an electrical engineer who’d already sold two companies: a car subscription service called Clutch and Experience, a live events platform acquired by Cox for $200 million. Neither company, in his uncle’s blunt assessment, really mattered. “Go build something that’s not so dopey,” his uncle told him.

He was living in Atlanta when a group of kids jumped out of a car one night in 2017 and started pulling door handles on parked trucks. This is a common form of organized crime in the southeast. Pull ten F-150 door handles and about three will be unlocked. One in ten of those trucks will have a firearm in the glove box. That night, one of those kids found a firearm, and the truck’s owner posted on Nextdoor: “Oh my gosh, I forgot my gun in my car, and it’s now gone.” The Atlanta Police Department responded, but there wasn’t much they could do.

Langley didn’t know much about law enforcement, so he started asking Atlanta-area police chiefs and HOA boards why more of these crimes weren’t being solved. The bottleneck was evidence, the police chiefs told him, and specifically license plates. Maybe this was his chance to build something that mattered.

He had stumbled onto a national problem. In the United States, more than half of murders go unsolved. For robberies, the figure is nearly 70%. For motor vehicle theft, 86%. These numbers actually overstate how well the system works, because the majority of crimes are never formally reported to the police in the first place. According to the Bureau of Justice Statistics, only 26.8% of household thefts even generate a police report.

Roughly 70% of US crimes involve a vehicle, so a clear license plate or vehicle signature is the fastest way to create an actionable lead. Automated License Plate Recognition (ALPR) systems had been around for years, but a single deployment ran about $50,000 due to construction costs (trenching for power and fiber, permits) and specialized hardware (infrared cameras). Rich cities could afford a handful at strategic chokepoints, but these systems were out of reach for everyone else. Safety, in other words, was a privilege.

Langley wondered if modern computer vision, running on cheap mobile processors, could do the same job. And if you powered the system with a solar panel, you didn’t need to trench for power or fiber at all.

Matt Feury was an engineer who had worked at Langley’s previous company, Experience. Together they spent a weekend building a prototype from an Android phone, a solar panel, and a battery they purchased for a combined price tag of less than $1,000.

The smartphone supply chain had made high-end image sensors cheap. The camera in a one- or two-year-old phone outperformed most standalone sensors and cost around $50. Computer vision had also reached a threshold where a competent engineering team could build a production-grade license plate recognition model without machine learning PhDs. And the whole thing could run on solar power with LTE connectivity, which meant no trenching, fiber, or electrical hookup. With some creativity, a box of parts from Best Buy could now beat the $50,000 incumbent technology.

The prototype looked, by Langley’s own admission, “like garbage,” but they put it up in their neighborhood anyway. Two months later, there was another break-in, but this time, they had an actionable lead. They could see every car that passed through the neighborhood that night. A few hours later, the police made an arrest.

The five o’clock news wanted the story, and the next morning, Langley had five emails from other neighborhoods. He, Feury, and their third co-founder Paige Todd incorporated Flock Safety later that year. Nine years later, the company operates across 49 states and is valued at more than $8 billion.

The Growth Engine

Flock’s first customers were HOAs, which turned out to be the perfect beachhead market. Boards met every month, controlled discretionary dues, and could approve a $2,000-3,000 annual camera subscription without a lengthy procurement process.

Josh Thomas, Flock’s VP of Marketing at the time, soon figured out a more scalable way to acquire these customers. Local news stations were starving for content, and a story about new crime-fighting technology with a privacy angle was editorial gold. Every time Flock sold a camera, they would pitch the story to local media. “Flock Safety Installs New Technology to Help Solve Crime, But What’s the Risk to Privacy?” was the headline Josh wrote in his pitch. The local affiliate would send out a crew to interview local community members and show the product. The story would air, and website traffic would spike.

“That’s basically how we grew for the first few years,” Josh told us. “Only PR and word-of-mouth.”

Law enforcement was mostly an afterthought for the first two years. Agencies would make arrests with evidence from neighborhood cameras, then call Flock asking to buy directly. The demand was there, but when Flock tried to turn it into a repeatable sales motion, the standard SaaS playbook of sales development representatives, account executives, and cold outbound produced almost nothing. For twelve months, the pipeline with law enforcement was essentially zero. The breakthrough came when Langley started flying to Los Angeles every week and doing door-to-door, relationship-based sales. When he could reliably fly to a city and come back with a signed deal, he brought on a Chief Revenue Officer who scaled the playbook across a hundred reps.

Flock’s vehicle search was already more useful for investigations than the legacy systems. The old ALPR technology relied on infrared reflection off a plate’s reflective coating, but Flock’s computer vision captured much more: make, model, color, the presence of roof racks, and bumper stickers. The company calls the twenty-four distinct characteristics it captures a “vehicle signature.” When Flock ran side-by-side comparisons with the $50,000 incumbents, customers were confused. “Something must be wrong with your camera. You’re seeing more cars,” they said. Flock was capturing 30-40% more vehicles, because it picked up all the cars with plastic license plate covers, temporary tags, or no plates at all. Before Flock, a criminal could beat the old systems by removing their plate. With Flock, that car still exists in the system as a white Ford F-150 with a roof rack.

Subscribe now

As Flock signed more police departments, they kept hearing the same request. Cops maintained hot lists of stolen cars, wanted vehicles, and missing persons, and they wanted to know the moment one passed a camera.

When Flock built the real-time alerts system, crime dropped 40-60% in Flock’s first major county deployment. Almost all of it was driven by alerts on repeat offenders who were already known to law enforcement. The cops just needed to know when those vehicles were moving through their jurisdiction. Six months later, Flock’s VP of Strategy Bailey Quintrell was invited to a meeting he hadn’t organized. A large agency convened every law enforcement entity in their million-person county and put a map on the wall showing where Flock cameras were deployed and where they needed more. Because Flock allowed agencies to share camera access with each other, every new department that joined made the network more useful, which created an incentive for word-of-mouth growth.

Despite this traction, raising venture capital was a struggle. Every investor saw the same thing: the biggest departments bought 10 to 20 cameras each, small departments couldn’t afford any, and the TAM was maybe in the low $100 millions. Langley frames the skepticism as a “quadrant problem” – plot sales cycle against annual contract value, and local governments sit in the quadrant venture capital won’t touch. Long sales cycles, low annual contract values, city council approvals, procurement bureaucracy. Every company in the space before Flock had been bootstrapped, and it took them decades to reach meaningful size. But what Flock was seeing looked nothing like this. The smallest agencies in the country were buying 50, 75, 100 cameras. The market wasn’t small. The problem was that nobody had built something good enough and cheap enough to find out how much demand actually existed.

The Market Everyone Missed

The deeper Flock got into the market, the clearer it became why nobody had tried. The infrastructure was stunningly fractured and old.

America has 18,000 municipal police departments. The United Kingdom has forty-five, Australia has eight, and France has two. The US model gives you hyper-local accountability, where you might know the officer who patrols your neighborhood, but it also means criminals cross city lines with impunity while agencies share information via PDF attachments and, until recently in some states, filing cabinets. The FBI’s national stolen vehicle database updates once a day via what Langley describes as “a CSV file that gets sent around on FTP servers.” If your car is stolen in South San Francisco, it takes twenty-four hours for the national database to catch up. As recently as 2022, it was illegal for law enforcement agencies in Florida to host data in the cloud. A top-100 American city was running entirely on paper records two years ago.

Flock’s cloud-based platform cut through much of this. Where legacy systems ran on local servers and CSV exports, Flock’s architecture lets any department share data with a neighboring department through opt-in agreements, instantly and on terms each agency sets. When San Francisco adopted Flock, Oakland’s crime went up as criminals shifted across the city line. Then Oakland adopted Flock too. Once both departments agreed to share data, a stolen car that crossed the border didn’t vanish anymore.

Then there’s the police staffing crisis. Early retirements surged in 2020 and never recovered. Many replacements are in their early twenties and haven’t actually solved homicides. One city Langley visited will buy you a house if you agree to be a cop.

For departments running at 40% staffing, Flock fills the gap. An investigator who used to spend hours navigating dozens of browser tabs and filing cabinets to build a case now gets much of that work pre-assembled. Flock calls this “amplified intelligence.” AI handles the data aggregation and pattern matching, but a human decides whether to pull someone over, and a judge still issues the warrant.

What Cameras Can Do

A 911 call in a major American city reported an attempted homicide. The only detail the caller provided was that the suspect was wearing white Converse sneakers. A crime center operator used FreeForm, Flock’s natural-language search tool, to query the camera network and identify the individual on a nearby feed. The operator pushed the video to the nearest officer, and the suspect was arrested. All of this took seventeen minutes.

In a town in Colorado, an armed robbery at a Levi’s outlet triggered a drone launch that tracked the fleeing suspect from 400 feet. The suspect, oblivious to the fact that he was being tracked, drove home, and as soon as he pulled into his driveway, was arrested by two officers. In Elk Grove, California, drones reach a 911 call in sixty-eight seconds, compared to the department’s previous seven-and-a-half minute average. For vehicular pursuits – one of the most dangerous things police can do – departments are sending drones instead.

Flock now serves over 12,000 customers across law enforcement and private industry. The original solar-powered license plate camera has been joined by situational awareness cameras, gunshot and crash detectors, HD video with pan-tilt-zoom, and American-designed drones produced at a 97,000-square-foot factory near Atlanta. FlockOS ties everything together as a real-time intelligence platform that integrates Flock’s own devices, third-party cameras, dispatch alerts, body-worn cameras, and traffic data.

Corporate customers compound the network. Hotels, bars, HOAs, and retailers install cameras for their own security, and when a crime happens nearby, they can share feeds with local police, becoming force multipliers the department doesn’t pay for. The customer decides whether to share, and on what terms. Some grant 24- or 48-hour access for a specific incident and then revoke it; some share nothing at all. Every camera that goes up is the customer’s, but when customers choose to share, even more crimes can be solved.

The Evidence

San Francisco is one of the clearest examples of what happens when a city commits to the full safety bundle: prosecution, modern technology, and political will. In 2024, the city deployed 400 license plate readers and 80 drones across the city. Major crimes fell 44% from 2023 to 2025, and car thefts dropped 54%. Evan Sernoffsky, a spokesperson for the SFPD, credited “tools like drones, ALPR, and public safety cameras.” The decline shows up across nearly every crime category:

Flock supported over 1 million criminal cases in America last year. In some cases, Flock provides the initial lead. In others, it’s the entire evidence chain outside of the officer physically making the arrest. Flock now serves over 6,000 communities, well north of 50% of the American population by coverage, helps resolve roughly 20% of all reported crime in the regions it’s active, and returned over 10,000 missing people last year – roughly one every hour.

The greatest deterrent to crime isn’t the severity of punishment but whether you think you’ll get caught. “Certain punishment means no punishment,” goes an old Chinese proverb. Raise the probability of getting caught, and the system punishes fewer people, not more. A study of Denmark’s DNA database expansion found that simply increasing offenders’ probability of being identified reduced recidivism by 43% in the first year, with effects persisting for at least three years. In 2022, a Southern California rap song warned listeners to stay away from Flock cameras. More recently, a well-known criminal in the Bay Area announced on a podcast that he was getting out of the crime game because license plate readers, cameras, and drones made it too hard not to get caught. He was going to get a job instead.

Las Vegas tested how much could be achieved with how little. A public-private partnership with Flock added less than one percent to the police budget, and police shootings of suspects dropped sharply. Vegas now has the highest murder clearance rate in the country at well over 90%.

The Cost of Cancellation

Since the start of 2025, at least 30 American cities have either canceled their Flock contracts or deactivated their cameras. The strongest version of the opposition’s argument isn’t about any single camera. It’s about what a network of cameras can become when each jurisdiction sets its own access rules. Surveillance infrastructure built for one purpose can be repurposed for another. And the controls Flock has built into the system – auditable searches, thirty-day default retention, per-jurisdiction access – only work if they’re enforced.

Some of the strongest answers to those concerns have come from outside the company. In January 2026, a federal judge in Norfolk, Virginia ruled that the city’s Flock camera network did not violate the Fourth Amendment. Judge Mark Davis found that the cameras captured “discrete pictures at different locations” rather than the kind of continuous tracking the Supreme Court has flagged as constitutionally problematic. The ruling joins more than thirty state and federal decisions that have upheld fixed-location ALPR systems.

But Langley doesn’t dispute the broader concern. Every powerful technology has been misused. Airplanes have been used for terrorism. Cars kill 40,000 Americans a year. The smartphone in your pocket streams continuous location data to dozens of apps and data brokers without a second thought – what you search, who you message, where you sleep – stored indefinitely and sold to advertisers. His argument isn’t that Flock can’t be misused. It’s that the answer to a powerful technology that can be misused isn’t to ban it but to build mechanisms that catch misuse and hold the misusers accountable. “Every search on our system is logged with the officer’s name, the reason, and a timestamp,” Langley says. “Every jurisdiction’s policy is public. Misuse leaves a trail by design, and that’s how every documented case of misuse has been found.” Last month, for example, a routine audit of Richmond Police Department’s Flock usage caught a sergeant who had shared a vehicle image with an FBI agent investigating a homicide outside Virginia, which was a violation of state law. The sergeant’s access was revoked, and the department disclosed the violation publicly.

Beyond accountability mechanisms, Flock pushes policing in a direction that should satisfy civil liberties advocates rather than alarm them. Oakland adopted the system specifically because they wanted to stop officers from driving around looking for suspicious people. Historical policing was deeply prejudiced: you went to dangerous neighborhoods, looked for suspicious people, and arrested them. That approach perpetuated cycles against specific communities. A system that flags stolen vehicles instead of suspicious people is, in his view, a step forward for civil rights, not backward.

The consequences of removing cameras are harder to debate. In 2021, Denver had 96 homicides. The police department mapped the specific blocks driving most of the violence and concentrated resources there: more officers, better lighting, license plate readers. By 2025, homicides had fallen to 37, the lowest in over a decade, and the largest drop of any major American city. Yet, on March 31, 2026, all 110 license plate readers were decommissioned after the city council declined to extend Flock’s contract over concerns that data could reach federal immigration authorities. Now the numbers have started moving the other way. “In the last probably 10 days, we’ve had six homicides, I believe,” Police Chief Ron Thomas said in an interview 21 days later. “We’re 50% above where we were this same time last year.” He stopped short of blaming the cameras’ removal directly, but acknowledged that going “blind for some period of time” without Flock had “hampered our investigative efforts.”

When cities remove their cameras, they are choosing fewer crimes solved, fewer missing persons found, and fewer cases cleared. The people who pay the price for that aren’t the activists who organized the removal but the residents whose stolen cars don’t come back, whose break-ins don’t get solved, whose missing children don’t get found. The political leaders who made that decision should be held accountable for those outcomes.

The Long Game

“If I woke up in ten years and all we had done was put a lot of people in prison,” Langley says, “we’d have failed.” A million arrests means a million victims. It also means a million people entering a system with a 70% recidivism rate that will likely make them more dangerous than when they went in.

The majority of crime Flock solves is nonviolent, and even a juvenile convicted of a nonviolent offense can end up in the penitentiary system, where data shows they are more likely to go from nonviolent to violent within months and never escape the cycle. Langley is exploring whether Flock can help build a system that gives nonviolent offenders a path back without increasing their likelihood of reoffending.

Flock launched the Thriving Cities Fund in 2024, investing $10 million in restaurants, nail salons, and other businesses in cities where Flock is deployed – the kinds of businesses where a sixteen-year-old can get a job instead of pulling truck door handles. In Las Vegas, a program called Hope for Prisoners pairs people coming out of prison with mentors and job placement, and a UNLV study found participants reoffend at a fraction of the national rate. The Vegas prosecutor’s office now sends some first-time non-violent offenders straight there instead of prosecuting them. In Atlanta, four @Promise Centers run by the Atlanta Police Foundation reach kids earlier, before the first arrest, offering workforce training, mentoring, therapy, and GED support. Langley wants Flock to help more cities do the same.

If you’re affluent, you can hire private security or move to a safer neighborhood. If you’re not, crime is simply the cost of being alive. That’s the gap Langley has been trying to close from the beginning.

In 2018, The Outline wrote a profile on Flock titled Snitching-ass startup raises $10 million to privatize the surveillance state. “I framed it,” Langley said. “It’s a reminder that when you work on important things, some people will disagree, and that’s okay. It means you should work harder.”

His uncle told him to go build something that wasn’t so dopey. Nine years later, his business has helped solve over a million criminal cases and brought home more than ten thousand missing people. The founder of the largest public safety network in the country is spending his time thinking about how to keep people out of prison, and how to make sure the next neighborhood that needs a camera doesn’t have to be rich to get one.

It’s hard to call any of that dopey.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Today’s post is by Will Bryk, cofounder and CEO of Exa. Exa announced their Series C this morning, led by a16z.

AI agents will search the web more than humans this year. In a few years, they’ll search 1000x more than humans. This new AI world needs search that’s redesigned from the ground up.

My cofounder Jeff and I built our first search engine together as college roommates. We felt the web was filled with an immense amount of valuable information, but traditional search only gave surface-level understanding. This is a serious problem, since the information we consume is the bedrock of our society. If we can’t fully grasp our chaotic world, then that leads to more chaos.

We started Exa five years ago to solve the same problem, but for an AI world. We knew AI systems would one day need a new kind of search engine. They would ask longer, more complex questions than humans ever had. AI products would require a latency profile and high search volume that the world’s best search engines weren’t designed for. AI was getting smarter each month, faster than anyone expected, but search wasn’t catching up. This felt like an opportunity to redesign our information ecosystem.

Traditional search engines were built for people clicking through a few links and reading a few webpages to learn something. In contrast, AI agents are doing real work on top of the web. Traditional search’s entire architecture is built around short keyword phrases, ranking shaped by human clicks, and an SEO ecosystem to win position in the ranking. None of this helps the AI agent. An AI agent wants an API to access comprehensive, trusted information so it can perform actions optimally. The higher quality the search, the more valuable it is to an agent. Luckily, comprehensive, trusted information is exactly what my cofounder and I wanted for ourselves. Our mission and business model were therefore aligned.

Since those early days, AI search has gone from nonexistent to a significant force today; it’s grown more than 1000x just in the last few years. And it will grow immensely from here. By the early 2030s, I believe agentic search will be a larger business than Google search ads is today.

Along the way, we’ve learned that search engines are extremely difficult to build. I underestimated this myself when we started Exa (and that’s probably why we were willing to start it). To build a truly useful search engine for AI, you need to crawl and process hundreds of billions of everchanging documents, devote years of research into retrieval models that handle complex queries, and design new vector databases for the extreme volume that agents need. There’s a reason there are more space programs than true search engines.

That’s also why most companies selling “AI search” are in fact wrappers over existing search engines. That’s fine for basic queries. But those underlying search engines were built for humans. Actually building real search for AI is much harder.

Exa has done the hard thing of becoming an AI lab for search. We track hundreds of billions of URLs, train novel embedding models on a GPU cluster we assembled ourselves, and design custom vector databases for extremely high throughput. We offer the fastest search in the world—under 200 milliseconds—for real-time agents, and the highest-quality search at 10x faster and cheaper for asynchronous agents that need comprehensive data. Recent hires include a head of retrieval infrastructure from Meta, a former head of search backend at Yandex, and a research team out of Google.

While we’re not used to search engines improving, Exa’s search gets better each week. For example, six months ago, we were worse than Google at code search; today nearly every coding agent uses us.

Exa now serves hundreds of thousands of developers and thousands of enterprises, including Cursor, Cognition, HubSpot, and Monday.com.

These are only the early days. We just raised $225M at a $2.2B valuation, led by a16z. Our goal is perfect search for AI agents: your AI system should have truly comprehensive knowledge of the world, no matter how complex the request.

The bottleneck to AI is increasingly not intelligence - we have plenty of that now. The bottleneck is finding the right information. Exa is what I think the search engine for the new world looks like. We have a lot more to build.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

For the last two decades, search was built for humans.

Google’s PageRank turned the web into something navigable: a system of links, rankings, and interfaces, designed to help people find the information they wanted. Entire industries emerged around this model, from SEO to advertising to content optimized for human consumption and focused on maximizing clicks and time spent per page.

But we’re entering a new era for search, built from the ground up around what AI can do.

Increasingly, every meaningful AI workflow, from coding agents to research systems to enterprise copilots, depends on external information as a core input. LLMs are frozen in time; a search engine for AI exists precisely to compensate with fresh, long-tail, real-world context. If the underlying data is stale, incomplete, or incorrect, everything downstream breaks. As one leading AI company put it: “In the limit, if we could search 100% of the time, we probably would. It just comes down to GPUs, latency, and cost.”

It’s easy to underestimate how difficult building search for agents is. Basic search is a commodity; simple keyword searches are cheap and easy. But agents don’t ask head queries. Their alpha comes from probing long-tail, constantly shifting information. They write complex queries that can span paragraphs long. In some cases, like with low-latency voice AI, agents need results instantly; but in other cases, as with KYC (Know Your Customer), they need to scan millions of pages to synthesize the most comprehensive answers. There’s a Pareto frontier in cost, latency and comprehensiveness for every search query, one that traditional search engines were not designed for.

We’re also marching into a different scale paradigm: agents will search 1,000s of times more than humans will. Exa’s CEO and Co-founder Will Bryk puts it simply: “We’re organizing the world’s knowledge, but this time for AI.” Doing that reliably and economically, at agent-scale, is a genuinely hard problem. It requires building and controlling the entire search stack – something only a few companies have ever done.

That’s why we’re excited to lead Exa’s Series C, to back their ambition of perfecting web search, and making it ready for the age of intelligent agents.

The signal from the market was consistent. Exa excels with the hardest queries – the long tail of high alpha searches where traditional engines fail. It stands out on low latency where it matters, especially time to first token, which matters acutely in user-facing agent flows. What struck us most was the default behavior: developers and agents are reaching for Exa first. As one customer put it, “this is the default for getting agents to do web search now.”

No ordinary team would take on such a grand vision: building a search engine from scratch for AI. Will has been obsessed with perfect search for years. He and his roommate, and cofounder, Jeffrey Wang built a search engine in their dorm at Harvard a decade ago. Years before ChatGPT or the AI boom really started, Will and Jeff were inspired by the transformer breakthrough and believed deeply that AI will fundamentally change the way we access information. So they set out on the path to build the search engine for a future where agents become the primary consumers of the web. What stands out is not just their technical depth, but their clarity of focus, developer-first instincts, and hustle.

Today, Exa is being used by the leading AI companies, from startups on the frontier like Cursor and Cognition to large enterprises like Hubspot, Monday.com and many Fortune 500s. It’s also serving hundreds of thousands of developers who are building agents that are reliant on the accuracy and reliability Exa provides.

As AI both consumes and produces the internet, the web will become bigger, faster, and noisier. Search will become the compass for agents to navigate it. The first search wars were won by organizing information for people. The next will be won by organizing information for agents.

We’re thrilled to partner with Will, Jeff, and the Exa team as they build the perfect search engine for agents and usher in a world of abundant information for all.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

We’re excited to welcome General (R) Bryan P. Fenton who is joining a16z as a special advisor. Read more about it here. - AD

Who knows something important and misunderstood about AI adoption? This guy:

I don’t mean cartoon Ben & Marc here of course; I mean the people you usually associate with this kind of ad: plaintiff attorneys. And I don’t say that because they’re adopting AI tools, or because their work is going to change in some way. I actually mean the opposite. I think their business is already in the future, and has been for some time.

Personal Injury succeeded as a business category by finding an economical way of making a value-aligned results-based appeal to lots of people. It took a thing previously out of scope for most people (legal services), and with clever marketing, it discovered unit economics that made that possible at scale. To me, this sounds a lot like how AI has been positioned, as an accelerator or enabler of new consumer business models!

Here are a few non-obvious (but actually kind of obvious, in hindsight) things I found interesting about Personal Injury that seem applicable here:

The business model of leadgen plus underwriting.

When you see all of those highway billboards for personal injury lawyers with extremely catchy names, one minute’s reflection will tell you, this is obviously a leadgen business. This business was born on Cable TV, reinvented itself around Google Adwords, and will have no trouble reinventing itself here.

Our personal injury lawyer intuitively understands this business model, and he’ll figure out how to juice it to the gills with AI that’s really good at sourcing & handicapping. He understands how to cast wide nets of possibility, pick promising leads, and size his bets.

When people say “everyone with AI is a product manager / portfolio manager”, these skills are what people mean. It means “are you good at bet sizing, and can you apply it to your task at hand?” AI can trivially get you an avalanche of information as an input to any problem; it can trivially carry out (some) tasks if given instruction. This makes you, the user, the underwriter, and you have to learn a kind of pragmatism and relative ruthlessness to succeed here.

There are so many industries that rely on this business model, one way or another. Insurance (of course), loans, legal services, para-health care, and auto sales & financing are all examples where the hardest and most lucrative part of the job is getting qualified leads. There are also endless personal niches like home rooftop solar installation (to give one example) where, once you have someone interested, anyone can really do the job, and it’s a valuable job, so you’re gonna shop it. There’s no doubt in my mind that AI accelerates this business model.

CAC is about recall, not merit. If you think about the stereotypical personal injury lawyer billboard, it’s all about personalities and being memorable, and it’s about recall at critical moments. (“Hurt in a car? Call William Mattar!”) The entire game is, are you the first jingle that comes to mind? This feels suspiciously relevant for a world where coding agents or harnesses could be choosing which tool to call or approach to take based on a version of “who does the LLM recall first?”

Acquiring customers for these kinds of services may end up less about comparative merit or feature advantages, and much more about “who comes to mind for the LLM first?” than most people would like to admit.

There’s some poetic justice here, in the sense that recall has always been one of the most important factors guiding our decision-making. It’s just that we pretend it isn’t. Everyone believes themselves to be a rational, informed decision-maker (when in reality, we often subconsciously choose our favorite, and rationalize it in conscious thought later.) Now with LLMs, there may be a bit of mystery in the black box (why did it call this tool, as opposed to that one?) and I suspect that “who does it recall in a time of immediate need” is the right mental model for marketers here.

There are actually a lot of businesses that fall into this category: e.g. plumbers and locksmiths, appliance repair, emergency health care or veterinary services, towing, funeral homes (a grim one), pest control, short term moving and storage, and many more. The theme here is obvious: in these critical moments, you’re going to be just distracted (and probably embarrassed) enough that, when you Google search for some options, the subconscious “trust” you have in a name is going to beat price or value comparison shopping, 9 times out of 10.

Contingency fees are the “aligned business model”. The average Silicon Valley tech person pays for AI on some combination of subscription and pay-for-tokens, and those are both fine business models, but they don’t cover everything. For most consumers out there, I suspect that contingency fees - “pay if we successfully get the job done” - are the right model for a lot of tasks, especially zero-sum ones where you either win or lose.

Beyond personal injury law and insurance, there’s a wide world of “helping transactions happen” where we can look for inspiration. Realtors, while not “pure” contingency, are an obvious one. There’s also a whole industry of tax refund recovery firms (helping you identify missed or non-necessarily-guaranteed credits), ranging into esoteric but profitable niches. (In Canada, for example, there’s this lovely cottage industry of “SR&ED consultants”, helping startups claim government benefits, that are mostly paid on this basis.) And finally, up in the elite tiers of the economy, isn’t this what a lot of investment banking is?

A lot of business model choices are downstream of this choice - for example, how much volume of business any given service provider (e.g. any given personal injury lawyer!) can take on. And that matters because, the other thing we think AI will unlock in a lot of places is:

There’s no upper limit on consumption. This kind of law (e.g. personal injury suits) is maximally Jevons Paradox-compatible. Businesses services that are “win a fight with a perceived adversary” are the kind of product offerings where, the cheaper you can make it, the more of it people are going to consume. If legal costs come down, they will almost certainly get consumed more by the general population.

Today, it only makes sense to take on a case whose expected payout will be above some threshold - the fixed costs of bringing a case to trial won’t net out if the case is too trivial. But we could find a new equilibrium where far, far more cases get taken up for consideration.

Brazil is an instructive example of what this looks like. For a variety of reasons, Brazil has evolved a set of legal institutions and social norms where consumers sue companies for a variety of everyday things - like your food delivery order being late. These disputes are taken up in a special small claims court, the Juizado Especial Cível (JEC), where complaints can be brought for a small fee and without necessarily requiring a lawyer. And there’s a thriving secondary market for buying and selling the rights to consumer claims: the claims backlog is able to clear, in large part, because the secondary market helps it clear (by consolidating claims, getting them settled out of court in bulk, or even ‘bought back’ by the company accused in the dispute.) The point here is, we don’t have to imagine a world where Jevons Paradox has made certain kinds of legal complaints 10x more prevalent; we can go to where the future has already materialized, and observe how institutions and market mechanisms react.

Finally, maybe the most important point I want to make here:

The “everyman” user isn’t always a hopeful-builder type. Most of the “regular people using AI” user stories we hear in Silicon Valley are “trades guy learns to build new kinds of stuff” feel-good positive sum stories, and to be sure, those are great. But there are going to be some people out in the world who aren’t naturally inclined to go out and build stuff; but they are naturally inclined to get things, or win zero-sum fights, and these people will also use AI. They are a less romantic version of the “empowered individual” but there are a lot of them out there.

(Imagine your average person out there who’s like, “AI seems cool, but I don’t really have any ideas of stuff to build.” That’s actually a lot of people! You know what’s a lot less hard for them to understand? “Use AI to get $500 out of your airline for cancelling your flight.” This is a large user category.)

In some ways, these ‘uncouth’ customers are better signals (or, at the very least, complementary signals) for builders and investors than the hopeful, romantic types. Their motives are naked and pure: they’re in a fight, and they want to win the fight! When AI really penetrates a lot of consumer experiences - for example, submitting an insurance claim - we’ll see an escalating arms race on both sides of the dispute: the AI-equipped David consumer, versus the AI-equipped Goliath claims adjuster. (Insurance companies literally use a back end system called “Colossus”, if that helps paint the picture.) Serving these customers is a bs-proof frontier for any technology: it either works, or not.

These kinds of user journeys, in some sense, are actually pretty pure examples of “the empowerment of the individual.” If we imagine AI as empowering both motivated individuals and entrenched systems in their everyday interactions (which it likely will), these are your prototype customers. Plaintiff attorneys may get flak as being “ambulance chasers” or some other dismissive label, but fundamentally they do represent the individual, and now the individual has ChatGPT.

Post-script: Get out of your comfort zone

If you want a head start on the future - for investing, for career advice, for anything - find people who are already living there, and study them. Those people are not always who you expect. Silicon Valley has a tendency to herd around a few specific archetypes of people that are “the AI user of the future” (e.g. the romantic tradesman that vibe codes himself a better CRM, or something). And one way to escape the groupthink is to identify other kinds of user stories, specifically those that are less beloved or romantic.

I think it’s important for everybody to spend some time reading, meeting, and thinking about the kinds of people that are not the “classic paragon of virtue” in your milieu. Odds are, if you’re reading this, that the people around you are in fact not representative of middle America, the global consumer class, or of your median small business entrepreneur. This is why it’s a good idea for everyone to have a bit of experience in sales, for instance. Not only do you meet people at their level and learn what they really want, but you also learn from other people who understand something fundamental about how to sell to people - like our friend the plaintiff attorney, who may just have some timeless people wisdom.

Thanks to for inspiring this post.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Memory to the moon

First it was power and chips, and now memory has entered the bottleneck chat:

Prices for both DDR5 and DDR4 memory chips have skyrocketed, since the end of last year, not coincidentally with a ~25% rise in PPI for the “computers” category.

The bidding war isn’t limited to DRAM (especially High Bandwidth Memory), which is particularly important to model training, but traditional steady-state NAND, as well:

As of March, monthly contract prices for DRAM had more than tripled yoy, while NAND prices grew by a more modest ~2X.

Truthfully, there is no real hierarchy in the list of AI-infrastructure related shortages, from power, and skilled labor, to chips (GPUs and CPUs) and memory. An historic surge in demand, backstopped by ~$1T in capex, will have that sort of effect, but memory is the most prominent recent addition to the list.

The run on memory has led to a pretty epic quarter for the leading memory-makers (Samsung, SK Hynix and Micron):

Memory-makers are expected to sextuple(!) their operating income in 2026. Micron itself earned more in ‘Q1 than it had in any single year before 2025.

Net-income gains for memory are so massive (and off a relatively lower base), they’ve broken the y-axis for the AI supply chain winners:

So, memory is the new belle of the ball. But, the funny thing about memory is that it used to be one of the most commoditized parts of the chip-stack. Despite being fairly ubiquitous in phones, PCs, and gaming, memory never really enjoyed any pricing power (and the success of those memory chipmakers was very much tied to the consumer goods cycle). Memory competed on price, and the name of the game was “how low can you go?”

But the sheer volume of demand, and the increasing complexity and power requirements of AI-related memory needs (where efficiency is at a premium) has changed everything.

It’s turned even previously sleepy low-cost memory-makers, like Western Digital, from horizontal lines into vertical ones:

Western Digital has done great. But if we put its DRAM specialist spinout (Sandisk) on the chart, it makes OldCo WDC performance look more like a modest hill than a legendary spike.

But, will this run on memory—and the extraordinary pricing power the memory chipmakers now enjoy—ever end? If demand is so high, then why not make more memory chips?

Well, it can take years to bring new capacity online, and in the meantime, manufacturers are prioritizing higher-margin, and more silicon-intensive HBM, which puts even more pressure on the supply of consumer memory chips. Plus, hyperscalers are now signing 5-year agreements—the industry standard used to be 1-year—to ensure their place at the top of the supply queue. As a result, even the more commoditized consumer chips are getting much more expensive, and those costs will flow through to either iPhone, etc.’s margins and/or consumer prices—expectations are that prices on PCs and phones will increase ~10-20%.

At the same time, everyone knows the saying “every shortage becomes a glut,” and no one knows that better than cyclical manufacturing businesses.

They are traditionally very hesitant to invest big bucks in new capacity to meet even substantial surges in demand, knowing that once the surge fades, they still have to foot the bill for all that new capacity. Plus, necessity is the mother of invention, and when costs rise, so too does the innovation-drive for efficiency (which we’ve already seen, e.g. with Google’s TPUs and Amazon’s Trainium, as part of their effort to be somewhat less dependent on NVIDIA’s GPUs, and more recently, with the release of ‘TurboQuant’).

On the other hand, the upward inflection in computational demand may just be getting started in the new-AI-normal, and that would make memory a lot less cyclical than ever before. So, is it a step-change or a one-time thing? That’s a multi-hundred billion question right now.

Affirm’s AI Turn

Rather than replacing workers, there’s every reason to think that AI will function as a force-multiplier, making at least some workers more valuable than before. SWEs are a perfect example of an “AI Augmented” role, and it’s not surprising that demand for SWEs appears to have increased.

It flew under the radar for whatever reason, but the publicly-traded BNPL company Affirm recently gave some very concrete evidence of the AI-driven productivity lift, and the positive impact on Affirm’s hiring intentions:

Earlier this quarter we took a full week as a development team to re-tool Affirm software engineering to be truly AI-first. You can see the difference below, with agentically-written code dominating non-agentic essentially overnight after this effort. The chart shows pull requests per week (each pull request is a product feature, a major bug fix, etc) as a function of time.

We think an order of magnitude gain in software development productivity is entirely plausible, yet we expect to modestly grow our development team – the limiting factor for Affirm has always been engineering cycles, not ideas for what to do with them.

When Affirm “re-tool[ed] software engineering to be truly AI-first,” PRs more than doubled, and roughly two-thirds of the output was “agentic.” As a result, Affirm plans to “modestly grow our development team.” More productivity, more hiring. Checks out.

But the real money quote came at the end: “the limiting factor for Affirm has always been engineering cycles, not ideas for what to do with them.” In other words, we are nowhere near peak-demand for software, and rather than put SWEs out of work, lowering the cost of development is expanding the frontier of what gets built.

Affirm is, of course, just one company, but there is mounting evidence of productivity lift across “AI Industries,” as a whole:

Output/employee inflected upwards in 2025, but especially amongst the “High AI Industries.” And within those industries, it’s tech-related industries doing most of the work.

Now, productivity is notoriously hard to measure in the aggregate, and a lot of this additional output simply reflects the surge in demand (and pricing) for AI-infrastructure. Higher prices for the same thing leads to more measured “output,” without any increase in output. But, at the same time, it also means that we’re very early in this transition, and even industries that stand a lot to gain from automation, have barely begun to scratch the surface of what they’re eventually going to do.

Agents, especially, are extremely early in their lifecycle, and some of the biggest AI adopters, are still barely entering their agent-phase, at least according to some data published by Microsoft:

Nearly 1 in 5 firms are using agents of some kind, but that also means that only 1 in 5 firms are using agents of some kind. Manufacturing is a surprise outlier for agent-deployment. While manufacturing firms comprise less than 10% of the firms using agents (and relatively few “prompts per user”), they account for ~18% of the agents, higher than any other sector.

The future isn’t even here yet, it’s definitely not evenly distributed, and we expect a lot more to come.

Shrinking SmallCaps

While many people are probably aware that there are fewer public companies now, it was somewhat surprising to learn that the shift is driven entirely by the disappearance of small public companies:

The number of large- and mid-sized public companies has stayed relatively constant since 2000 (but really, since 1990). It’s the number of microcaps (and to a lesser extent smallcaps) that has really gone from boom-to-bust.

The SEC attributed the change to three factors: (1) lots of M&A; (2) smallco delistings; and (3) fewer IPOs. Other factors include regulatory costs, which generally don’t scale with market cap (i.e. regulatory costs are a larger share of market for smaller companies), and of course the growth of private equity, which has made it easier for companies to raise capital without recourse to the public markets, at all.

These are very rough estimates, of course, but research has repeatedly shown that the fixed compliance costs of being public are proportionately higher for smaller companies. In that sense, compliance is a regressive tax.

There are other structural reasons for smaller companies avoiding the public markets, as well: small caps don’t benefit from passive flows, less liquidity makes it harder for public market investors to maneuver without big impacts on price, and smaller float makes smallcaps less interesting to sell-side coverage (which leads to less liquidity).

The net result of all of this is that smallcos have retreated from the public markets (not entirely of course), but they haven’t disappeared: it’s just that the action has moved to the private side. Of course, the decision to go public or not is ultimately specific to each company, but the broader trend is real, and it’s a function of structural changes that have made public markets less-appealing, relative to their private alternatives. If we want more companies to go public, then perhaps public markets ought to up their game.

California’s Healthcare Jobmaker

Come on California, we can do better than this:

It’s time to build.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Yesterday we heard from Seema Amble with Is Software Losing Its Head?, today we continue on the same topic with this piece from Gio, Steph and Alex. Enjoy.- AD

Here’s one way you can think about system of record stickiness:

For a long time, the valuable part of social media businesses was the friend graph. When you opened Facebook back in the day, the thing you interacted with was people’s profiles, and the data graph across the profiles was a powerful, durable asset. It was hard to foresee what could disrupt such an obvious network effect.

Then the news feed came along. The news feed gave us a new place to go: “Here’s what happened today; here’s where you catch up and take action, all in one place.” This started out as a complementary layer to the friend graph, but in time, the graph became “just one of many inputs” to the feed serving you relevant content. While it never went away, it’s no longer the important layer - the feed algorithm is, and all kinds of things feed into it. Your social profile, posts and likes are primarily consumed “at the internal API layer”, so to speak; the newsfeed is its consumer.

We think this is starting to happen to one of the supposedly “least disruptable” parts of the enterprise: the CRM. The CRM isn’t going to go away, just like the friend graph never went away—but it’s turning into just an input; one of many inputs, into the systems of intelligence which we use to get work done.

At firms across the country, the typical account executive now opens his laptop in the morning and finds, waiting for him, a small collection of software agents he had no part in programming — a research agent that combs 10-Ks and recent earnings calls before his first meeting of the day; a dialer that coaches him on objections in the moment; an orchestration layer that listens to his calls and writes structured notes back into the CRM without his lifting a finger. None of this, by itself, is earthshaking. But taken together, you recognize what this is: this is the newsfeed. It’s the valuable thing now.

There’s no doubt: owning the system of record has been the winning play for go-to-market software for twenty years. It’s sticky, valuable, and hard to leave. And we can’t imagine the SoR incumbents going away anytime soon: Salesforce and HubSpot still sit on some of the most valuable datasets in the industry, they’ve realized that it matters, and they’re quickly coming up with API-first offerings that bring AI features in their own walls.

But we think we’ve seen this movie before. In the next decade, you want to own the system of intelligence that pulls from the system of record, becomes the user’s one-stop shop for gaining context and taking action, and turns the SoR into something that’s primarily consumed at the API layer. The reasoning layer that sits above the database, and that increasingly treats the database as infrastructure, is where a new generation of companies is being built, and it’s where the majority of the next decade’s enterprise value of GTM software will end up.

Why the database won

Over the last thirty years, software companies have produced an unbelievable number of products to help companies manage themselves. A thousand companies were founded to help salespeople sell; but almost all the value ended up accumulating in just two names: Salesforce, today valued at around $140 billion, and HubSpot, valued around $9 billion. As the line goes, “First prize is a Cadillac. Second prize is a set of steak knives.”

The reason, everyone in the industry has long understood, is simple: Salesforce and HubSpot own the database. And the database is where all the value resides. Every call note, every pricing precedent, every contact, every stray observation about why a deal had stalled is entered into the system, and the cost of leaving it behind becomes enormous. Once that database has accumulated a few years of operational context, switching costs become, as our colleague Alex Rampell has put it, high enough that users are “hostages, not customers.” Every app in the Salesforce AppExchange and every tool in the HubSpot Marketplace is, in effect, paying rent for the right to plug into someone else’s database.

Then, Salesforce and HubSpot do what every dominant platform owner in every era does: they expand outward. They add features like marketing, service, analytics, and commerce: each new module built on the same data spine, and each further raising the cost of any decision to leave.

One of the more counterintuitive findings from our GTM survey is that CRM usage has actually risen since AI tools began to be adopted at scale. The agents that listen to calls and write structured notes back into the system are, for the moment, giving reps fresh reason to consult it, because the data sitting there has become dramatically richer than it used to be.

Orchestration is the new gravity well

AI agents, acting on behalf of sales reps and alongside them, are taking over a steadily widening share of the GTM workflow. Sometimes the rep instructs the agent directly: research this account, draft this outbound sequence, qualify these inbound leads, update this deal record after the call. Sometimes the agent works in the background, listening to a meeting recording and writing the structured fields back into the CRM on its own.

And the agent doesn’t need a drag-and-drop pipeline view. What it needs is structured data it can read and write with low friction. The CRM, from the agent’s perspective, is a database. A very large and carefully curated database, hosted by a trusted vendor, with excellent integrations and a decade of accumulated customer trust; but a database, nonetheless. The opinionated workflows on top become, progressively, legacy furniture - a bit like the lovingly created UI of your Facebook profile; once paramount, now an afterthought.

In the software era, the gravity in enterprise software came from data accumulation: that is, from the fact that every valuable piece of sales context had to live in one place because the human operating on that context could only look in one place at a time. But in the AI era, gravity will come from orchestration. An AI agent doesn’t find it difficult to pull dozens of signals simultaneously from the CRM, the calendar, the shared inbox, the call recording, Slack, the enrichment API, the billing system, and the product telemetry. Nor does it find it difficult to synthesize information across all of them before actually taking any actions.

Subscribe now

Switching costs shift accordingly. “All of our customer data is in Salesforce” becomes “all of our workflows, our reasoning, our accumulated institutional context live in our AI layer.” The CRM used to tax every app that wanted access to its data; now the system of intelligence has become the hub, and the CRM is one of the many systems of record that it orchestrates across.

At the technical core of the new stack sit the foundation models. But a foundation model is not, by itself, a GTM application, any more than Oracle’s database engine was a CRM. Between the model and the customer sits an enormous amount of unglamorous and domain-specific work: orchestrating context across dozens of connected systems, encoding the actual logic of how sales and marketing teams operate, handling permissions and compliance, integrating with the chaotic reality of a Fortune 500 IT environment. That work is the new GTM application layer. It is where the new GTM companies are being built.

Go-to-market has, for decades, been a category in which software was the junior partner to labor. Historically, software made up between 5 and 10 percent of total GTM spending in a typical enterprise; the rest is payroll. Salesforce dominates the software slice, but the software slice has always been a thin wedge of the pie. What AI opens up, for the first time, is the prospect that software companies can meaningfully reduce costs while opening up new high ROI use cases.

The natural question is whether this comes at the expense of sales headcount. So far, it has not, or at least not in a straightforward way. While roles within the GTM team may shift, we’re seeing teams spend even more on people. The ROIs on these agents are strong enough that the total pie grows rather than the labor budget shrinking. Reps using these tools are hitting attainment and quota at noticeably higher rates than those without them; the return on every GTM dollar is rising, rather than merely holding steady.

The next wave

There are two observations worth making about the AI-native GTM startups that have emerged over the last few years. The first is that they are clustering, for now, around a few relatively narrow and high-frequency workflows: in all of these workflows, inputs are structured and outputs are measurable.

And while some of them are doing an existing job in a new way, many of them are inventing new jobs entirely: they are doing things that nobody was quite doing before.

Consider, for a moment, the position of a VP of Sales at a typical enterprise software firm a few years hence. She no longer begins her day by opening Salesforce to a static account list and deciding where to focus. She begins it in a prioritized feed generated by her system of intelligence: which of her accounts had material news overnight, which prospects in the territory are suddenly in market, which deals in the pipeline have gone quiet in ways that ought to be investigated. The daily prioritization decision — which used to consume real cognitive effort from every rep and every sales leader in America — has been quietly offloaded to the intelligence layer. Her reps spend more of their time actually selling.

And, when they sell, they arrive better prepared. Prep that used to happen case by case, if it happened at all, now happens every time as a matter of course. The rep who would never have read the 10-K is walking in with a briefing drafted for him; the new hire six weeks into the job is, by certain measures, better equipped than the ten-year veteran at the desk next to her.

More importantly, the VP of Sales has an honest picture of what her team is doing. At the moment that picture is whatever gets logged into the CRM, which is often incomplete and occasionally fictional. With call transcripts, email threads, and calendar data flowing in automatically, analyzed continuously, she can see, at any given moment, who is running disciplined discovery and who is skipping steps, which accounts are getting coverage and which have been quietly neglected. A system of intelligence that has ingested every interaction across a sales team can surface patterns no human manager, however committed, could see unaided.

The longer-run implications push further still, and begin to open up categories of job that did not really exist before. Every company bleeds institutional knowledge when its reps turn over — context on accounts, the history of what worked for whom, the texture of relationships built up over years. A system of intelligence that has been quietly ingesting that context for the duration of a rep’s tenure can, when she leaves, hand the whole of it over to her successor. Institutional memory becomes something a company can actually ship.

None of this, it should be said, is bad news for the CRM. Salesforce still owns its database; HubSpot still owns its database; the customer data continues to live where it has always lived, for the reasons it has always lived there. But the locus of value is migrating upward, into the layer that reads and writes to the database and does the actual thinking. The pie is getting larger in the process, not smaller. Just as the feed increased the TAM of social media to “everything of interest”, the agent revolution expands what software can plausibly charge for, and does it without gutting the labor budget that funds most GTM work today.

A new generation of companies is being built on top of this emerging layer. The next decade of go-to-market software will be written there.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

A system of record (SOR) keeps track of the atomic units of a business — in banking, the holy grail SOR is the core banking system. It stores every account, every transaction, and is the single source of truth that every downstream workflow depends on.

In the United States, FIS and Fiserv became two of the most durable, defensible software businesses ever built because they sit at the center of every bank’s operational universe and banks can’t replace them (their data is trapped!).

The result of this incumbency is that financial institutions are sitting on decades of infrastructure debt. This debt has become the single biggest obstacle to AI adoption. Intelligent agents cannot run on systems without clean APIs and data.

Enter Stitch.

Stitch is building an operating system for modern financial institutions, starting in the Middle East. Its API-first platform enables banks, fintechs, and non-financial institutions to launch, scale, and operate financial products across a modular stack, including a general ledger and core database, product-specific software for cards, loans, and bank accounts, and a customizable workflow builder with a marketplace of integrations including KYC/KYB.

Convincing banks to swap their core is no small task — but a rare greenfield opportunity exists in the Middle East, where a nationwide focus on expanding outside of oil is feeding one of the fastest-growing financial services ecosystems in the world; the industry is growing 10% YoY. Saudi Arabia has plans to add hundreds of new financial institutions over the next five years, all of which need core banking software for the first time.

In the last six months alone, more than $5 billion has been transacted on the Stitch platform. Customer numbers grew 10x in 2025, revenue grew 20x, and Stitch has seen pull beyond the GCC into Africa and Southeast Asia. Marquee regional customers like Lulu Exchange and Raya Financing have already launched products up to 80% faster, thanks to Stitch’s flexible infrastructure.

As Stitch becomes the infrastructure layer, it can support these financial institutions in their AI transformation, adding AI-powered customer support, automating loan origination, and eliminating the messy inbox workflows that consume operational headcount at every financial institution.

Mohamed, Stitch’s founder, has the rare combination of relationships, sales instinct, and grit to win in a market where trust and distribution matter as much as product. And he’s hired a team of world-class operators from the likes of FIS and Barclays to bring his vision to life.

We are thrilled to lead Stitch’s Series A — and honored to make this our first investment in Saudi Arabia. After years building conviction around the global fintech infrastructure opportunity, we believe Stitch represents the clearest expression we’ve found of what the next generation of this infrastructure looks like. It is built natively in a market with massive tailwinds, led by a founder who understands both the technology and the regulatory landscape, and designed from the ground up to be the system of record that makes everything else (i.e. AI) possible.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Is software losing its head?

Last month Salesforce announced it would open its APIs and launch a headless product, essentially betting that in an agentic world, its value lies in the data layer, not the UI. It’s a smart repositioning. (Although it’s worth noting that not much appears to have changed technically: the APIs Salesforce is now marketing as a “headless product” have largely existed for years. In other words it was a classic Salesforce marketing launch.) The idea behind the new product is that agents can access the data from the system of record without needing to interact with the UI, which is designed for humans to track workflows.

The announcement is a useful prompt for a more interesting question: if you strip away the UI and expose the database, what are you actually left with? How is that different from a Postgres database, a well-designed schema, and an API? Do the classic factors that make systems of record durable persist, or is there a new set of criteria? In the SaaS era, the system of record was defensible because humans lived in the interface. In the agentic era, that advantage weakens. The defensible layers shift downward into data models, permissions, workflow logic, and compliance, and upward into networks, proprietary data generation, and real-world execution.

When software goes headless, where does defensibility move?

The UI was the product

A system of record is the authoritative source of truth for a given domain of business data. It’s the place where the official version of a customer relationship, an employee record, or a financial transaction lives, the system other tools read from and write back to. A CRM is the system of record for revenue. An HRIS is the system of record for people. An ERP is the system of record for money. What makes them powerful isn’t just that they store data, it’s that they become the shared reality an entire organization operates from.

For the last two decades, Salesforce sold a way for sales leaders to run their teams. The dashboards, pipeline views, forecasting tools, and activity feeds were what people were buying. Their business model was predicated on selling seats to users that provided access to these features. The database underneath was critical, but incidental.

Which means the UI drove stickiness. It enforced data hygiene. It created shared vocabulary: Leads, Opportunities, Accounts. It made thousands of sales reps enter data they otherwise wouldn’t have. The UI has been the mechanism that kept the data coherent. The product is so sticky that many sales leaders insist on bringing Salesforce with them to new jobs not because the UI is good, but because it is muscle-memory.

Agents are beginning to upend this model. Instead of interacting through the UI, they can read and write directly to the underlying data, which has prompted a wave of new tools and workarounds that bypass the interface entirely (Salesforce isn’t the only example here: we recently wrote about how SAP is seeing an entire AI-friendly ecosystem proliferate around it). Computer using agents also make the traditional human level factors like preferences, training, undocumented context obsolete over time. In other words, the requirements to be a durable system of record are evolving.

The historical scorecard

Before asking what changes in an agentic world, it’s worth being precise about what drove stickiness for a system of record in the first place. The first few really focus on how humans interact with software and their preferences. Software was made sticky largely by the UI, habit, human workflow and embedded process.

How frequently is it accessed? A CRM is used every day by the GTM team and beyond. That frequency makes it critical infrastructure, and the human layer built on top of it, such as rituals, muscle memory, management cadences built over years, is often the hardest thing to migrate, because it’s not even recognized as something that needs migrating.

Is it write-only or read-write? A sticky SoR is a read-write SoR. A CRM, for example, isn’t a write-only archive; it’s being read constantly. Every call logged, every stage updated, every task created was inputted by someone (who presumably cared about what they were doing). That bidirectional flow means any replacement has to handle live operational data, not just a historical export. There’s no safe moment to cut over, which means that enterprises tend to stick with a provider once they’ve onboarded. On the flip side, an applicant tracking system (ATS) tends to be write-only, there are limited reasons to return to the data after the hire has been made.

Subscribe now

How many undocumented SOPs are there? Business-critical context doesn’t live in any wiki; it’s encoded in workflow rules built up over years by admins and system integrators. In the sales example, the undocumented context is that enterprise deals over $100K need VP approval, EMEA deals require privacy review, and strategic logo discounts can bypass finance only at quarter-end. And this context is often what makes the difference between something getting done in a timely fashion (or without violating some important practice) or not happening at all. Migrating means reverse-engineering every automation, or losing the institutional memory entirely.

Are there a lot of internal or external dependencies? The core question is how many internal systems, team processes, or outside stakeholders depend on this system of record. Internal connectivity refers to other software or workflows downstream of it. External connectivity refers to outside parties like auditors, accountants, or regulators who need direct access to the data for something like an ERP. The higher the connectivity on either dimension, the more that needs to be untangled during a migration.

How critical is the data from a compliance perspective? The core question here is simply: is this system compliance-critical? Compliance-critical systems like payroll, ERPs, and HR data require a legally defensible source of truth, strict admin access controls, and direct involvement from auditors and regulators in any migration. That makes them significantly stickier. Sales data and customer support tools like Zendesk sit at the other end: you care about continuity and context, but there’s no regulatory exposure if data moves or someone gains access.

Not all systems of record have carried the same switching cost. Score a CRM against an Applicant Tracking System (ATS) across these same dimensions and the gap is stark. An ATS is a workflow tool for a bounded process: recruiting. Once a candidate is hired or rejected, that record is largely write-once. The integrations are narrower. The user base is small and concentrated.

An ERP sits at the other extreme: the ledger is the audit trail, and your accountants, auditors, and regulators become direct stakeholders in any migration. Replacing your ATS is painful but survivable. Replacing your CRM is open-heart surgery. Replacing your ERP is open-heart surgery while the patient is running a marathon.

Traditionally, systems of record have not taken advantage of moat-creators like proprietary data or network effects; the workflow typically created enough of a moat. If anything, consumer businesses have brought together tools and a network; historically SORs have not.

Proprietary Data - While many systems of record collected customer data, they didn’t really do much with that data (and often contractually couldn’t). So while a CRM has a rich set of data and could aggregate across customers to generate cross-customer insights, they never did in a meaningful way (although there have been some attempts like Salesforce’s Einstein).

Network Effects - The holy grail would have been network effects. The CRM becomes more valuable because software sellers can find buyers. Like data, network effects have been weak at best for systems of record historically.

So if the UI disappears — and agents arrive — what’s left?

An agent doesn’t need a browser. It needs an API, context, instructions, and the ability to act. Two things made this possible at scale: LLMs became capable enough to reason. As such, an agent can now read context, form a plan, select tools, execute actions, and review output, without a human in the loop for most tasks. And MCP standardized tool access, giving agents a common interface to call external capabilities. An agent with MCP access to your platform can do what a human user does in milliseconds, at scale, without a browser. With the right context, computer-using agents should be able to navigate incumbent software interfaces without even needing APIs.

Simplistically looking at it, there are now three paths for a software buyer:

1) Incumbent system + agents. Use the incumbent’s CLI and APIs — either through their native agent product (Salesforce’s Agentforce, SAP’s Joule) or by building your own agents on top. (Suspend disbelief that APIs are complete and usable and that going headless is not as operationally complex as it is.)

2) DIY the system of record entirely. Build your own data model, operational logic, and things like permissioning, audit trails, integrations, etc. and your own agents from scratch (likely leveraging third party agent building and database tools).

3) Buy an AI-native replacement. Buy the new generation of software built from the ground up for the agentic era, designed for machine readability, with agent orchestration as a first-class feature rather than a bolt-on. This could be headless.

So, what stays from the old scorecard? The elements driven by human behavior and preferences fade away like frequency of access or read vs. read-write which are related to human muscle memory. Agents may kill muscle memory as a moat, but they do not kill operational logic and context as a moat. If anything, they make that logic more important, because agents need explicit rules, permissions, and process definitions in order to act safely.

Undocumented SOPs stay important in the short-term. The institutional logic encoded in your workflow rules is exactly what agents need to operate correctly on your behalf. It’s also the hardest thing to reconstruct. That doesn’t export cleanly, yet, especially when there are still humans involved in some part of the process. However, capturing context is becoming easier, and as agents replace more labor this becomes less relevant.

Connectivity is still hard to unwind and extends further. The connectivity factor shifts. It’s less about keeping up with humans and more about maintaining connectivity across traditionally siloed functions and software. A CRM agent needs to stitch together data and context across sales, billing and customer success. And if your platform is also the node through which agents from multiple external organizations transact buyers, sellers, partners, the dependency deepens further. An incumbent with agents is going to have a tougher time working across the primitives of various underlying software, as would a DIY database and set of agents.

Compliance-critical data remains important. Data for regulators or with regulatory or legal risk needs a single trusted source of data. A customer is less likely to switch if they trust their existing product. Take payroll and accounting data - an agent might want to access this data, you’re less likely to build and maintain this in-house. In a fully agentic world, one of the hardest unsolved problems is: which agents are authorized to do what, on whose behalf, with what auditability? A system of record that becomes the identity and permissioning layer for agent-to-agent interactions has a structural role that’s genuinely hard to displace, not because of the data it holds, but because of the trust architecture it enforces.

Going forward an increasingly relevant set of factors become important for driving defensibility for AI native startups:

How hard is it to recreate the SoR? - Data is going to matter more in a few ways. First, in the near term, the ease in extracting and recreating the data that underlies the system of record. AI is making this easy with a number of tools that enable a user to do this. In the near term, incumbents can and will make this harder by making APIs painful, gated, incomplete, or economically unattractive, if APIs are even provided at all. But as the extraction tools get better, particularly as computer using agents improve, they will make it even easier. Simultaneously, of course, new companies are recreating a richer set of data from emails, phone calls and voice agents, and internal documents. AI lowers the cost of recreating the first 80% of a system of record. The remaining 20%, which are the exceptions, approvals, compliance requirements, and edge-case workflows, is still what separates a useful wedge from a true replacement.

Is there meaningful proprietary data? - Second, the data itself becomes more interesting. The defensible data is not the data you import; it’s the data your product uniquely causes to exist. We talk about walled gardens of data - data that is either proprietary, regulated, or constantly needs to be updated. A software provider that has invested in collecting authoritative and complete data has an advantage over general-purposed providers or competitors that don’t have this data. Another vector here around data is when the data depends on internally generated actions. The best businesses won’t just warehouse data entered elsewhere. They will generate new data exhaust through being in the loop and include things like observed behavior, response rates, timing patterns, process outcomes, benchmarks, exception patterns and agent performance traces. The key thing here is that the data is the context now.

Does it own the action layer? - In the old world, storing the record was enough. In the new world, agents take action and defensibility may shift toward products that can operate in a closed loop from taking the action, to capturing, the outcome, to using that feedback to improve future decisions. For an ERP, this could be approving spend, triggering payroll, reconciling invoices, sending notices, etc. Products that close the loop are more defensible because they sit inside execution, not just observation: they generate unique data, improve with use, and become harder to remove without breaking the workflow. The value here increases of course with more context gathered and more edge cases handled.

Is there a real-world execution element? - Business models that have connectivity into real-world operations that will not be fully automated. The obvious examples are businesses with an operations network built out, like DoorDash, which historically were not systems of record but are instructive here. More broadly, any software business that closes the loop into services, fulfillment, logistics, field operations, or payments has a different kind of defensibility than pure SaaS. These companies do not just store the record or recommend an action; they dispatch people, move goods, or complete the service.

For builders, this suggests opportunity in markets where software can increasingly decide and agents can increasingly coordinate, but where the final mile still requires execution in the real world. For example, vertical software tied to field services.

Are there network effects? - Historically, network effects were weak in most systems of record because the software was primarily internal. But in an agentic world, network effects may become much more important if the system is embedded in a multi-party workflow. If the system mediates recurring interactions between multiple parties like buyers and sellers, employers and employees, companies and auditors, vendors and customers, payers and providers, then each additional participant can make the network more useful to the next.

One way is via shared workflow coordination: the product becomes the place where both sides of a process transact, exchange context, and resolve exceptions. Another is through benchmarking and intelligence: the system can surface norms, anomalies, and recommendations based on patterns observed across the network, which works together with the data point above. And a third is through trust and standardization: once counterparties begin to rely on the same rails for approvals, handoffs, compliance, or payments, the product becomes harder to displace because it is no longer just a database, but part of the coordination infrastructure for the market itself.

How technically capable is the buyer? - In a world where anyone can theoretically build their own agents, there is still a wide range in buyers’ actual ability to do so. Especially in vertical end markets and among functional buyers that have historically not had strong internal engineering resources, the odds that they will build, maintain, and continuously improve their own database, workflow logic, agent stack, and governance layer remain low. Cost matters here too: DIY may reduce software licensing in theory, but often shifts spend into implementation, maintenance, and internal complexity. This means there is real opportunity in categories where the buyer’s operations are operationally complex but technically underserved which is true of much manufacturing, construction back-office, industrial and field-service workflows, or for areas like accounting.

There are a few other important factors, which will also be table stakes for software. For example, the ontology needs to be different. A lot of “DIY database” thinking underestimates how much value lives in the object model itself. Incumbent software was built for dashboards, reports, and humans, capturing workflow. This would be opportunities, tickets, candidates, etc. Agentic schema needs to capture reasoning, actions, state tracking, exception handling, delegation, and coordination across systems. The native object model might become tasks, intents, threads, policies, or outcomes instead.

Similarly, permissioning needs to be updated for managing agents, not just humans. This includes: who can do what, through which agent, under what policy, with what approvals, with what audit trail, and with what rollback / exception handling.

And of course, all of this is in the context of cost (e.g. how much it costs to build and maintain agents/database, how much the API access costs), which again circles back to things like how hard it is to recreate the data and the number of dependencies.

So where does this leave us?

As incumbents go headless, they are making an implicit bet that the data layer will remain the source of value. In some categories, especially those that are deeply compliance-bound like financial services, that bet may hold for some time and going headless may be farther away. For the software builder, the opportunity to compete against the incumbents as they go headless and build durable software changes. The next generation of systems of record are already starting to look different: not just repositories of data collected to log human work, but agentic such that they capture the context, initiate the work and record the data exhaust. Moreover, the most interesting businesses will extend into real-world execution, coordinating the likes of field workers, logistics providers, service teams, and physical assets, or sit between multiple parties. They will be mixing the business models of the old world, and the core of the traditional system of record, the data, will be what sits in the background.

A big thank you to @astrange for her thought partnership on this!

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

When DUDE44, an American F-15, was downed over Iranian territory, the United States expended considerable resources—logistics, risk, cost—to recover the stranded Americans. Not because the math favored it in any narrow operational sense, but because as Americans we believe in “no man left behind” as a tenet.

This deeply held belief—that we are a nation that does not leave our warfighters behind—shapes our R&D agenda and our military capability stack. It shapes what war is, when we choose to fight it. The engineering requirements of “how do we bring them home” are a technical constraint for us, and they are the good kind of constraint: one that links our tools and our values to each other.

Our commitment to “no man left behind” matters a great deal in how we prepare for autonomous warfare. The moral case for autonomous warfare is not despite our values, but because of them. A democracy that refuses to replace soldiers with machines when machines can do the job is breaking its covenant with the people it sends to fight. America must now build autonomous systems at both the quality and quantity required to win. We have the talent advantage. We are losing the production race. The stakes are whether the United States arrives at the next conflict with overwhelming superiority in autonomy, or whether we cede that advantage to adversaries who are already building it.

From the Stick to the Swarm

The entire history of warfare can be read as a single, accelerating motion: extending the standoff between weapon and target. The first man to pick up a stick made a two-foot standoff. The first to throw a rock made it a hundred. The American sharpshooter at Saratoga, firing accurate rounds from cover at three hundred yards, opened a gap British line infantry could not close. The British called it cowardly when they saw the standard panic response of an army watching its officers fall to a weapon it cannot answer. But Britain fielded its own riflemen within a generation. This is how it goes.

The same charge has followed every standoff weapon since. The airplane moved the standoff to hundreds of miles. The Italians used it for reconnaissance in 1911, and by 1941 carrier aviation let Japanese pilots fly 230 miles to Pearl Harbor and bomb a fleet whose faces they would never see. The bomber over Hiroshima killed roughly 80,000 people in a flash and ended the war in the Pacific without the million American casualties projected for an invasion of the home islands. An infantryman asked to do that work one body at a time could not have done it. The technology did the work the human could not. It also brought a generation of American servicemen home alive.

The nuclear bomb extended the standoff to civilizational scale and, by being too costly to use, anchored it there. The United States built the largest arsenal in human history and used exactly two warheads. Cold War deterrence did not eliminate war. Proxy fighting continued in Korea, Vietnam, Afghanistan, and dozens of smaller theaters. But it kept the great powers from direct confrontation for eighty years. The lesson, drawn correctly by the strategists who set policy in 1945 and after, is that capability prevents catastrophe more reliably than restraint without it.

The next iteration of that machinery is already here. Cheap, swarming, AI-targeted autonomous systems will dominate the battlefield of the next century the way airpower dominated the last. Ukraine and Russia have been fighting an autonomous war for three years, both sides building, fielding, and adapting in real time. The first-person-view (FPV) drone costs $400 to $500 per unit, roughly the same as a single 60mm mortar round. For the price of an unguided munition, the modern infantry unit now fields a precision-guided one.

Operation Spiderweb, June 1, 2025, was the proof of concept at strategic scale. One hundred and seventeen FPV drones, concealed in wooden sheds bolted to the backs of trucks, driven across Russia by civilian drivers who didn’t know what they were carrying, launched remotely from inside Ukraine, guided by AI trained on museum models of Soviet aircraft. Forty-one strategic bombers destroyed or damaged. An estimated $7 billion in losses for an operation that cost in the thousands of dollars. No Ukrainian operator was within a thousand miles of the kill zone. No Ukrainian casualties.

In contrast, Russia fed infantry into Avdiivka in human waves, at casualty ratios that would end any American commander’s career, absorbed as acceptable cost because the underlying obligations are simply different. A military that treats soldiers as mass does not face the engineering problem of how to extract a casualty from a drone-saturated environment. It never asks that question, therefore it never builds that answer. The nation that goes back for its people just builds different machines than the nation that does not.

The Usual Concerns

Some make a serious case that autonomy is structurally different from prior leaps in standoff, and that the differences make it war-promoting rather than war-deterring. Three versions of that claim deserve direct rebuttal.

The first holds that autonomy lowers the threshold for democratic leaders to authorize force by removing the political cost of casualties: the flag-draped coffins coming home to Dover. The argument gets the moral framing wrong. We have never said Kevlar encourages war, or armor encourages war, or that the medevac helicopter encourages war. America must defend our interests. Casualties in that pursuit are a tragedy to be minimized. Refusing to minimize them, in the name of preserving political restraint, is not a coherent moral position.

The second holds that AI removes the psychological resistance to killing. The idea traces to S.L.A. Marshall, a World War II combat historian who claimed that most soldiers, as many as 85%, never fired their weapons in battle, held back not by fear but by a deep human resistance to killing another person. Dave Grossman later built on that finding in On Killing, arguing that modern military training works precisely by conditioning soldiers to overcome that resistance. The concern, then, is that autonomous weapons do the same thing, permanently and at scale. But the argument conflates two different resistances. Marshall’s research, contested as it is, was about the soldier on the trigger, the rifleman who hesitates at close range. The decision to go to war is made by commanders and political leaders who have never been on the trigger and have never carried that resistance. Lincoln sent Sherman through Georgia. Truman authorized Hiroshima. The wall Grossman names is the wall the trigger-puller hits. It has never been the wall that restrained the order to fire. Removing it changes who pulls triggers. It does not change who orders force.

Subscribe now

The third holds that a nation with mature autonomous capability, facing one without, has every incentive to use the asymmetric window before it closes. The historical record contradicts this. The United States held a unilateral nuclear monopoly from 1945 to 1949 and used the weapon zero additional times. Asymmetric capability has rarely produced inevitable use. What it produces is leverage, exercised through signaling and credible threat rather than detonation. The more pressing question is what happens when an adversary holds asymmetry over us. Building decisive autonomous capability is how we ensure no adversary holds that asymmetry, in the same way the nuclear arsenal ensured no adversary held nuclear asymmetry over us. The argument for building autonomy parallels the argument that has held the peace under nuclear arms for eighty years.

The moral argument for autonomy eclipses any of these weak arguments against it. In a democracy, the state asks young men and women to put their bodies between the country and its enemies. The state makes that request under a covenant: it will send them only when necessary, and when it does send them, it will give them every advantage available to bring them home. The covenant binds the elected leadership of a democracy more deeply than any other form of government. A democracy that refuses to develop technology capable of substituting machines for human soldiers in lethal roles, when that technology is available, is breaking that contract. It is choosing to send a 19-year-old to die in a place where a machine could have completed the mission.

Not just a weapon

Autonomy, however, means more than lethality. That framing, dominant in the public debate, misses the full extent of what the technology enables and why it matters.

The engineering stack that makes Spiderweb possible—autonomous navigation, sensor fusion, real-time decision-making in denied and contested environments—is the same stack required to put a machine into a hot zone to extract a wounded soldier. Not a related technology. The same technology, expressed in a different mission. Where autonomy is most visibly transforming standoff in its offensive dimension, its application as a life-saving tool is older, less discussed, and in many respects more foundational. That application did not begin with the current war. It predates it. And as automated lethality on the battlefield increases—as FPV drones make open-ground movement a death sentence—the imperative for autonomous capabilities grows in direct proportion to the threat: Casualty evacuations, logistics & resupply, reconnaissance of chemical or radiologically contaminated areas, disposal of explosive ordnance—the dull, dirty and dangerous work better performed by machine than human.

Long before the current era of drone warfare, explosive ordnance disposal units were fielding unmanned robotic systems for the same reason that animates the entire history of standoff: missions deemed too dangerous for human intervention. The calculus was unchanged. What has changed is the capability of the machine and the range of missions it can perform.

In the Bakhmut and Avdiivka corridors, vehicle exposure time in the open during daylight was measured in seconds before FPV strike. Conventional casualty extraction became effectively impossible in certain pockets of the line. Ukrainian units improvised: agricultural drones fitted with stretcher frames, and remote-controlled ground vehicles used for resupply and the recovery of human remains along routes that would otherwise kill any driver who attempted them.

Before, the alternative to these improvised autonomous solutions was not a better conventional solution. It was leaving people behind.

A case study in the Pacific

So then how does this look on the battlefield? Let’s use Operation Detachment, the U.S. invasion of Iwo Jima, February 19th, 1945, as a case study.

The Marines were tasked with seizing and securing Iwo Jima. Specifically: capture Airfield No. 1 (South Field), Airfield No. 2 (Central Field), and deny the island to Japanese forces permanently in order to eliminate Japanese early-warning radar and interceptor bases disrupting B-29 raids, establish P-51 escort fighter bases, and provide emergency landing fields for damaged bombers returning from Japan.

Ten days of pre-landing fires were requested but only three were approved, based on faulty intelligence. Using naval guns and B-24 and B-25 bombers, the Americans softened their target before landing 30,000 Marines from the 4th and 5th Divisions on February 19th, 1945. Over 36 days, more than 70,000 Marines fought at close quarters through volcanic terrain, tunnels, and fortified ridgelines. The mission was accomplished. 26,000 Americans were wounded or killed. The Japanese suffered over 21,000 killed.

Now take the same mission. The same objective. The same enemy, with the same tunnel network, the same 21,000 defenders, the same fortified ridgelines. But fight it with tomorrow’s technology.

The operation itself begins not on the beach, but weeks before anyone considers putting a Marine ashore. Persistent ISR drones, small and quiet, operating around the clock, feed a continuous stream of imagery and signals intelligence into the AI targeting system. Every tunnel entrance is located. Every hidden artillery position is identified by its acoustic signature, its heat profile, its pattern of use. The 18 kilometers of tunnel network that made Kuribayashi’s defense impenetrable in 1945 becomes a database. The AI platform simultaneously constructs and refines the strike plan, categorizing targets by type, depth, and interdependency, identifying which positions are command nodes, which are logistics hubs, and which are mutually supporting and must be struck together. It presents the commander with a prioritized, resourced, and sequenced strike plan rather than a target list. A target list tells you what to hit. A strike plan tells you what to hit first, with what, and in what order to produce the effect you need.

Pre-landing fires no longer need to be the blunt instrument of naval bombardment. At Iwo Jima in 1945, Marine commanders begged for ten days of shelling before the beach assault, but they got three, courtesy of Admiral Blandy. Even so, the limit of three days was never really the problem. Ships were firing at a Japanese tunnel network they couldn’t fully see, at defenders who simply went underground and waited out the shells. The real limiting factor was always the inability to find and fix targets faster than the enemy could adapt. An AI targeting system assigns munitions to individual positions and coordinates simultaneous engagement across the entire island. Not sequential. Not sector by sector. Everything at once, with each strike timed to prevent defenders from repositioning between impacts.

In 1945, American commanders believed the pre-landing bombardment had destroyed most of the Japanese defenders. But it had not, and Marines died for that mistake. The AI targeting system cannot make the same error. It requires confirmation of destruction before marking a target complete, and it knows the difference between a position that has been suppressed and one that has been destroyed. The hardest targets to confirm were the ones underground. But the tunnel network that made Iwo Jima so costly presents no special problem for autonomous systems. Where soldiers cannot go, micro-drones can. Systems small enough to navigate underground passages enter the tunnels directly, reaching defenders that no surface assault could touch.

And if any individual drone is shot down, the mission continues without interruption. There are no people on board to lose.

In this scenario, the beach landing—the moment that cost 5,300 Americans in the first 58 hours—does not happen. The Japanese built their entire defensive theory around one assumption: that the Americans must come ashore on the southeastern beaches. That assumption was correct in 1945. It is no longer correct. Without a beach landing, the volcanic ash terraces, the interlocking fields of fire, the months of tunnel construction oriented to cover the shoreline, all of it is a fortress that the attacker simply never has to enter.

The mission is accomplished. A small special operations element then moves to seize and secure the airfields. Not a division. Not 30,000 men across seven beach sectors. A team.

The island is denied to the enemy. The objective is identical. The human cost is not.

Kuribayashi was a brilliant commander who correctly identified the single point of American vulnerability and built his entire defense around exploiting it. Autonomous systems guided by AI don’t eliminate brilliant adversaries. They eliminate the vulnerability.

Peace Through Engineering

America does not wage war the way other nations wage war. That is not an accident. It is doctrine, built from values. We have always been willing to spend disproportionate resources to recover a single soldier, sailor, airman, or Marine, because we believe, at the level of national covenant, that the people we send to fight are not inputs to be optimized. They are citizens. And the state that sends them bears an obligation to bring them home. That belief has shaped our weapons, our tactics, our training, and our tolerance for operational cost in ways that distinguish American warfighting from our adversaries. Autonomous systems are not a departure from that tradition but its fullest expression.

For the first time in the history of warfare, the technology exists to honor that covenant completely, to accomplish the mission without sending the man. The United States must now field it at the quality and quantity that delivers a decisive advantage. Historically, the American defense establishment has struggled to achieve dominance on both of those two axes.

The first one is quality, the intelligence layer: sensor fusion, target discrimination, electronic warfare resilience, swarming behavior, on-board decision-making in degraded communications environments. This is the part of the problem the United States understands. Our AI talent pool, our frontier model labs, and our operational experience integrating advanced systems are durable advantages.

The second is quantity. Mass. The capacity to produce attritable platforms at the volume the modern battlefield actually consumes. This is the part the United States has done poorly for fifty years. Our procurement system rewards exquisite, high-cost platforms in small numbers. The F-35, by every measure of capability, is the finest fighter aircraft ever built. By every measure of cost-per-unit and units-fielded, it is a cautionary tale. China is not making the same mistake. Chinese drone production capacity, by open-source estimates, exceeds the entire NATO production capacity by an order of magnitude.

Both axes are required. Quality without quantity gives you boutique capability: beautiful systems too rare to use. Quantity without quality gives you dumb mass, attrited at the same cost ratios that destroyed Russian armor in Ukraine. The United States needs both, and the procurement system has to be restructured to deliver both. That means treating attritable autonomous platforms as a different procurement category from crewed aircraft and surface combatants. They are. They should be procured, fielded, and replaced like ammunition, not like aircraft carriers—expendable, not durable items. We have the talent advantage, but we are losing the production race.

The strategic logic for autonomy is the strategic logic the United States has faced in every major technology race of the last century. We did not build the atomic bomb because we wanted to use it. We built it because the alternative was a Reich or a USSR that built it first and used it without our restraint. The same logic applies, in sharper form, to autonomous warfare. China is racing toward decisive autonomous capability. The Chinese state operates without American legal constraints, without American norms about civilian casualties, and without American restraint about authoritarian uses of surveillance and force. A future in which the People’s Liberation Army has decisive autonomous superiority over Taiwan, the South China Sea, or American forces in the Pacific is a future in which the rules of force projection are written by Beijing.

If China moves on Taiwan, the form will not be a 1944-style amphibious assault. It will be cyber, information operations, autonomous logistics interdiction, and drone swarms over Taiwanese command nodes: the full toolkit, applied at strategic scale, designed to break Taiwanese will before any visible kinetic engagement begins. The only thing that prevents that future is a credible American autonomous capability that makes the Beijing calculation come out the other way. Recent American operations against Iranian and Venezuelan targets demonstrated autonomy at strategic effect. Beijing’s silence in response was calculation, not indifference. The window in which we hold decisive advantage is the window in which we deter. We must not waste it.

The standoff that started with a stick has reached its current form. There will be further milestones; autonomy is not the last. What will not change is the obligation: the obligation of the state to its citizens, the obligation of the leadership to the men and women it might one day send into harm’s way. We owe them our best engineering, our largest arsenal of attritable systems, and our most capable AI. We owe them Peace through Strength written into the budget, not just into the speech. The technology that keeps them home is a moral asset, not a moral hazard.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Other Income

Much has been made of the massive earnings growth in the public markets, that is large to begin with, but expected to get even larger, this year and next.

There is, however, a fun little wrinkle underneath the earnings story that’s not something you see everyday. Not all income comes from the same place. And the share of hyperscaler income attributable to “Other Income” was exceptionally high:

“Other income” was more than a third of net-income in Q1, even though historically, it’s ~5-10% (give or take).

Other Income can mean a lot of different things, but in this case, the hyperscalers (but really Amazon and Google, mostly) explicitly attributed nearly all of the gains (~$53B) to their private market investments. Alphabet’s CFO said, “Other income and expenses was $37.7B . . . primarily due to unrealized gains in our nonmarketable equity securities portfolio,” while Amazon flagged in its 10-Q its $15.6B gain (net of expenses) “from our investments in Anthropic.”

The “other income” story is an investment returns story. The hyperscalers are good at venture, one might suppose.

Putting aside, though, Google’s and/or Amazon’s recent successes in their private market tech investments, the truly staggering thing is how much everyone invests in tech—you might even say (as we are wont to do) that tech is the cycle.

KKR estimates that tech-related capex is the only kind of capex currently contributing to growth (and it’s contribution is growing)—in fact, tech capex contributed 1.9% of the 2% total GDP growth in Q1, i.e. basically all of it.

But, tech investing is bigger than just capex, and its role in the economy is bigger than its recent contributions to GDP.

By the BEA’s measure of total business capital expenses (which includes R&D and software, in addition to capex), tech is now 55% of all business investment in the US:

Tech’s share of capital expenses has been steadily climbing for quite some time, and there’s good reason to think it will continue to climb (perhaps even more quickly). As per Yardeni Research:

Before the Age of AI, economists were taught that there are only three factors of production, namely, Land, Labor, and Capital . . . Now, economists should recognize that there is a fourth factor of production, namely, Data . . . The Digital Revolution increases the incentive to create more Data (a.k.a. Information), especially now that AI tools can process so much more of it, increasing its value . . . All the data increases the demand for “compute.”

In other words, the more useful data becomes, the more we invest in it (and the tools around it), and AI has made data even more useful than before.

Good on Amazon and Google for doing VC, but the truth is that we’re all tech investors now.

The Slop Surplus

Great news, everyone: there are now so many more e-books to read, thanks to AI:

Monthly releases of Amazon e-books has tripled since ChatGPT’s release.

There are two ways to interpret this chart:

First, is the easy one: AI showed up in late 2022, the slop tsunami began, and Amazon is now drowning in machine-generated junk. By late 2025, new e-book releases were running at over 300,000 per month (roughly triple the pre-ChatGPT baseline).

Second, is the slightly more nuanced one: yes, there’s a slop tsunami, but there are still more “quality” books than before.

A new NBER paper from Imke Reimers (Cornell) and Joel Waldfogel (Minnesota) suggests that the supply increase is large enough that even with average quality dropping, the absolute number of moderately good books rose. Reimers and Waldfogel calibrate a nested logit demand model and find that the 2025 choice set delivered about 7% more consumer surplus than a human-only counterfactual would have. Not earth-shattering, but positive and rising every year. A 2023 reader was barely better off; a 2025 reader, meaningfully so.

In fact, one of the biggest beneficiaries of adding AI to the mix are incumbent authors (the ones publishing before LLMs existed). Incumbents got much more productive after 2023:

AI hasn’t just ushered in robot-authors, but it has super-charged the human authors as well.

This is roughly the prediction Marc Andreessen made on David Perell’s podcast a couple years back: “It’s now so easy to write that we are absolutely awash in bad content . . . On the other hand, these tools are now so effective that there ought to be a giant explosion of high quality content that goes right along with that.”

The slop is real, but so is the surplus. And the writers who were good before LLMs are getting more done.

Call Centers, Not Dead Yet

David George just wrote a whole thing about how the AI jobs apocalypse is a fantasy. You should read it. It’s good.

One point he makes is to distinguish between AI “substitution” v. AI “augmentation”, whereby the former category of workers are certainly at-risk, while the latter become more valuable than before. One ready example of a job in the substitution category is customer service—that makes sense: AI can handle all the Q&A, with infinite patience to boot.

Maybe so, and it does seem likely that customer service will face substantial substitution, but however logical that may sound, apparently someone forgot to tell the customer service reps:

Per Apollo, IT and business processing industry employment in the Philippines (the call center capital of the world) rose from 1.15 million in 2016 to 1.9 million in 2025—straight through every major leap in AI capability. The industry’s trade group is projecting another 70,000 jobs added in 2026 (+3.7% YOY).

Subscribe now

It’s not just the Philippines that seems relatively immune from the great customer service replacement—it’s true in the US, as well:

Demand for customer service reps has perked up, more so than the field:

Indeed’s job-posting data shows customer service jobs are not only increasing, but they’re running well-ahead of the (negative) headline figure, growing ~10pp faster YOY.

Even more striking, the flippening is fairly recent (August 2025).

Does that mean that everyone is wrong and actually AI is a massive tailwind for customer service reps? Well, probably not.

The story here is really about the relative costs of text-based LLM output v. voice. The latter is much more expensive, and it’s still too expensive to justify fully automating the function. Goldman Sachs actually ran their own internal experiment and estimated the head-to-head costs of humans v. AI call center reps, as being roughly comparable:

Goldman’s all-in estimate for an AI rep is $92/day, which is slightly more expensive than the $90/day a human costs. That stands in comparison to a coding agent, which relies entirely on text, and is orders of magnitude cheaper than a human—of course, the big difference between code and customer support, is that there is much, much more latent demand for code than customer service. Demand for SWEs has also accelerated, substantially.

The anecdata supports the notion that AI may eventually substitute-out call centers, but for now, the juice isn’t worth the squeeze (in many cases). In early 2024, Klarna announced that it had replaced 700 customer service agents with AI. The CEO said the bot was doing the work of all of them, and it was the most-cited example of “AI is replacing humans” in the service economy.

By May 2025, the CEO had reversed course, and Klarna started rehiring as service quality dropped and customers were getting generic, repetitive answers. “We focused too much on efficiency and cost,” he told Bloomberg. “The result was lower quality, and that’s not sustainable.”

This won’t last forever. API costs are dropping fast, companies like Decagon are scaling extremely quickly, and the parity number probably looks different in 18 months.

In AI, Great Products Win Fast

AI continues to go mobile at an incredibly rapid pace:

All of downloads, monetization, and time-spent inflected upwards in Q1—monetization and time-spent both nearly doubled yoy.

Maybe people are spending less time on social media because they’re spending more time vibe-coding killer stuff on their phones? That wouldn’t be so bad.

Bonus chart: per YipitData, the number of panelists using 2-5 and 6-9 AI vendors both continue to rise—at this point, less than 20% are using just one vendor.

There’s no winner-takes-all in the B2B AI market, for now, at least.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

For more news from the a16z policy team, subscribe to their Substack.

Almost everyone in American public life learned how a bill becomes a law from the same place: a three-minute animated segment on Saturday morning television. The Schoolhouse Rock version is not wrong, exactly. But it leaves out nearly everything that matters in determining whether a bill becomes a law in today’s Congress.

Members of Congress have introduced hundreds of AI bills. What is their path to becoming a law? What kinds of bills are likely to survive the legislative process?

We’re fortunate at a16z to have a whole team of people who know exactly how an AI bill becomes a law, including Ben (one of the authors of this piece) who served as floor director for House Majority Leader Steve Scalise for several years. This piece breaks down the whole process from start to finish, including:

• The funnel dynamics of proposing and passing bills (only 1.5% make it through!)

• The structural hurdles and constraints of Congress, and how they shape AI legislation

• How to evaluate the political viability of an AI proposal

• What to watch for, if you think a bill will move (or won’t)

“Don’t just stand there, do something!”

Since the public emergence of generative AI tools such as ChatGPT in late 2022, there has been a frenzy among lawmakers at both the state and federal level to regulate artificial intelligence. While the states have successfully enacted a growing web of AI laws, the federal government has been slower to act. This is not for a lack of ideas. AI regulation is one of the most discussed topics among policymakers in Washington. In just the past two years, over 100 AI-related bills have been introduced in Congress, yet only one stand-alone bill, the Take It Down Act, has been signed into law.

At first glance, this gap invites familiar explanations. Some point to industry lobbying as the primary obstacle, arguing that technology companies have successfully blocked or weakened legislative proposals. Others cite partisan gridlock or the inherent difficulty of regulating a rapidly evolving technology. These arguments are frequently raised, and each may describe a real dynamic in specific cases. But none of them accounts for the more fundamental pattern: even proposals with broad bipartisan support and no organized industry opposition struggle to advance through Congress. The barrier is not simply political. It is structural, rooted not in how the legislative process looks on paper, but in the reality of how it plays out.

The US Congress is not designed to translate policy interest into law quickly or easily. It is a system defined by multiple inflection points, fragmented jurisdiction, and procedural hurdles that make legislative success the exception rather than the rule. Any given bill must navigate committee gatekeeping, secure leadership support, survive floor consideration in both chambers, and ultimately be reconciled into a single text before it reaches the president’s desk for signature. These features are built into the design of the Constitution.

But they have particular consequences for emerging technologies like artificial intelligence, where policy cuts across numerous committee jurisdictions, lacks settled partisan alignment, and often requires comprehensive rather than incremental solutions. Even widespread agreement that “something should be done” does not translate into legislative outcomes.

The legislative filter

A sub-1.5 percent enactment rate

The modern federal legislative process operates less as a pipeline than as a filter. Consider the numbers from the 118th Congress (2023-2025): 19,297 bills or resolutions were introduced across the House and Senate. Of those, only 1,809 were reported out of at least one House or Senate committee. 1,029 bills or resolutions passed the full Senate, 995 passed the full House and, ultimately, just 274 bills cleared both chambers and were signed into law.

That is an enactment rate below 1.5 percent. At each stage of the process, proposals are winnowed down through a combination of institutional constraints and political incentives. For most legislation, these hurdles are substantial. For AI legislation, they are particularly acute. The interesting question, then, is not what bills are being introduced. It is what separates the 274 that made it through from the 19,000 that did not.

Why so many bills are introduced if most will die

It is not difficult for a bill to be introduced. Any one of the 535 members of Congress need only reduce an idea to writing, sign the upper right corner of the cover page, and submit it while the chamber is in session. This low barrier to entry helps explain the sheer volume.

It also reflects a deeper reality about how Congress operates. Most bills are not introduced with the expectation that they will become law. In the modern Congress, members often introduce what might be called “messaging bills”: legislation designed to make a political statement, generate media attention, or signal responsiveness to particular interest groups or constituencies. A messaging bill can function as little more than a press release in legislative form. It can also serve a more strategic purpose: by introducing legislation on a narrow issue, a member can stake out early ownership of a policy area and position themselves as a relevant voice in subsequent negotiations. Even proposals with little chance of passage can help establish credibility and secure a seat at the table when broader legislative efforts take shape.

More serious proposals intended to move through the legislative process typically require more preparation. Members work closely with their staff and with legislative counsel to draft statutory text, often incorporating input from committees, leadership, and external stakeholders. When done in an intentional manner, this process can take months; sometimes years. But even well-developed AI proposals face significant obstacles once they leave the introduction stage.

What do the bills that make it through have in common?

Before examining each stage where bills stall, it is worth asking what the small number of proposals that emerge from the legislative funnel tend to share. The patterns are remarkably consistent.

Bipartisan commitment. Not merely bipartisan co-sponsors on a press release, but genuine bipartisan investment in moving the bill forward. There is a meaningful difference between a member lending their name to a bill and a member spending political capital on it. Successful legislation typically involves a bipartisan coalition who have decided to co-own the issue and drive toward its enactment.

External urgency. Legislation tends to advance when the political cost of inaction exceeds the cost of acting. This can take several forms: a crisis that demands a response, a looming lapse in funding or authority, a court decision that creates regulatory uncertainty, a growing state-level patchwork that imposes compliance burdens on national industries, or international competitive pressure. In the AI context, the proliferation of state AI laws, international competitiveness, and growing public concern are generating pressure on Congress to act. Whether and when that pressure reaches the threshold necessary to move legislation is one of the central questions of the current moment.

Subscribe now

Committee and leadership buy-in. Bills generally do not move unless the relevant committee chair or ranking member wants it to. Committee chairs control the hearing calendar, the markup schedule, and whether a bill is ultimately reported favorably to the full chamber.Following a successful committee process, Leadership must then decide which bills should and should not be considered by the whole chamber. This is perhaps the most underappreciated variable in legislative outcomes.

Executive branch buy-in and shepherding. While the legislative process formally ends with the President’s signature or veto, the executive branch often shapes legislation from the outset. Through policy frameworks, public statements, and agency engagement, the White House can signal to Congress that it is seeking legislative action in a given area, helping to define the contours of the debate and catalyze movement on Capitol Hill. Sustained executive branch engagement is often essential for any major proposal to survive the process.

Connection to a legislative vehicle. A bill does not need to pass on its own if it can attach to something that must pass. This dynamic is important enough to warrant its own discussion below.

Where AI bills get stuck

The committee problem

The committee stage represents the first major bottleneck. Committees serve as gatekeepers, deciding which bills receive hearings, markups, and ultimately a vote to be reported to the full chamber. Most bills never make it past this stage.

For AI legislation, committee gatekeeping is particularly consequential because of the issue’s inherently cross-cutting nature. Artificial intelligence touches on commerce, consumer protection, national security, intellectual property, labor, and more. Jurisdiction over AI policy is fragmented across multiple committees in both the House and Senate.

Compare this to a policy area with a single committee home. Higher education legislation moves through Senate Health, Education, Labor, and Pensions and the House Education and Workforce committee. Tax legislation moves through Senate Finance and House Ways and Means. A single chair can drive the process from hearing through markup to floor. For AI, no single home exists. A comprehensive AI bill falls within the jurisdiction of several committees, each with its own priorities, expertise, and political incentives. Committees may compete for jurisdiction, decline to act, or advance conflicting approaches. Rather than producing a single, coherent legislative pathway, this structure disperses authority and makes standalone AI legislation difficult to advance.

This fragmentation is not a bug that can be fixed with better coordination. It is an accurate reflection of how broadly AI implicates existing regulatory structures. The implication is significant: comprehensive AI legislation is more likely to move through leadership-driven coordination across committees than through any single committee acting on its own.

It’s important to note that bills are not strictly required to move through a formal committee process before receiving a floor vote in either chamber, but it is relatively uncommon for sweeping policy proposals—such as a federal AI framework—to advance without some committee action. That said, AI as a topic has already been extensively debated at the committee level through hearings and markups of bills with a more limited scope.

The floor: agenda control and scarcity

Even if a bill is reported out of committee, it must still secure time on the chamber floor. This stage is where many observers, including those who follow AI policy closely, underestimate the difficulty of enactment. A bill can have strong committee support, bipartisan backing, and genuine policy merit, and still never reach a vote in either chamber. The constraints differ meaningfully between the House and the Senate, but in both, floor time is a scarce resource allocated through collective judgment, not policy merit alone.

In the House, floor access turns first on whether a bill appears capable of securing a simple majority of support for passage—218 votes when the chamber is at full occupancy and attendance. That is only part of the threshold question. An equally as important question is whether the measure is supported by a substantial majority of the majority party. House Leadership often tries to operate under what is sometimes referred to as the “Hastert Rule” after former Speaker Dennis Hastert: bills that significantly divide the majority party’s own conference or caucus are rarely brought to the floor, even if they could pass with bipartisan support. A Republican Speaker who would need to rely on 20 Republicans and 200 Democrats to pass an AI bill would be falling short of core party leadership obligations in scheduling that vote. The same dynamic applies regardless of which party holds the majority.

The Rules Committee adds another layer of control and complexity. Unless 2/3rds of the House agrees to suspend the rules and pass a bill, the Rules Committee issues a special rule governing the terms of floor consideration: how long the bill will be debated, which amendments will be allowed, and under what conditions. The majority party controls the Rules Committee, and the rule it produces can shape the outcome of floor consideration as much as the substance of the bill itself. A restrictive rule can block amendments that might fracture a delicate coalition; an open rule can expose the bill to politically difficult votes. For a cross-cutting issue like AI, where members may have competing priorities and comprehensive bills bring with them wide germaneness, the structure of the rule is itself a significant variable.

As a result, the House floor is not simply a venue for considering bills that have left committee. It is a screening mechanism for measures that are both passable and manageable within the majority party’s internal and external politics.

In the Senate, the constraints are different but no less severe. Floor time itself is often the binding constraint. The chamber’s rules permit extended debate and, absent the unanimous consent pipeline, effectively require multiple votes at a super majority threshold to advance legislation. Because any single senator can object to a unanimous consent request, and because invoking cloture consumes up to 30 hours of post-cloture debate time, Senate leadership must be highly selective about which measures receive consideration. The opportunity cost of bringing any bill to the floor is measured in terms of everything else that cannot be considered during that time. For a Senate Majority Leader with a limited number of legislative weeks, this calculus is unforgiving. A bill that would consume a week of floor time for uncertain prospects will almost always lose out to measures with clearer paths to passage or more pressing deadlines.

The filibuster compounds this dynamic. Even a bill with simple majority support in the Senate can be blocked unless 60 senators agree to end debate. For AI legislation, which lacks settled partisan alignment and where individual senators may have strong views shaped by their state’s technology sector, labor constituency, or ideological orientation, reaching the 60-vote threshold is a significant challenge. This makes the Senate the chamber where promising AI legislation is most likely to stall, not because of opposition to the substance, but because of the procedural cost of moving it.

These dynamics interact in particular ways for AI legislation. In the House, assembling a stable coalition of 218 votes requires resolving competing interests across committees and constituencies before a bill ever reaches the floor. Members from different committees may have advanced different or conflicting approaches, and reconciling those approaches for floor consideration is itself a negotiation. In the Senate, the scarcity of floor time favors proposals that are either broadly supported enough to pass quickly or important enough to justify the procedural effort.

Finally, before a bill can be sent to the president’s desk, both chambers must agree on identical text. Reconciling differences between the House and Senate requires additional negotiation, sometimes conducted through formal conference committees, but more frequently through informal bicameral negotiations of the kind that produced the FY2025 NDAA. This work is most effective when it begins before either chamber commits to a final position, and it adds yet another stage at which legislation can stall or fail entirely.

How laws frequently get made: the must-pass vehicle

Given these constraints, it is worth being direct about how much federal law actually gets made. For many policy areas, the likeliest path to enactment runs not through standalone bills but through attachment to legislation that Congress must pass.

The National Defense Authorization Act, annual appropriations bills, and budget reconciliation packages are the workhorses of enacted legislation. The NDAA alone has been enacted for 64 consecutive fiscal years. According to the Congressional Research Service, the bill is “frequently used as a vehicle for legislation under the jurisdiction of committees other than the House and Senate Committees on Armed Services.” The FY2025 NDAA, for example, spanned 794 pages and incorporated the Department of State and Intelligence Authorization Acts along with provisions touching the Departments of Justice, Homeland Security, and Veterans Affairs. These vehicles must move on a predictable schedule, and they create opportunities for provisions that might never survive the process on their own.

This is not a cynical observation. It reflects the design of the system. Congress faces collective action problems in moving standalone legislation on cross-cutting issues, and must-pass vehicles provide a mechanism for overcoming them.

The political landscape

Stakeholders and outside coalitions

Beyond formal procedure, federal legislation typically must also survive negotiation among outside groups, industry stakeholders, advocacy organizations, and political constituencies. For most major policy areas, these actors play a meaningful role in shaping legislative outcomes, as Congress often draws on outside expertise and firsthand industry knowledge to inform policy.

For artificial intelligence, this landscape is unusually broad. AI policy implicates technology companies of vastly different sizes and business models, labor groups concerned about workforce displacement, civil society organizations focused on civil rights and consumer protection, national security stakeholders, creators worried about intellectual property, and consumers. Each brings distinct and often competing priorities. Crafting federal legislation that can attract sufficient support across this landscape is inherently difficult. A provision that satisfies large technology companies on preemption may alienate state attorneys general. A transparency requirement that satisfies civil society groups may impose disproportionate compliance burdens on startups. A licensing or registration regime that benefits established incumbents may raise barriers to entry for smaller competitors.

Larger legislative packages provide a mechanism for managing these tensions. By combining multiple provisions, each addressing different stakeholder concerns, such packages allow policymakers to assemble a broader base of support than any one proposal could achieve on its own. This is one reason, beyond the procedural dynamics discussed above, that coordinated, multi-issue frameworks may be better positioned to advance than narrow, standalone proposals.

Transparency and the cost of legislating

Federal lawmaking takes place under conditions of intense and increasing public visibility. Congressional activity is subject to constant scrutiny from the press, advocacy groups, and an increasingly engaged online audience. The 24-hour news cycle, combined with real-time commentary on platforms like X, ensures that even early-stage proposals are quickly amplified and debated in full public view.

Political science research has examined how this level of transparency affects legislative outcomes. Tim Groseclose and Nolan McCarty, in their influential study “The Politics of Blame: Bargaining Before an Audience,“found that public-facing negotiations incentivize position-taking over compromise, reducing legislative efficiency. When legislators know their moves will be immediately visible to outside audiences, the incentive shifts from finding workable deals to staking out defensible positions. The result is that the same transparency designed to promote accountability can make it harder to assemble the kind of pragmatic, cross-partisan agreements that complex legislation requires.

For AI, where public interest is high and the stakes are easily dramatized, this visibility raises the political cost of every legislative choice. A vote on an AI bill is not just a policy decision; it is a public statement that can be framed, clipped, and circulated. Members must weigh not only the substance of a bill, but also how it will be received by constituents, advocacy groups, and national media.

The contrast with state legislatures is instructive. States have become increasingly active in AI policymaking, and their processes receive far less sustained national attention. Proposals can sometimes advance with fewer external pressures, even when the underlying policy questions are similarly complex. This is not an argument against transparency. It is an observation about the asymmetry it creates: Congress faces higher political costs for legislating on AI than state legislatures do, which partly explains why the states have moved faster.

What to watch for

Given these structural dynamics, how should someone following AI legislation distinguish signal from noise?

Signs a bill might move

The indicators that matter track the structural attributes described throughout this post. A bill is more likely to advance when it has bipartisan lead sponsors who are actively coordinating, not merely co-sponsoring. When it has the engagement of a committee chair who is personally invested in moving the issue. When a companion bill exists in the other chamber, with sponsors working together across the bicameral divide. When it’s being discussed for attachment to a must-pass bill. When an external forcing function is generating urgency, such as a growing state patchwork, a court decision, competitive pressure from abroad, or a salient public harm. And, most relevant now, when Administration and Congressional leadership coordinate on a single legislative push.

Signs a bill might not

Conversely, a bill introduced by a single-party sponsor, with no committee action, and no companion bill, is more likely a messaging bill than a legislative effort, regardless of how much attention it receives. A bill introduced in direct response to a news cycle, without the months of preparation that serious legislative text requires, is another signal. So is the absence of committee leadership engagement after introduction. These indicators do not make a bill irrelevant; messaging bills can shape the terms of debate and position members for future negotiations. But they should calibrate expectations about near-term outcomes.

For technology companies, startups, and others in the AI ecosystem, these indicators provide a practical framework for deciding where to focus attention and engagement. Not every bill that generates a headline warrants a response. The proposals that matter are the ones that exhibit the structural attributes associated with legislative success, and they are almost always a small fraction of the total.

So how does an AI bill become a law?

By navigating a process that is designed, by its nature, to prevent most bills from doing exactly that.

The structural constraints described in this post are real: jurisdictional fragmentation at the committee level, the scarcity of floor time in both chambers, the complexity of bicameral reconciliation, an unusually broad stakeholder landscape, and the heightened cost of legislating under public scrutiny. These are not temporary obstacles. They are features of the system.

But the current legislative environment has several features that make federal AI legislation more plausible than it has been at any prior point. The White House released a National AI Framework in March 2026, signaling executive branch interest in shaping the legislative agenda. House and Senate Committee leadership are actively scoping comprehensive AI legislation. The state-level patchwork of AI laws continues to grow and the high-stakes competition with China for AI leadership is sharpening, generating increasing pressure on Congress to establish a federal baseline. Many stakeholders, including firms like a16z, have publicly advocated for Congress to pass a federal framework to govern AI. And public demand for AI regulation remains high and bipartisan.

The question is no longer whether Congress is interested in AI legislation. It is whether the political conditions will align to move a proposal through a process that filters out more than 98% of everything that enters it. The most likely path runs through a highly coordinated, comprehensive framework of bipartisan provisions, supported by committees of jurisdiction in both chambers, and guided through the process by White House and Congressional Leadership. That may not be precisely how Schoolhouse Rock described it. But it is how an AI bill is most likely to become a law.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

The loudest voices in AI want to talk about what happens if AI goes wrong. We want to ask a different question: what if AI goes wildly, unbelievably right?

That is to say: what if AI doesn’t just reduce the cost of work, but improves the experience and NPS of working? What if, instead of making people feel more interchangeable, AI could make each of us more discoverable and ultimately more valuable? What if AI leads to opportunity becoming truly abundant?

Think back to the moments that changed the arc of your career: someone took a bet on you; an unexpected email landed in your inbox; a call from a friend led to a project that led to a company. Almost every pivotal leap forward in a person’s working life comes down to the same thing: the right opportunity finding the right person at the right time.

History suggests that human intelligence is one of the most important drivers of better life outcomes. Across decades of social science, one conclusion appears again and again: when people’s capabilities are better recognized and better developed, outcomes improve.

Every person has a comparative advantage somewhere in the economy. The problem has never been a shortage of human capability. It’s been that since most of the capability is unstructured and invisible, we simply haven’t had the infrastructure to surface it.

Enter Ethos.

Ethos is building AI-powered infrastructure for human opportunity.

Using AI voice agents, Ethos captures the knowledge, expertise, and nuance that traditional professional profiles miss, making the world’s expert intelligence machine-readable. From there, it matches people to a wide range of opportunities across the economy: market research, expert calls, fractional roles, and full-time jobs. The result is more doors, opening more often, for more people.

The company is already seeing remarkable momentum. More than 35,000 experts are joining every week, spanning the full economy: lawyers, doctors, consultants, investment bankers, electricians, plumbers, and many more. Clients include hedge funds, private equity firms, leading foundational AI labs, and enterprise consultancies. And for some members, the platform is already creating meaningful income: top earners are making more than $10,000 per month on Ethos.

Ethos celebrates human greatness in the age of AI. At a moment when a lot of the conversation around AI is defensive, James and Daniel are building from the opposite premise: AI shouldn’t make you replaceable, but irreplaceable.

They are, in their own right, exactly the kind of human greatness Ethos exists to celebrate. James cut his teeth at SoftBank helping transform companies like Arm, and has spent his career at the intersection of talent, capital, and operating. Daniel spent nearly a decade as a researcher at Google DeepMind working on the frontier of reinforcement learning and agentic systems. It’s a rare combination: a commercial operator who has seen how global firms actually source expertise, and a research scientist who can build AI that understands a person at depth.

The last great improvement in the NPS of employment was Deel, making work global. The next will be Ethos: making work abundant.

If you are an expert — in anything, from law to HVAC — you can build your Ethos profile at askethos.com.

If you are a company that needs access to human intelligence at scale, reach out to the Ethos team directly.

If you want to help build it, Ethos is hiring across engineering, product, and go-to-market.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

The AI Alarmist, “Permanent Underclass” panic isn’t a convincing story. It isn’t even a new story. It’s the “lump-of-labor” fallacy, with updated branding.

The “lump-of-labor” fallacy claims there is a fixed amount of work to be done. It assumes a zero-sum competition between existing workers, and anyone or anything that may do the same job—whether that’s other workers, machines, or in this case, AI. If there is a fixed amount of useful work that needs doing, then if AI does more, humans must do less.

The problem with that premise is that it defies everything we know about people, markets and economics. Human wants and needs are anything but fixed. Keynes famously predicted almost a century ago that automation would lead to a 15-hour work-week, but of course Keynes was wrong. He was right that automation created a “labor surplus,” but rather than just sit back and enjoy the ride, we found new and different productive endeavors to fill our time.

Of course AI will absolutely eliminate some tasks and compress some roles (and there’s some evidence that that may already be happening). The shape of the labor market will change, as it always does when a transformational technology is unlocked. But the claim that AI will produce economy-wide, permanent unemployment is unhelpful marketing, bad economics and worse history. To the contrary, productivity gains should increase demand for labor, because labor becomes more valuable.

Here is our argument why.1

“Checkmate, humans?” Come on.

We agree with the doomers—and, frankly, anyone with their eyes open—that the price of cognition is collapsing. AI is getting better and better at what, until recently, was considered the exclusive domain of the human brain.

The doomer argument goes, “If AI can do our thinking for us, then humanity’s ‘moat’ evaporates and our terminal value goes to zero.” Checkmate, humans. Apparently, we’ve done all the thinking we’re ever going to need or want, and now that AI will carry an increasingly large share of the cognitive load, humans slide into obsolescence.

Here’s the thing, though: precedent (and intuition) shows that when the cost of a powerful input falls, the economy does not politely stand still. Costs fall, quality rises, speed rises, new products become viable, and demand moves outward.2 Jevons Paradox reigns supreme. When fossil fuels first made energy cheap and plentiful, we did more than just put whalers and woodchoppers out of business; we invented plastics!

Contra-doomers, there’s every reason to expect that AI will have a similar effect. Now that AI will carry an increasingly large share of the cognitive load, humans are free to tackle even more ambitious frontiers than ever before.

If history is any guide, we can expect that technological transformation will enlarge the size of the pie.

Every “dominant economic sector” has given way to an even larger successor . . . which, in turn, has made the economy only that much larger.

Tech today is bigger than finance, railroads, or industrials ever were, but still smaller as a fraction of the economy or the market as a whole. Far from being negative-sum, productivity gains have been a positive-sum force on steroids. The net result of having delegated so much of our efforts to machines is that the economy and labor market have only gotten bigger, more diverse, and more complex.

Doomers want you to ignore the history of innovation, freeze-frame the collapsing cost of cognition, and call it the whole movie. They see task-substitution and just stop.

“We’re going to 10x our cognitive output, but rather than do more thinking, we’re going to pat our tum-tums and hit lunch early, and so is everyone else,” reflects not only a massive failure of imagination, but of basic observation. Doomers call it “realism,” but it’s just not what happens, ever!

Luddite Fails

Let’s take a look at what does actually happen, when great leaps forward in productivity ripped through the economy.

Agriculture

In the early 20th century before widespread adoption of farm mechanization, roughly a third of U.S. employment was in farming. By 2017, it was about 2 percent.

If automation caused permanent unemployment, the tractor should have broken the labor market forever. Instead, farm output almost tripled, which supported a massive increase in population—and far from being permanently unemployed, those workers flowed into previously unimagined industries, factories, stores, offices, hospitals, labs, and eventually services and software.

So, sure, you could say that technology upended the career prospects for the median farmhand, but in doing so, it unlocked a global labor (and resource) surplus, and an entirely new economy.

Electrification

Electricity tells a similar story.

Electrification did not just swap one power source for another. It replaced shafts and belts with individual motors, forced factories to reorganize around entirely new workflows, and created entirely new categories of consumer and industrial goods.

This is exactly what we expect to see during the distinct phases of technological revolutions, as documented by Carlota Perez in Technological Revolutions and Financial Capital: huge upfront investment and financial interest, huge declines in the costs of durable goods, and then a generational run for durable goods manufacturers.

It took time for electricity to work its productive magic. At the turn of the 20th century, only 5 percent of American factories used electricity to power their machines, and fewer than 10 percent of homes had electricity at all.

By 1930, electricity supplied almost 80 percent of manufacturing power, and labor productivity growth doubled for decades.

Far from destroying demand for labor, more productivity meant more manufacturing, more salespeople, more lending, and more commercial activity—not to mention the second-order effects of labor-saving devices, like washing machines and cars, both of which pulled more people into higher value endeavors than was previously possible.

As prices for cars fell, both auto production and employment exploded.

That is what a real general-purpose technology does: it reorganizes the economy and expands the frontier of useful work.

We see this again and again. Did VisiCalc and Excel doom the bookkeepers? Emphatically, no. Vastly more efficient computational technology led to an explosion of bookkeepers, and created an entire industry of FP&A.

We lost ~1M “bookkeepers” and gained ~1.5M “financial analysts.”

Those new service-sector jobs

It’s of course not always the case that task-substitution leads to job-growth in some adjacent part of the economy. Sometimes, the productivity surplus materializes as net-new job-growth in an entirely unrelated industry.

But what if AI means that some people will become fantastically wealthy, leaving the rest behind?

Well, at a minimum, those fantastically wealthy people will need to spend their money somewhere, creating whole new service industries from scratch, just like they did before:

Massive productivity gains and subsequent wealth-creation led to entirely new lines of work that may never have come to fruition without rising incomes and worker availability (even though they were technologically possible, well before the 90s). However one feels about service industries that cater to the wealthy, the net result left everyone better off, as more demand led to a massive ramp in median wages (leading to more “wealthy” people).

Ernie Tedeschi, Stripe’s in-house economist, offers a fascinating “all-in-one” example of a job disrupted, transformed, and remade with technology: travel agents.

Did technology reduce demand for travel agents? Yes, absolutely:

Travel agency payrolls are today about half of what they were at the turn of the century, almost certainly because of technology.

So, does that mean technology was a job-killer? No, again, because travel agents didn’t just end up permanently unemployed. They found work elsewhere in the economy, which overall has about the same employment:population ratio now, as it did in 2000 (when adjusted for aging).

Meanwhile, for the travel agents who did remain in the now tech-enabled industry, increased productivity meant higher wages than before:

“Average weekly earnings at travel agencies were 87% of overall average weekly earnings back in the heyday of 2000. By 2025, the ratio had reached 99%, meaning travel agency wages had outpaced the rest of the private sector over that span.”

So, even then, while it’s true that tech devastated travel agent employment, in the aggregate, working-age people are just as employed as they were before, and the remaining travel agents are doing better than ever.

Augmentation > Substitution (and the Jobs that Don’t Yet Exist)

That last point is very important, and reflects yet another way that doomers are only telling one small part of the story.

For some jobs, AI is an existential threat. True. But for others, AI is a force-multiplier—which will make those jobs that much more valuable. For every job at-risk of AI-Substitution, there are other jobs that stand to benefit:

Goldman’s estimated “AI Substitution” effects are more than balanced-out by the effects of “AI Augmentation.”

Management teams also appear to be much more focused on augmentation than substitution, for what it’s worth:

As of now, AI-as-augmentation out-mentions AI-as-substitution on earnings calls by ~8:1.

Subscribe now

While Goldman doesn’t even include them on their “augmentation” list, software engineers are probably the perfect example of an AI Augmented role.

AI is a force-multiplier for coding. Not only are git pushes skyrocketing (as are new apps and new business formation), but it appears as though demand for SWE is inflecting upwards:

Software Development jobs (both by count, and a percent of the overall job market) have been increasing since the beginning of 2025.

Is that because of AI? Truthfully, it’s probably too soon to tell, but AI most definitely augments the work of software engineering, not to mention that AI is top-of-mind for every executive at every company.

With everyone trying to figure out how to incorporate AI into their businesses, it stands to reason that there would be substantial hiring efforts underway to make that happen, making certain employees more valuable, not less:

AI-exposure seems to be driving above-trend wage-growth (which is especially true for systems design).

Those gains may be somewhat narrow for now, but it’s still so, so early. As expertise widens, so too will the opportunities. In all events, it’s not the data that the doomers want you to see.

Meanwhile, according to Lenny Rachitsky (of Lenny’s Newsletter, one of the great tech-insider communities), open PM jobs continue to climb (off their rate-driven collapse) and are now more plentiful than they’ve been since 2022:

Hiring growth in both software engineers and product managers is a concise example of why the “lump of labor” fallacy is wrong. If AI substituted thinking 1:1, then you might plausibly expect, “PMs need fewer engineers,” or you could argue “engineers need fewer PMs,” but that isn’t what we see. We see demand for both continuing to rebound, because what matters is people are getting more work done.

That’s why the doomer failure is primarily a failure of imagination. They focus on the tasks that get automated away, and ignore a new frontier of demand that will create jobs we haven’t even conceived of yet:

The majority of new jobs created since 1940 didn’t even exist in 1940. And in 2000, it was pretty easy to imagine all the travel agents that would be out of a job, but it was probably much harder to imagine an entire middle-market tech services industry built around “cloud migration,” since, of course, the cloud was more than a decade away.

What does the current scoreboard show?

Up until this point, we’ve been focused mostly on theory and precedent because both theory and precedent favor the bulls:

It’s true. With every productivity unlock, we get an increase in demand and/or a reallocation of the surplus to somewhere else in the economy. That means more jobs, including a whole bunch that will get a lot more valuable, and still more that we’ve never even heard of yet. If somehow this time it’s going to be different, the doomers have to make a much stronger case than frantic handwaving.

That “job substitution” is not a civilization-killer (but the opposite really) makes sense. Humanity, by its nature, does not get complacent. We finish one job and look for another.

But, theory and precedent aside, what does the actual data show, with respect to AI and employment? With the caveat that it’s early (for better or for worse), the weight of the data does not support the doomer claim. If anything, the data shows “no change, one way or another,” but there is also emerging data that points in the other direction: AI is more job-maker than job-taker.

First, let’s start with some academic research—this is not an exhaustive literature review, but just a sampling of recent papers:

• AI, productivity, and the workforce: evidence from corporate executives (NBER Working Paper 34984): “Taken together, these results suggest that while AI adoption has not yet led to meaningful changes in total employment, it is already beginning to reshape the allocation of tasks and occupations within firms. In particular, routine clerical and administrative activities appear more exposed to substitution, while analytical, technical, and managerial tasks are more often described as being complemented by AI.”

• Firm Data on AI (Fed Reserve Bank of Atlanta Working Paper 2026-3): “Across the four surveys, more than 90% of firms on average estimate no impact over the last three years.”3

  

• Microstructure of AI Diffusion: Evidence from firms, business functions, and worker tasks (Census Center for Economic Studies, Working Paper CES 26-25): “AI-driven employment change instances remain modest, with only about 5% of AI-using firms reporting any headcount impact—distributed nearly equally between increases (2.3% firm-weighted and 3.7% employment-weighted) and decreases (2.0% firm-weighted and 2.4% employment-weighted).”

  

• Tracking the Impact of AI on the Labor Market (Yale Budget Lab, April 16, 2026). “While anxiety over the effects of AI on today’s labor market is widespread, our data suggests it remains largely speculative. The picture of AI’s impact on the labor market that emerges from our data is one that largely reflects stability, not major disruption at an economy-wide level.”

You get the idea. The recurring refrain from the most recent research is “no change on net, but some evidence of reallocation between jobs and tasks.” In one case, the net-effects of AI implementation on hiring were positive.

There is one notable exception to the “no change” story. Researchers at Stanford, the Dallas Fed, and Census all found (to varying degrees) that entry-level roles with high “AI exposure” are increasingly difficult to find. Before anyone concludes that “AI is killing entry-level jobs,” however, it bears mentioning that these researchers also variously found an increase in entry-level roles where AI is augmentative (and an increase where AI has no impact at all).

But, even if we stipulate for a moment that AI is “killing” certain entry-level roles (as opposed to the effects of broader cyclical hiring trends, as well as “aging in place”), in the bigger scheme of things, the data is showing pretty clearly that the aggregate effects of AI on employment are basically null.4

This is probably the most succinct view of the scoreboard with respect to AI’s impact on employment:

“Still no statistically significant relationship between AI and unemployment or employment growth.”

There is, perhaps, some pull towards AI Augmented roles, and some push from AI-substitute roles:

Hiring growth appears stronger (and unemployment weaker) for “AI Augmented” Industries, while the opposite is true for industries at higher risk for “AI substitution.”

In other words, the aggregate picture is neutral, but not unchanged: some job-destruction, and some job-creation, some jobs deprecated, others now with a premium.5 At this rate, job-postings for devs will exceed the pre-pandemic level in less than two years. AI may have already single-handedly saved the SF Office market, as well.

That’s where we started: AI will definitely kill and/or compress some roles (and businesses), but it’s a mistake to think that’s the end of the story. Reorientation (and eventually growth) of the labor market—as opposed to widespread unemployment—is exactly what we should expect from a transformational technology. It’s happened before, and it will almost certainly happen again (and it looks like it’s already underway).

Knowledge Work is Just Getting Started

It sounds trite, but it’s true: this isn’t the end of knowledge work—if anything, it’s the beginning.

Automation strips out the repetitive layer and pulls human work up the stack. The reason why is simple: humans want to expand! When one layer of scarcity falls, people move up to the next one. When food gets cheaper, we spend more on housing, health, education, travel, entertainment, convenience, pets, safety, beauty, and longevity.

The same thing happens in labor markets. New work keeps appearing because human ambition does not stop, and conquering old frontiers reveals new frontiers to conquer.

New business formation is already exploding, with a pretty decent correlation to AI adoption:

New apps are hitting the app store at a 60% YOY clip:

There should be no reason to think of the modern economy as some kind of museum of yesterday’s roles. Instead, it’s a creative allocation machine, enabling new jobs, new work, new goals, and new inventions, all the time.

Much of robotics has been considered science fiction, because the computational demands in a dynamic environment were too high. But AI is bringing an entirely new robotics industry into scope:

Robotics data sets have exploded, going from tenth to first, in just two years.

There’s a universe of robotics jobs no one has ever needed, until AI unlocks that need.

To repeat, none of this means every role survives intact. The BLS expects customer service representatives and medical transcriptionists to decline, and perhaps that decline is already underway:

Some jobs will disappear, others will shrink. There will be adjustments and painful transitions, and it may take some time for productivity gains to wind their way through the economy (in fits and starts). We should be sympathetic to those changes and pour effort into making them as smooth as possible, including, among other things, with active job retraining (an initiative that a16z is proud to support).

Productivity is supposed to eliminate drudgery, and that is what it will do this time around too. But, the AI job apocalypse story only works if you assume human wants and ideas suddenly freeze at the exact moment intelligence gets cheap. That is absurd. I, for one, reject the Wall-E meme, and I don’t think I’m alone:

The macro story is not a jobless future, where we retire fat and complacent to our Netflix-scooters.

The future is cheaper intelligence, bigger markets, new firms, new industries, and higher-order human work. There is no fixed amount of work, let alone a fixed amount of cognition, and there never was. AI is not the end of work. It is the beginning of abundant intelligence.

LFG!

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

5

It’s actually not the case that the aggregate employment picture is neutral—the effects on “AI exposure” and “AI Adoption,” are basically neutral—but AI has likely been a net-jobmaker, thus far. None of Goldman’s (or the other cited research) accounts for all the hard-hat jobs that AI Capex has created:

Data center construction (and the electrification of everything) have created a massive run on the skilled-trades that’s expected to persist at least through the end of the decade.

Putting aside the longer-term effects of AI on the labor market, the sheer magnitude of AI Capex makes AI a jobmaker, for now. There’s a good argument that AI Capex is the cycle now, and outside of healthcare, it’s the only jobmaker in town.

America | Tech | Opinion | Culture | Charts

A few weeks ago, we announced the a16z Growth Engineer Fellowship—an 8-week cohort of growth engineers, AI-native GTM leaders, and tech-savvy agent builders shaping the future of company growth.

We received thousands of applications from talented people around the world. Today, we’re excited to share more about the 65 fellows we’ve selected to join the inaugural cohort who are leading growth at companies like OpenAI, Replit, Notion, Coinbase, ElevenLabs, Perplexity, and Clay.

Many of these fellows are household names in the growth × AI space who have helped to write the original growth playbooks—like Luke Harries from ElevenLabs, Raman Malik from Perplexity, Ben Shanken from Coinbase, and Drew Bredvick from Vercel.

Some are AI-native engineers now harnessing their technical acumen to build massive growth engines at scale—like Aaron Lu at Notion, Ramya Durvasula at Partiful, and Vaibhav Aggarwal at OpenAI.

Others are translating a depth in product strategy to design powerful new growth systems—like Alkarim Nasser at Google, Ravish Agrawal at Gamma, Emily Gavrilenko at Cursor, and Ajay Prakash at Superpower.

You might recognize Yash Tekriwal from his educational Clay content—or Jordan Crawford from his thoughtful write-ups on growth engineering via Claude Code.

But we’ve also got young hustlers like Diana Lim who are fresh out of college, trained in computer science, and already tastefully automating everything in sight at early-stage startups.

Across the board, these fellows are architecting growth engines behind iconic tech products, engineering custom loops most people wouldn’t believe are possible, and operating at the bleeding edge of what growth can be.

We hope that you’ll join us in celebrating everything this group has already done, and everything they’re about to do next.

Congratulations and welcome, a16z Growth Engineer Fellows!

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

When you fill a prescription, the inventory system that tracked that drug from manufacturer to pharmacy was almost certainly running on the German software giant SAP. Most of the food in your kitchen passed through one too. So did the parts in whatever you drove this morning; and so did the electricity in the building you’re sitting in. About three-quarters of the Fortune 500 run their core operations on SAP. SAP is a huge and incredibly important company: and it dominates the industry that keeps much of the world’s most important software running.

That industry is system integration. Accenture, Deloitte, IBM, Infosys, and TCS dominate the top of the market, and tens of thousands of smaller firms work in their wake. Their work is the connective tissue of enterprise IT—not only the business process redesigns that follow when a company adopts a new platform of any kind, but also the migrations and integrations between systems that were never really designed to be compatible.

Every Fortune 500 firm has hundreds or thousands of these consultants embedded in its operations. Software licenses themselves add up to roughly $300 billion a year globally; the work of making that software actually function inside large companies costs several times more. Total IT services spend is around $1.8 trillion annually, and integration of enterprise applications accounts for a substantial chunk of it.

This is one of the largest pieces of enterprise technology that has barely changed in twenty years. The delivery model the global SI industry uses today is essentially the one its predecessors built when SAP itself was a startup. The largest SAP migrations at Fortune 500 companies take three to five years and cost between $100 million and $500 million, and roughly 70 percent of large transformation programs fail to deliver what they set out to accomplish.

The reason this work has resisted automation for so long is that it is genuinely difficult.

A typical SAP environment contains data from dozens of overlapping systems, many of them with inconsistent schemas and missing documentation, in which basic concepts like “customer” or “order” mean entirely different things. And they run millions of lines of custom code – much of it written by people who left the company long before the current CIO was hired.

To make this concrete: picture a global manufacturer that has run on SAP ECC for fifteen years. Their procurement system has been heavily customized over time to handle a specific contract structure with their Tier 1 suppliers, encoded in tens of thousands of lines of ABAP that nobody currently at the company fully understands. Revenue gets booked through a custom recognition workflow that was built for a regulatory environment which has since changed twice. Upgrading this company is its own multi-disciplinary project. Every custom flow has to be analyzed and classified. The new environment then has to behave correctly across every edge case the old one handled. Audit has to pass. The cutover has to be executed on a weekend that cannot run long, because the factories cannot stop. Roughly 10,000 ECC environments globally face some version of this problem, and most of the work has to happen in the next three years.

Subscribe now

For decades only humans could deal with this kind of mess. Software wasn’t capable of the judgment, the contextual reasoning, or the code comprehension the work demands, so a global industry grew up to do it manually. The economics worked out beautifully for the firms at the top and badly for almost everyone else. CIOs know perfectly well that their core systems are aging and expensive to maintain. They also know SAP is ending standard support for ECC in 2027, which puts an enormous amount of upgrade pressure on their desks right now. But the alternative to staying on legacy is a multi-year, multi-hundred-million-dollar program with a failure rate north of two-thirds, so most companies do the rational thing. They wait, patch what they can, pay the extended maintenance fees, and hope the problem can be deferred for another year.

I started Tessera because the technology to do this differently has only just arrived, and because the window to build a new kind of company in this space is fairly narrow. I came at the problem from research rather than consulting. What struck me looking at the SI industry from outside was how unusually well its core work matches the strengths of modern AI. The outputs have structure, success criteria are reasonably well-defined, there’s an enormous body of prior code and documentation to train against, and the feedback loops are clean. The parts of this work that have been hardest for humans for thirty years are exactly the parts AI has just gotten good at.

Until very recently it wasn’t. Brittle scripts could shuffle data between identical schemas, but they could not read 20-year-old ABAP and figure out what business problem it had been written to solve. That changed in the last two years, more dramatically than people outside this field tend to appreciate. Modern frontier models can read custom enterprise code, including the obscure dialects like ABAP that most of the world’s mission-critical software is written in. Mapping fields across heterogeneous schemas, which is most of what data migration actually consists of, has become newly tractable. Business logic can be inferred from observed system behavior when documentation is missing, which is the usual situation. The long tail of judgment-heavy structured work, that historically required teams of $300-an-hour consultants is now within reach of a well-orchestrated agent system, at quality that holds up to the same kind of review human work gets.

This is the technical change that makes a different kind of company possible, and Tessera is what comes after it. We are an AI enterprise transformation platform. Software does the work humans used to do, and our people focus on the parts of transformation that genuinely benefit from human judgment: architectural decisions, stakeholder alignment, and the strategic question of how a business should actually run on the other side of a migration.

Our first focus is the type of nitty-gritty ERP work in which SAP has excelled for decades. Our agents ingest a customer’s ECC environment line by line and field by field, producing a comprehensive map of what is there and how it behaves. From that, they classify each component as keep, replace, or retire, and then they build the actual S/4HANA implementation: writing the new code, generating the test suite, validating the data movements, and orchestrating the cutover itself. A six-person team at Tessera can now deliver work that would historically have required sixty consultants, working over two full years.

There is a great deal of activity in the broader space of “AI for services” right now, and I want to be precise about how Tessera differs from most of it. The dominant pattern in the market today is a kind of “autocomplete for consultants”: tools that help an existing human bill faster, while still leaving the underlying delivery model intact. We are doing something more structural. We are building software that does the work, and we have organized the company around that economic reality.

A reasonable question is why the firms that already dominate this market, like Accenture and Deloitte, aren’t the ones doing this. The answer is mostly the standard Innovator’s Dilemma. Their entire business model rests on billable hours. Doing the same work in software, with outcome-based pricing, would compress their revenue per migration by something like an order of magnitude in the scenarios where it works best. Leadership at any of those firms can see clearly that the future of this market is software. The firm itself cannot get there, because everything about how those companies hire, train, sell, and account for revenue is calibrated to a fundamentally different shape of business.

The longer-term vision goes well beyond SAP. Every large company runs on a constellation of enterprise systems: ERPs, CRMs, HRIS platforms, supply chain tools, financial reporting stacks, all of which need continuous integration, modernization, and migration as the business itself evolves. Almost all of that work can become software. When it does, the cost of running enterprise IT falls by an order of magnitude, freeing up enormous amounts of capital and management attention currently absorbed by maintenance and modernization projects. The speed at which large companies can adapt to new circumstances also accelerates dramatically. When a Fortune 500 CFO wants to consolidate two divisions onto a single financial reporting stack today, the project takes 18 months and a small army of consultants. In the world Tessera is building, it takes weeks. Most of the changes that currently get framed as “system integration projects” — territory reorganizations, post-acquisition integrations, subsidiary spin-outs, compliance overhauls — become things companies just do, rather than year-long programs they have to plan around.

The companies that adopt this model first will have a structural advantage over the ones that don’t, and that gap will open faster than most people expect.

This is the largest market in enterprise software that the AI shift hasn’t touched yet. Every business of meaningful size depends on these systems, and the friction of modernizing them is a real drag on productivity for the companies and the economies that depend on them. We are going to take that friction out.

If you are an engineer, a researcher, an enterprise software veteran, or anyone who has looked at a $1.8 trillion industry still operating on a delivery model from another era and thought there has to be a better way, we would love to hear from you.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

They say you should invest in big markets.

It’s hard to think of a larger market than the one for SIs (system integrators). Large companies can have 50+ ERPs across their many divisions, and spend hundreds of millions each annually on ERP transformation (as well as other systems of record like their CRM and HRIS). There is a vast amount of services spend both internally and externally on consultants and SIs to do migrations and maintenance - companies spend over $500 billion dollars annually on this.

They say you should invest in great people.

Kabir is a whiz kid. After starting college at 13, he graduated from UC San Diego with a CS degree at 17, earned his PhD there at 20 (where his advisor described him as “the strongest research student I have ever advised/mentored in my career” with an “unprecedented … slate of research accolades”), and went on to do research at Meta before founding Tessera. In the 18 months since, he’s closed multimillion-dollar ACV contracts with large enterprise CIOs and recruited a team of more than 30 people (including six people from Netflix, where he was an intern). Kabir’s slope has been one of one.

When you find a great person building in a big market, partnering is a no brainer.

That is what we found in Tessera, where Kabir is building an AI-native system integrator for enterprise transformations. ERP transformations are a massive pain point for enterprise IT teams, often costing companies millions of dollars in consulting costs to analyze custom code, map and validate data, redesign business processes, test integrations, and manage cutovers. Tessera is initially focused on SAP upgrades, including ECC to S/4HANA, but the broader vision is far bigger: to automate the management and modernization of enterprise software systems end to end.

We’re thrilled to partner with Tessera Labs for their Series A.

If you are a talented engineer or operator and interested in deploying frontier intelligence into one of the world’s largest markets at one of its fastest-growing startups, you should consider joining them.

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

For more news from the a16z crypto team, subscribe to their Substack.

Crypto cycles tend to follow a pattern. A wave of speculation pulls in attention and capital. Some of it gets wasted. Some of it funds infrastructure that wouldn’t otherwise get built. When the noise dies down, what’s left is usually more useful than it looked at the peak, and more durable than it looked at the trough.

You see this in every cycle if you look past prices: what actually gets built, and what people keep using when the hype fades. We’re at one of those quieter moments now. And the signal coming through is one of the most encouraging it has been in years.

The clearest evidence is stablecoins. Trading volumes go up and down with the market, but stablecoin usage has kept climbing even through downturns. People are using them to save, to send money across borders, and to pay for things, often exposing just how slow, expensive, and unreliable the alternatives are. Their growth looks less like speculation and more like network adoption: usage compounds because the technology is useful, not because of expectations about price action.

Blockchains are also proving their worth in capital markets. Since the last cycle we’ve seen meaningful growth in perpetual futures for price discovery, prediction markets for surfacing truth, and onchain lending for stablecoin credit markets. Traditional assets are starting to move onchain, and onchain finance is being used for assets beyond network tokens. A new financial system is taking shape that runs continuously, settles nearly instantly, costs almost nothing, and is open to anyone with internet access.

Regulation is moving in the right direction too. The GENIUS Act is a good example of what thoughtful policy can look like: clear definitions, strong safeguards, and room for builders to build. We expect more regulatory progress for the rest of the crypto market through legislation and rulemaking. This gives consumers protection, builders certainty, and mainstream institutions a path to participate.

It’s worth stepping back and asking why this matters now in particular.

Software is getting more complex and harder to trust. AI systems are powerful and largely opaque. The infrastructure the internet runs on is more consolidated than ever. In that environment, the properties that crypto networks were designed to provide become more valuable, not less:

• Systems that are transparent and verifiable

• Networks that are global from day one

• Economic models that align users, creators, developers, and operators

• Infrastructure that doesn’t depend on a small number of intermediaries.

These properties are showing up in real products: in payments, in financial services, in creator platforms, in decentralized infrastructure, in new ways for people and machines to coordinate. Much of this is being built by startups, and increasingly adopted by financial institutions, tech companies, and others to deliver faster, cheaper, and more reliable services.

In practice, this means sending money globally in an instant, holding dollars without relying on a bank, tokenizing assets so they can move frictionlessly anywhere, tapping into composable networks that others can build on, and embedding these capabilities in applications everywhere. It also includes new models that weren’t possible before: where users can own their assets and identities directly, and hold inviolable digital property rights; where swarms of software agents can decide, act, and transact on a user’s behalf, acquiring compute, data, and services as they go; and where increasingly autonomous networks can fund, govern, and evolve themselves through code.

That’s why we’re announcing our new Crypto Fund 5; it is built for this moment. The founders we’re backing with this $2.2 billion fund are working on the part of the cycle that gets less attention and we believe produces more of the lasting value: turning new infrastructure into products people use every day. That is how every important computing platform has eventually mattered, and it is how crypto will too.

a16z crypto

your guide to the next internet from the team at a16z crypto

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

The Fastest V-Shaped Recovery Ever

S&P 500 sold-off by ~10% during the Iran conflict, and just 11 trading sessions later, it made a full recovery.

It constituted the fastest V-shaped recovery ever on record.

Last Friday, the S&P 500 closed at an ATH 7,165.08, up 30% yoy and almost 9% over the past month alone.

For situation monitors, the whiplash is a thing to behold. For everyone else, they may not have even noticed their 401(k)s took a header in the first place.

It could just be a coincidence, but either way, these fresh new market highs seem to have revived an old debate: are stocks too expensive? This comes only weeks after wondering ‘Is Tech Cheap?’ What a world.

Well, two of the most successful investors of the past fifty years recently weighed-in, and for their part, they’re inclined to say “yes, stocks are too expensive [for them, at least]”

In a CNBC interview during the March correction, Warren Buffett was asked whether stocks looked cheap. “No,” he replied. “Three times since I’ve taken over Berkshire, [the market has] gone down more than 50%... this is nothing. If they’re 5-6% cheaper, we aren’t in it to make 5% or 6%.”

Buffett is certainly very busy not-putting his money where his mouth is.

Berkshire is sitting on roughly $373 billion in cash – the largest pile in its history – still waiting for that rainy day to go bargain-hunting.

The second hall-of-famer to cast some doubt on stock prices is Paul Tudor Jones. In an interview with Patrick O’Shaughnessy, PTJ pointed out just how high stock prices have become, in relation to the economy writ-large:

“We’re at 252% of stock market cap to GDP. In 1929, we were 65%. In 1987 we got to 85-90%. In 2000, 170% . . . Let’s say we mean revert to the past 25 or 30-year P/E. That would be a 30, 35% decline. Well, 35% on 250% of GDP is 80, 90% of GDP.

10% of our tax revenues are capital gains, they go to zero. So you can see the budget deficit blowing up. You can see the bond market getting smoked. You can see this kind of negative self-reinforcing effect.”

Sounds bad.

Now, what would cause that mean-reversion to transpire? Neither PTJ or Buffett can say for sure, but they’ve been around long enough to take the possibility seriously. It should be noted that the S&P 500 is up more than 10% since Buffett’s interview, for those keeping score at home.

Far be it for us to question the wisdom of those two giants—and mean reversion can be a helluva drug—but there may be some reasons why the market is structurally different now (i.e. we’re going to question it, anyways).

One reason that prices may be structurally higher is that there is just a lot more cash around now than there was before.

Just to take an overly simplistic view of things, despite all the money currently at-work in equities, the balance of money market funds continues to rise:

MMF balances are at record levels of ~$8 trillion, nearly double pre-pandemic levels.

More cash for MMF, and more cash for equities, and more cash for everyone, make stocks go brrr? Could be.

Fed Chair nominee Kevin Warsh made a similar call-out in his confirmation hearings last week. The Fed’s balance sheet doubled during COVID in 2 years—from $4.2 trillion to a $9 trillion peak. Even after three years of quantitative tightening, it sits at $6.7 trillion today, and roughly 30% of all U.S. dollars in M2 existence today were created in the last five years. That’s a lot of fresh cash.

If you view the S&P 500 as a price-ratio to dollars, the shrinking value of the denominator may be doing more work than people realize.

“The big balance sheet has become an ordinary, recurring force that, I think, has been quite unhelpful,” Warsh told the Senate Banking Committee. He called it “fiscal policy in disguise,” and said the Fed is “not blameless” in the widened wealth gap and K-shaped economy that has emerged.

A second possibility, of course, is what we’ve alluded to before: the reason that stock prices are growing is because earnings are growing and margins are historically wide (and getting wider).

Generally speaking, more profits at fatter margins is a well-regarded basis for assigning higher multiples.

PTJ is certainly right that multiples are historically high and that the value of the stock market (relative to GDP) is too . . . but if you’re going to turn the clock back 30 years for a benchmark of what those values ought to be, then it bears mentioning that profit margins are roughly double what they were in the 1990s, and forecast earnings growth is more than double the long-term historical average.

Subscribe now

Now, perhaps those forecasts are illusory, but still, “the prices are too high because they’re higher than before,” is on its own, a strange objection.

Depending on where you look, you might even say that in some cases prices are relatively low:

On a growth-adjusted basis, software multiples have already mean-reverted to the previous decade’s average.

“This time it’s different,” are, of course, the most famous last words, and when GOATs speak, we listen . . . but certain things are, in fact, different. For now, at least.

First Semis, Then Software

While on the subject of the stock market, one thing we’ve observed before is that while tech is driving a lot of the growth, it’s really semis (and AI infrastructure) specifically that’s driving the lion’s share of returns. Software, for its part, is still scuffling, relatively speaking (again, as we’ve observed many times before).

The out-performance of the AI buildout v. the more subdued performance of rest-of-tech raises an interesting question: how can AI infrastructure trade at a substantial premium, while the putative beneficiaries of AI trade at little-to-no premium, at all? If all this capex is going to be justified, there will need to be tangible benefits that flow through to the token-buying customers . . . but as of now, few-to-none of those benefits appear to be priced-in.

Without going down a rabbit hole of the various possibilities that might explain that disconnect, for present purposes, the point we want to make is that it has happened before: semis went first, and then software did eventually take the cake.

The post-GFC market recovery was also a “semis and infrastructure, first, software and services, second” kind of thing:

During the mobile-tech cycle, it took about 5 years for software to emerge as the clear market leader. Before then, it was semis and infra (i.e. the enabling capex) that led the way.

More precisely:

• Semis like Qualcomm and ARM captured the picks-and-shovels returns from 2010 to 2012 because every smartphone needed chips before any apps could exist;

• the platform layer (Samsung and Apple) came next by selling all those smartphones; and then

• software and services came last, and won biggest—Google and Amazon (and the “app layer” more generally) pulled ahead, as value migrated up the stack to where the applications (now running on all that hardware) lived and breathed.

Is the AI cycle following the same semis-then-software trajectory? It’s not for us to say, but there are certainly some similarities.

First, of course, is the semis-led charge, although this time, instead of Qualcomm leading the way, it’s Nvidia.

Second, is perhaps the acceleration in cloud, one layer up in the stack, where all those chips are being put to work:

Reported revenue backlogs for the hyperscalers nearly doubled yoy in ‘Q4. That’s not supposed to happen to already fully-scaled businesses.

Third is . . . well, the third is where it gets interesting.

The previous semi-led cycle’s biggest winners came from the application layer—Uber, Airbnb, Instagram, WhatsApp, etc.—and enterprise SaaS, as well. They were preceded by all the hardware because without all the hardware, none of the software would have been possible at that scale. It did take ~5 years, though, from the beginning of the semis-charge to when software really took off.

In the AI cycle, we know that the application layer is, in fact, scaling—indeed, scaling faster than ever before, albeit mostly on the private side. Speed-to-revenue for AI companies is unprecedented.

Whether and when the application layer takes off in the public markets—and brings software back to the fore—only time will tell. But this isn’t the first time that atoms have led the bits.

Peak Social Media

Here’s another first.

Meta’s daily active users declined for the first time, well, not “ever,” but the first time since 2021, which makes this the second time ever.

3.5B+ daily users is still a lot, but it’s slightly less than before.

It turns out, however, that “less time on Facebook” may be part of a broader theme.

Time spent on social media globally peaked in mid-2022 and has been falling since.

According to GWI’s tracking of 250,000 adults across 50+ countries, daily social media use across the developed world fell by about 10% from a peak of ~2.5 hours in 2022 to ~2.3 hours.

The decline is sharpest among the demographic most associated with platform addiction: 16-to-24-year-olds. If you’re in camp “social media and the phones are bad for the youngs,” then this is good news.

There is one notable exception, however:

North American social media consumption is still rising. By 2024, it had climbed to roughly 15% more than European levels and is still rising.

Why do North Americans appear to be uniquely addicted to social media? Your guess is as good as ours, but it will be interesting to see how AI plays into this dynamic. AI-generated content is exploding all over the internet, and social media is no exception. Whether standalone AI-content apps for image/video generation, or the tools that platforms are rolling out for creators, AI is increasingly in the creator’s seat (or perhaps the co-pilot’s seat).

YouTube CEO Neal Mohan wrote in his 2026 letter that more than 1 million channels used YouTube’s AI creation tools daily in December, and 9 of the top 100 fastest-growing YouTube channels globally are entirely AI-generated.

Maybe consumers will get tired of it . . . or maybe social media’s hold on our attention spans is just getting started.

Your SaaSflation Is My Opportunity

Here is yet another observation related to the current software malaise: the spillover into the credit markets.

For a decade, software was considered a fantastic thing to lend against. SaaS companies had predictable cash flows and sticky, recurring revenue, which are exactly the kinds of things that lenders like to see. The private equity industry built an entire playbook around it: buy mature software businesses, lever them up, and let the revenue service the loans. Lenders, for their part, were happy to harvest the yield.

But with the big compression in software multiples, that playbook has broken down. It’s had some big implications for private credit, and the publicly traded Business Development Companies that specialize in this sort of thing.

Pricing on software-backed debt has plummeted, relative to comparable credits from other sectors:

Average bids on software loans have dropped to ~90 cents on the dollar, and software now comprises ~35% of distressed loans (usually defined as loans trading with a ~20%+ discount).

Software gets less valuable, so too do software-backed loans. Makes sense, right?

Well, sort of. It certainly makes sense that the equity would reprice, but equity is not debt. And from an actual credit perspective, the loans are still performing about as well as they were before:

Default rates in either of the syndicated loan market or private credit both look no worse for the wear.

Now, present default rates aren’t necessarily predictive of future default rates, but the point still stands: software debt is selling off, even though software lenders are still getting paid, and that’s a bit odd. Just because the value of all future cashflows of a company may be worth less than they were before, doesn’t mean those cashflows are so impaired as to make the debt something less than money-good.

So, what’s going on?

Well, there are, of course, anecdata to make lenders nervous, like this one: Thoma Bravo recently forfeited its SaaS company Medallia back to its creditors. Thoma Bravo bought Medallia in 2021 for $6.4 billion, financing the deal with about $3 billion in debt. As interest rates rose, however, Medallia’s annual debt service roughly tripled from $100 million to $300 million. Fast-forward to the present, and Blackstone and other lenders refused to extend a lifeline, so Thoma Bravo decided to write down its $5.1 billion in equity and hand over the keys.

Medallia’s core business is a feedback-and-survey product, which is exactly the sort of thing in the crosshairs of AI-disruption. A big discount on the equity probably makes sense. But will disruption come so quickly and so intensely such that the company will lack the cashflows to repay the debt, as well? Perhaps, but perhaps not, and what’s true of Medallia isn’t necessarily true of the entire universe of software-backed loans.

But, that hasn’t prevented the sector-wide pall. Asset prices have dropped, fundamentals be-damned. And for the credit-pickers willing to be discerning, that means it’s “go time.”

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.

America | Tech | Opinion | Culture | Charts

Today we’re introducing the a16z Design Engineer Fellowship - an 8-week cohort of AI-native design leaders at the forefront of how software will be designed and built in the future.

We’re looking for cracked designers who have absorbed engineering into their craft and are thinking in systems, not screens. People who are spending their weeks vibe-coding, shipping new tools, and creating new AI processes and workflows that scale across the org.

Right now, the best practitioners are scattered across companies. This fellowship is our way of bringing everyone together to demo their latest work and level up.

If this sounds like you, apply here by Friday, May 22, 2026.

Applications will be accepted on a rolling basis, and the fellowship will begin in June 2026.

Apply Now

How it works

• 8 weeks long

• Starts in June 2026

• Time commitment is flexible, all opt-in

What you’ll get

• Connect with AI-native design engineers who are redefining how software gets built

• Dinner series in SF (and more cities, by demand)

• Demos, feedback, and live working sessions

Who you’ll be joining

You’ll break bread with founding fellows and industry leaders like:

• Ian Silber - Head of Design at OpenAI

• Benji Taylor - Head of Design at xAI

• Yuliya Gorlovetsky - Head of Product Design at Stripe

• Joshua To - VP of Product Design, AR, AI & Wearables at Meta

• Dana Cho - VP of Design at Pinterest

• Ben Blumenrose - Co-founder and Managing Partner at Designer Fund

• David Hoang - VP of Design & AI at Atlassian

• Ben South - Co-founder/CEO of Variant, Prev VP at Postmates

• David Webster - Head of UX at Google Labs

• John Phamous - Design Engineer at Vercel

• Diego Zaks - VP of Design at Ramp

• Ammaar Reshi - Product & Design Lead at Google DeepMind

• Seth Weisfeld - Co-founder and Chief Design Officer at Function Health

• Jenny Wen - Design at Anthropic, prev Dir of Product Design at Figma

...and many more.

Apply Now

AI is revolutionizing design

We’re seeing three big shifts in this space.

1. Design is going from output-focused to systems-focused.

Teams aren’t making assets anymore. They’re building systems that generate assets continuously. Brands aren’t campaigns anymore. They’re living, dynamic systems. The future about designing how things get designed.

2. There is a new 80/20 rule for design.

Vibe coding gets you to 80% fast, but what matters most (and what separates “good” from “great” design) is how you close the last 20%. This is where judgment, systems thinking, and craft become the whole game.

AI is commoditizing execution - everyone can make stuff now. So, the defensible value is shifting from exploration to context and taste-making - choosing what meets (or exceeds) the bar, what’s on-brand, and what’s right for your company in a world of infinite output. A distinct POV and refined judgment are becoming the real differentiators.

3. Finally, the roles themselves are collapsing.

Traditional design and engineering roles are merging into new hybrid roles that combine taste, systems thinking, and technical fluency. Proprietary workflows - how you combine tools, structure prompts, build repeatable systems - are becoming the new craft.

Right now, the practitioners defining what design engineering actually means are scattered across companies. We’re launching this fellowship to connect and bring everyone into one place.

We’re looking for fellows who:

• Are closing the gap between design and engineering - whether they came through Figma and learned to build in Cursor, or came through React and developed real product sense

• Ship real products, not just prototypes - and understand the gap between the two

• Hold both truths: AI changed everything and broke things along the way - the security vulns, the unmaintainable codebases, the distance between demo and production, and more

• Bring taste, systems thinking, and technical fluency to everything they build

You might call yourself a design engineer, a designer who codes, a frontend engineer with product sense, or something else entirely. Titles don’t matter - what does is that you’re building at the intersection of taste and craft, and the work speaks for itself.

If you’re interested, apply here. We’re accepting applications on a rolling basis.

Questions? Contact Katie Kirsch. Thank you!

Apply Now

Subscribe now

This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.