America | Tech | Opinion | Culture | Charts

A billion machines will become autonomous or intelligent over the next ten years. Cars, trucks, tractors, mining haulers, defense systems, warehouse robots, humanoids—the physical economy will be rebuilt around software that perceives, decides, and acts.

The prevailing assumption about how we get there goes something like this: models keep improving, world models mature, foundation models for robotics arrive, and autonomy falls out the other end. Intelligence is the whole game; scale the intelligence and the machines will follow.

I’ve spent a decade helping Applied Intuition build the software infrastructure behind many of the world’s most ambitious physical AI programs, from software-defined vehicles and autonomous trucks to construction equipment, mining systems, defense platforms, and robotics. That vantage point has given me a front-row seat to where the industry is accelerating—and where it continues to slow itself down. I’m as bullish on the intelligence as anyone, but the prevailing assumption gets the math wrong. Deployed physical AI is a product of two variables: the capability of the models, and the capacity of the engineering system around them—i.e. how requirements become software, how software gets validated, and how validated systems get deployed, monitored, and improved. The industry has largely poured everything into the first variable while the second sits roughly where it was a decade ago, built for quarterly releases and hundred-person integration teams. Frontier intelligence running on a legacy engineering system doesn’t produce frontier outcomes, because the old engineering system is the limiting factor.

The contrarian bet, then, isn’t against intelligence. It’s that the next order of magnitude in physical AI comes from making the engineering system as intelligent as the models it carries. The industry’s roadmap, however, rests on a handful of assumptions that once made sense but no longer match where physical AI is headed.

Smarter Models Don’t Create Deployed Machines

The gap between a capable model and a certified, operating machine is enormous, and model quality alone doesn’t close it. A model that’s 20% better in benchmark terms still has to be integrated with many other software components, tested across millions of scenario variations, traced against safety requirements, validated on hardware, rolled out to a fleet, and monitored in the field. In a typical program, that pipeline — not the model — sets the tempo. Teams take delivery of a meaningfully better model and then spend two quarters proving it’s safe to ship.

This is why world models, as remarkable as they are, won’t get us to a fully autonomous future on their own. They are advancing faster than engineering organizations can keep up. Every leap in model capability relocates the bottleneck rather than eliminating it. The constraint moves downstream, from “can the machine perceive the world?” to “can we validate, integrate, and operate what the machine can now do?” A team whose validation cycle takes months is, in effect, throttling frontier AI down to the speed of its own process.

Here’s the implication the industry hasn’t priced in: as models commoditize toward the frontier, two companies with access to the same intelligence will have wildly different outcomes. The difference will be determined by how fast their engineering systems can absorb what the models can do. Intelligence is becoming ubiquitous. The ability to operationalize it isn’t.

Digital AI Doesn’t Transfer to Physical AI

The second assumption is subtler: that the agentic revolution happening in digital work will naturally extend to physical AI. Direct the coding agents and copilots at the autonomy stack, and the same productivity gains will follow.

They won’t, because most digital AI stops at documents, conversations, and code. Physical AI work doesn’t live there. It lives in drive logs and sensor data, in simulation runs and hardware-in-the-loop test rigs, and in requirements databases and validation reports. It lives in fleet telemetry streaming from real vehicles on real roads and job sites. An agent that has never seen a disengagement, doesn’t know why a perception regression matters, and can’t trace a requirement to a test case, is not a productivity tool in this domain. It’s a liability with a friendly interface.

Making agents genuinely capable in deploying physical AI itself is a frontier intelligence problem, one that is entirely different than training a bigger model. The agents need access to the actual data and tools of the trade—simulators, data pipelines, validation systems—through interfaces hardened enough to trust. They need embedded domain judgment and the accumulated knowledge of what “validated” actually means when the artifact ships into a multi-ton machine. And they need evaluation and governance built in, because in this domain a plausible answer and a correct answer can be separated by fatal consequences.

When I say physical AI needs its own agentic platform, I’m talking about a platform that combines state-of-the-art models grounded in the data layer, tooling, and domain expertise of physical systems, with evaluation and governance native to the platform. It is not a chatbot bolted onto engineering tools, and it is not something you get by fine-tuning a general-purpose agent. It’s a different architecture, and it demands as much AI innovation as the models themselves.

Speed Doesn’t Compromise Safety

When discussing agentic capabilities in physical AI development, the reflexive objection is that agents have no place in safety-critical engineering. Automation means moving fast and perhaps getting a few things wrong. It’s a reasonable position when the thing in question weighs several tons.

However, in safety-critical systems, the speed of your feedback loop is a safety mechanism. When validation takes weeks, teams test at irregular milestones. When it takes minutes, they test on every change. Problems surface earlier, when they’re cheap to fix. Coverage expands to orders of magnitude more scenarios. Requirements get implemented more accurately and verified more often. The slow, careful-looking process isn’t by its nature the safest one. It’s often the one where defects age quietly for months before anyone notices.

What actually makes agents safe in this domain isn’t slowing them down. It’s drawing the line correctly. Automate development, validation, and operations workflows, so safety-critical issues get resolved faster, and refuse to automate certification, regulatory sign-off, and final engineering judgment. High-stakes agents propose; humans decide. Anything touching production systems runs behind approval gates. This isn’t a temporary concession while the models improve. It’s the correct permanent architecture for physical AI, in the same way that a well-designed autonomous vehicle has a defined operational domain rather than unlimited authority.

When agents help build physical AI faster, that speed can make the end product safer.

Models Don’t Compound. Systems Do.

When you understand these assumptions are holding physical AI back, the logical next step is combining better intelligence with faster learning, creating an agentic flywheel where frontier models and frontier engineering systems feed each other.

A continuously turning flywheel looks like this. A machine underperforms in the field. Agents mine the operational data to find out where and why, or synthetically generate the scenarios that expose the gap. Findings become requirements, and requirements become test cases. Test cases become validated software, which is deployed to the fleet. The fleet generates new data, which makes the models better, which makes the agents better, which speeds up the next turn of the loop. Every turn used to take months and a room full of specialists. Each stage that agents accelerate doesn’t just save time; it increases the number of turns, and each turn compounds both the speed and the intelligence. Smarter models turn the loop faster. A faster loop makes the models smarter. That’s the compounding effect the industry is leaving on the table when it treats intelligence as the whole game.

We know this flywheel is real because we’ve been running it on ourselves. Applied Intuition has spent a decade at the frontier of physical intelligence — perception, simulation, validation, vehicle software across automotive, trucking, mining, agriculture, and defense. Over the past year we built an agentic platform grounded in that infrastructure. It’s called Dana, and our engineers have built more than a thousand internal apps and agents on it. With Dana, development cycles have become roughly 20x faster, with higher output quality. Deployments went from once every few weeks to multiple times a day. Applications that took months to build now take days or hours. And, tellingly, applications that would never have justified months of effort now get built regularly. When the cost of building drops by an order of magnitude, the set of things worth building expands by more than an order of magnitude.

The intelligence made the engineering system possible; the engineering system made the intelligence matter. Neither alone gets you there.

What This Means for the Next Decade

If the industry keeps betting on intelligence alone, the next decade of physical AI looks like a slow one: dazzling demos, decade-long programs, and a widening gap between what machines can do in a lab and what’s actually operating in the world. The models will be extraordinary and the deployment curve will stay stubbornly flat, because every improvement will queue up behind engineering organizations that absorb change at last year’s speeds.

Pair frontier intelligence with agentic engineering systems and the curve bends. Physical AI starts reaching its full potential. Farms that stabilize output through labor and climate shocks. Mines with continuous, safer extraction. Freight networks that self-route around disruption. Defense systems that hold under degraded conditions. Multi-hundred-billion-dollar markets converging on the same stack, with the learning loop at the center of all of them.

Software ate the world by making it cheap to build applications for the digital economy. Physical AI will do the same for the physical one, but only if building intelligent machines becomes as fast and iterative as building software, without compromising the discipline safety-critical systems demand. That takes the best models and a reinvention of how we engineer, and the second half is the one almost nobody is building.

Twenty years from now, we won’t remember which company had the best world model in 2027. We’ll remember which company figured out how to continuously turn intelligence into deployed systems. That’s the problem we’ve been working on.

Peter Ludwig is the co-founder and CTO of Applied Intuition, a leader in physical AI and the company behind Dana, an agentic platform for building intelligent machines. Since co-founding the company in 2017, he has helped grow Applied Intuition into one of the world’s leading AI infrastructure companies, serving the automotive, defense, trucking, construction, mining, and robotics industries.

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

That logo you’re chasing is costing you your market.

Founders with great AI products are spending months courting their first Fortune 100 customers, burning through funding rounds, and tying up their teams on deals that will never convert. They chase these logos not because the deals are big, but because a marquee name on a sales deck supposedly makes every deal after it easier. So founders focus all their attention on them, and often even offer steep discounts (or pay customers to use them!) to land them. The founders don’t care that the revenue barely registers. The logo is the whole point. Meanwhile the buyers who actually need the software, and would pay full price for it, have never heard the company’s name.

It’s not that there’s no logic to this. AI is new, so founders assume they have to educate the market, and they default to the Lighthouse strategy: win a few marquee customers, build social proof, and reassure the buyer who’s afraid of making the wrong call. But for a lot of AI companies that instinct is exactly backwards. Their buyers already understand the problem and aren’t afraid of getting it wrong. They just need the math to work, and every week spent chasing prestigious logos is a week a competitor spends selling in Ohio. Those companies would be better off pursuing the Landgrab strategy: win on math, move fast, and sign the largest number of customers possible, logo be damned.

There are two GTM playbooks for AI companies selling into the enterprise: the Lighthouse and the Landgrab. The difference between them isn’t about product quality or team strength. It’s about what you’re selling and who you’re selling to.

The Lighthouse

Category creation is when AI enables work that couldn’t be done before. This kind of solution has no precedent inside the buyer’s organization, no incumbent to displace, and no existing mental model in the buyer’s head. You’re not replacing Salesforce but inventing something new, and that often means you’re asking buyers to take a leap. Social proof is everything so you need lighthouse customers whose adoption signals that this category is real and you’re the one to bet on.

Harvey built AI for legal professionals, a category that now feels entrenched but only a few years ago had no players, let alone customers. Law firms were buying research tools from Thomson Reuters and LexisNexis, which surface information for an associate to interpret. Meanwhile, Harvey does all that work itself: drafting, research, and due diligence across thousands of documents. Lawyers are trained to avoid risk, so no firm wanted to go first, but when Allen & Overy signed in late 2022, followed by Paul Weiss in early 2023, every peer took notice. Today Harvey has hundreds of millions in ARR and an $11 billion valuation, but it first needed A&O and Paul Weiss to tell the legal industry AI-powered legal work was real.

Hebbia ran the same playbook in financial services, building an AI intelligence platform for firms whose teams spend 60+ hour weeks poring over high-stakes data rooms. In this world of highly confidential deals and guarded reputations, no fund wanted to go first either. Hebbia broke through with the world’s largest private equity firms, hedge funds, and consultancies as early customers and then expanded to more than 40% of the largest asset managers by AUM, including KKR and BlackRock. Just as with Harvey, marquee names went first and the market followed.

Capturing lighthouses is a small-team, founder-led, and high-touch effort. Applied in the right markets, it lands large deals: six-figure ACVs at least, often seven figures early on. Sales cycles can run three to six months (or longer) because you’re navigating POCs, custom work, and buyers wary of costly mistakes. The team that closes the deal is often the same team that delivers the product, which is expensive and unscalable by design. Going first can come with outsized rewards for buyers, so the best lighthouse sellers make the leap feel like a chance to win big instead of a chance to be wrong.

The Landgrab

The game flips entirely when the buyer already knows the problem and a mistake won’t cost them their job. They understand what you’re selling, and the pitch is simple: “I replace Y at a lower cost or with a better outcome.” You don’t need Stripe’s CTO vouching for you to get a meeting. You get the meeting by showing a VP of Support their current spend and saying “we cut this in half.” Social proof still helps you close faster here, but the market already believes in what you’re selling.

In these markets, sacrificing speed is deadly because founders are not only competing with other startups, but also with incumbents building AI. Sure, Zendesk adding an AI copilot is fundamentally different from Decagon replacing the entire support function with agents. But the incumbent already owns the customer, and the more AI it adds each quarter (whether it’s make or buy), the harder it becomes to pull that customer away. That is why speed is the whole game. In the words of our colleague Alex Rampell: you need to get distribution before the incumbent gets innovation. The way you do that is with a landgrab.

Take Stuut, which automates accounts receivable: collections, payments, cash application, and dispute resolution. SAP and Oracle ship AR modules in their ERP suites, and players like HighRadius have sold point solutions since the early 2000s. Yet teams still lose countless hours chasing invoices. Stuut finally solves this: customers increase cash flow by 40%, cut manual tasks by 70%, and lower their collection period by 37%. The company went wide early, leaning into the lower middle market over Fortune 100 logos, and now serves manufacturers, distributors, and logistics companies across Michigan, Ohio, Texas, and beyond, deploying in under a week against the 6 to 18 months of a traditional rollout.

Decagon used a similar motion to win in customer support. The founders ran roughly a hundred customer conversations in a month before building their product, then sold on rapid deployment and immediate ROI, scaling from 0 to 8 figures in ARR in 18 months. In 2025 alone, the company signed more than 100 new enterprise customers across travel, finance, health, and retail, tripling its valuation to $4.5 billion in under six months.

Landgrab selling is demo-driven and relies on a larger team. The product needs to be standardized enough that a customer can onboard fast and see value quickly. Implementation is driven by forward-deployed teams specialized in delivery, rather than discovery. The unit economics have to work at volume because volume is the whole strategy.

Picking Your Strategy

Enterprise sales is fundamentally about risk and reward. Behind every deal is a person who has to sign their name to the decision, and who wants the same thing all of us want: for the thing they approved to work, and to still have their job next year. Instead of evaluating your product in the abstract, they’re deciding how much personal exposure it creates, and what evidence would make that exposure bearable.

That’s why two questions help you draw the map and decide which strategy you should pick.

How exposed is the buyer who signs? Not every mistake costs the same. In customer support or AR automation, a faulty reply or misstated invoice annoys a customer and gets fixed - the buyer’s downside is a bad quarter, not a bad career. A buyer’s exposure climbs with three things: whether the industry is regulated enough that a vendor mistake becomes the buyer’s compliance problem, whether you’re replacing a system of record or adding a tool alongside one, and whether the output faces the outside world (a filed document or customer-facing answer vs. an internal draft someone reviews). In law and financial services all three run hot, and one fabricated figure can misprice a position or sink a deal. For those buyers, the ROI math is beside the point. They’re managing personal downside no discount can offset.

Does social proof travel? In some markets, reputation carries well, while in others it does not. In financial services and law, firms watch each other obsessively, status is legible, and landing two marquee firms moves the whole market: whoever went first has effectively done the risk assessment for everyone behind them. In AR at mid-market firms, the controller in Des Moines does not care that a household-name brand uses your product and is never going to hear about it anyway. Buyers there don’t watch each other closely, so each sale starts from zero and a marquee logo buys you less. Concentrated, status-driven markets carry proof while fragmented ones make you earn every deal on the math.

Put these two questions together and you have the map. When exposure is high and proof travels, you’re in lighthouse territory: a few credible buyers going first unlock the whole market, the way it played out in legal work and financial research. When mistakes are recoverable and proof travels less, you’re in landgrab territory: math closes the deal, and coverage wins the market, which is what is happening in customer support and AR automation. The other two corners matter less to enterprise sellers but are worth naming. When proof travels but isn’t required, you may not initially need a large sales team; the product spreads itself, for instance engineer to engineer, the way dev tools do, and category creation happens bottom-up. Finally, when the buyer needs proof but the logos never reach them, you’re in a hard market and you likely haven’t heard of the companies stuck there.

You’ll notice other patterns that seem to predict the motion: SMB skews landgrab, risk-averse industries skew lighthouse, additive tools move faster than system-of-record replacements. Trace any of these back far enough and you land on exposure or proof. The controller at a 200-person distributor is a landgrab buyer because her market is fragmented and her deal is small. The industry code on the account never entered into it.

Markets don’t always fall cleanly into one bucket, and the two questions above lead to the clearest conclusion at the extremes. For the ambiguous cases, a few other factors are also worth weighing. For instance, an existing budget and napkin ROI usually confirm you’re in landgrab territory, but only after the exposure question clears. A buyer can have the budget, see the math, and still refuse to move until someone credible goes first. When the two point in opposite directions, exposure wins every time.

Sales cycles can provide another gut check to identify which territory you’re in. If they run longer than 60 days, if you’re doing custom work to prove the concept, if the buyer asks “is this safe?” before “what does it cost?”, your buyers need proof, and you need to run the lighthouse. The more common mistake right now is the opposite: assuming buyers need proof just because the technology is AI and therefore feels new. But the controller whose worst case is a misstated invoice isn’t managing career risk. Show up with a marquee logo instead of a number and you’ve answered a question they never asked.

The Traps

Lighthouse Traps

Becoming a hostage to the logo. Everyone wants the same marquee logos, and you end up in a brutal fight for the same 500 accounts that every other AI startup is pitching while the logos extract concessions because they know you’re desperate. Lighthouse customers are a means to an end. Get a few trusted names, then run the table. The vast majority of revenue sits in companies no one’s heard of.

Prestige without payback. The wrong customer won’t collaborate on repeatable software (you become a consulting shop), won’t pay recurring (unsustainable economics), or won’t pay high enough ACVs (the math breaks). Worst case: a prestigious logo that teaches you nothing replicable and underpays, and you’ve traded time for a vanity metric.

Pilot purgatory. Big logos love pilots. Six-month POCs that never convert, burning your best people on deals that were never real. The fix is time-boxed pilots with clear milestones and auto-converting contracts.

A lighthouse for one. You over-rotate on your lighthouse customer’s requests and build a product perfect for them and useless for everyone else. You win the lighthouse but no other ships follow the light in.

Landgrab Traps

Dying of indigestion. When you can sell to anyone the discipline is saying no. Some customers are poison: hard to onboard, hard to drive results for, low ACV. Without qualification discipline and a deal desk, you wake up with 200 customers and 50 underwater.

Grabbing land you can’t hold. You’re selling at volume, and if you scale coverage before the product is ready you create detractors at scale. 50 unhappy customers means churn. 500 is a reputation problem.

Mistaking the valley for the market. Canvassing every bus stop in a major metro isn’t grabbing land. The real opportunity is figuring out how to show your ROI to the 50,000 companies outside of normal networks.

Sequencing: Lighthouse to Landgrab

The best companies don’t stay in one mode forever. They sequence from lighthouse to landgrab deliberately, getting a bellwether in one vertical, dominating that vertical, and then finding adjacent verticals that look similar. At Affirm (a story worth its own post), the breakthrough was Casper. Once they had one mattress company they got every mattress company, and then they went to exercise equipment, and then to things that look like exercise equipment but aren’t. Mattresses and Pelotons have nothing in common except that they’re big-ticket items people want to pay for over time, and Affirm saw that before the market did.

Eventually, new categories turn into recognized ones because the lighthouse customers define them. But you have to earn the transition. Going wide before the category is established burns cash and credibility. The signal that you’ve earned it is buyers approaching you with allocated budgets and asking for a demo rather than asking who went first. When that happens, the lighthouse worked, and it’s time to run the landgrab.

Closing

Every AI founder believes they’re inventing the future. And they are! But your buyer doesn’t purchase the future; they purchase either proof or math. If they need proof, go win the logo that gives it to them. If they need math, get on a plane and show it to them before your competitor or the incumbent does.

The founders who get this wrong won’t fail because they built the wrong product or picked the wrong strategy off a menu. They’ll fail because they never asked which game they were in, and in a market moving this fast, you only get to ask it once.

Thank you to , , , Spencer Wiedeman, Frank Golden, and for their thoughts on 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

Are Semis Cheap?

It’s pretty much common knowledge at this point that chips have been the “AI trade.” First, it was Nvidia’s GPUs (and the related players in that supply chain), and then it broadened out CPUs, too, but especially memory.

The reason is pretty straightforward: all that AI capex. Hyperscalers have been pouring their free cashflow (and, more recently, debt) into the computational infrastructure necessary to meet the computational demands of AI. That huge demand impulse has led to a generational run for the broader semiconductor industry, where demand has outpaced supply, leading to both historic growth and pricing power.

Those are the sorts of things that investors like, and they’ve responded accordingly.

Just under 50% of the S&P’s earnings growth is expected to come from semis alone.

AI Capex has been an earnings bonanza for the chipmakers.

With all that earnings growth, you’d think that semi multiples would re-rate in kind. And while they certainly did, for a bit, that’s no longer the case. Multiples for the semis sector are actually right around their 10 year average, and even dropped slightly below:

This chart was compiled a few weeks ago, when Forward P/E for semis was at ~19.5x, slightly below the 19.7x that the industry has averaged, including the 6 years before all this AI capex really began. As of earlier this week, the industry multiple had nudged higher to something closer to 21x (depending on how you calculate it, since semis don’t report as a group).

But in all events, for an industry on an historic run, forward multiples are historically unremarkable.

By way of comparison, recent semis multiples are not all that different than the very slow n’ steady healthcare sector:

Healthcare has been trading at more or less the same prices since 2013 (with some intervening wiggles), but recently crossed paths with the hottest hardware in the world.

Healthcare and semiconductors. These are the same, apparently.

If you dig beneath the aggregate industry-level picture, the story is even more peculiar:

Some of the biggest names in the semiconductor industry are valued at-or-below the industry average—despite consensus analyst expectations for earnings growth (on a calendar year basis) running from ~20% to 60%.

Micron, the crown jewel of the memory shortage, especially stands out. Analysts expect earnings to grow nearly 60% yoy, but as of July 21 the memory-maker was trading at just about ~6x earnings, the lowest in the batch!

So, why does it appear that semis are suddenly cheap? If there’s such a bottleneck around memory specifically, why does the premium appear so low?

Well, there too, the answer is pretty straightforward. Investors are wondering whether semis can keep it up. Chips are a historically cyclical industry, and it’s the nature of cyclical industries to have peaks and valleys. We’re in the peak, so surely the valley must be right around the corner (so the story goes).

It’s also partly why those industries tend not to ramp up supply even when it appears like there’s a shortage: because they’ve been around long enough to know that every shortage eventually becomes a glut. Bringing new chip-making supply online takes time and a lot of money, and the concern is that by the time all the new supply is online, demand will have waned (leaving a lot of slack capacity in its wake). The energy industry knows this perhaps best of all.

Plus, as it pertains to memory especially, it’s never really been that expensive before, but since necessity is the mother of invention, plenty of memory customers are now highly motivated to figure out how to get more memory-efficient. That’s yet another reason why at least some investors are wondering how long the run can last.

Consider just how unusual memory’s pricing power happens to be right now:

Napkin math says Micron’s gross margins are almost three times higher than its previous 5-year average.

Those margins reflect a dramatic supply-demand imbalance in Micron’s favor, but again, the question is: how long can it last? Sure, “this time is different,” but how different and for how long? Analysts seem to think the answer is “a good while longer,” but for now at least, the market appears to be less convinced.

IRL is So Back

There was a brief moment in the thick of the pandemic lockdown when some people boldly predicted that people would never hang out in-person again.

Social and professional interaction would be forever remote, everything would be virtual, and the future would be grimly animated avatars, mingling in Sim-style boardrooms and “common areas.” It got to the point that established businesses were paying millions of dollars for “real estate” in the metaverse to frontrun the inevitable shift to the global simulacrum.

Obviously, the Snow Crash future hasn’t arrived (not yet, at least), but the prediction was prescient in at least one sense: venues for irl socialization are, in fact, in serial decline:

The number of hang-out places per-capita has declined anywhere from a half to a third since the turn of the century.

Fewer bars, fewer bowling alleys, fewer marinas and fewer movie theaters—query who hangs out in a marina, and other than movie theaters, the pandemic does not appear to have materially impacted the decline, but either way, the point stands: irl gathering places are in short(er) supply.

The thing is, however, there’s some data that suggests that recently, at least, demand for irl experiences has surged. Whether it’s an emerging social phenomenon, or simply the rising cost of airfare, hospitality and dining out, consumer spending on experiences has accelerated far more quickly than spending on the broader category of services:

Consumer spending growth on experiences has ~tripled since 2024, and is similarly running ~3x the growth of service-spending more generally.

It’s not just consumer spending data that tells the “experiences so hot rn” story. Foot-traffic data from Placer.ai points in a similar direction: getting out on the town, in public, with people, is very much en vogue. If you look at foot-traffic to retail corridors, there are only two times of the week where people appear to be visiting more than they did before the pandemic:

Friday and Saturday nights have experienced a moderate increase in retail foot-traffic relative to 2019—that stands in contrast to every other day (and time of day) where fewer people appear to be milling around retail corridors than before. On a yoy basis, the trend is even more pronounced: foot-traffic for the 8-12pm block (and only the 8-12pm block) is higher across the board.

It’s hard to imagine why nighttimes (and especially weekend nighttimes) would be busier, if not for an inclination to socialize. Retail shopping as a whole has increasingly moved online, with ecomm penetration climbing steadily to ~30%—which is consistent with the daytime decline in retail corridor foot-traffic. If you need to shop, you’re increasingly doing it online. But, if the point isn’t to shop, but rather hang out with friends, then window shopping after work and/or a weekend night out on the town is apparently more popular than ever.

People are so desperate to get off their couch and away from their monitors, that the Mandalorian and Grogu inspired the biggest above-trend trip to the movies for the year:

Memorial Day may have had something to do with it, but either way, in the two weeks following Grogu’s big screen debut, trips to the movies were ~80% higher than the weekly average.

Perhaps it won’t last, and there’s still no accounting for taste, but while AI may ultimately disrupt a whole slew of incumbent businesses, if there is one secular theme that AI cannot disrupt—looking at you, AI girlfriend—it’s demand for actual human connectivity and experiences.

Gen-Z’s Small Business Surge

We’ve written previously about the recent surge in small business formation, and the potential rise of the AI-enabled solopreneur. Yes, more businesses are being formed, yes, they tend to be “low propensity to hire,” yes, they appear to be growing more quickly than before, and yes, the fastest growers appear to invest more in technology than the others.

That’s not a definitive case for the AI-enabled solopreneur, but one might say the data is triangulating.

Here’s another wrinkle on that theme. According to data from BofA, Gen-Z specifically is increasing new business applications more so than any other generation:

In March, Gen-Z new business applications more than doubled relative to the year prior.

To be fair, Gen-Z new businesses are increasing off a relatively small base, but as a share of total, Gen-Z is taking a larger share of the pie:

Gen-X is still responsible for the lion’s share of new businesses, but the ratio of Gen-X to Gen-Z formations has gone from ~26:1 in 2020 to just ~4:1 in 2026.

What accounts for the Gen-Z surge in entrepreneurship? It’s hard to say. It could be an AI-native generation putting technology to work, or it could be a byproduct of a tough job market for entry-level hires (or some combination thereof, or something else entirely).

In all events, it remains the case that small businesses do appear to be deploying AI to meaningful effect:

According to the Cleveland Fed’s small business survey, 71% of respondents said AI increased productivity, while 39% and 31% said AI increased quality and sales, respectively.

Take that for what you will, but for now, we’ll add it to the AI-driven small business renaissance file. If Gen-Z is joining the party, then we’re here for it.

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

There’s a certain rhythm in the history of computing: we can call it the specialization turn. When a workload is varied and shifting—usually early in the lifecycle of that workflow—people pay for flexibility: nobody knows what tomorrow’s program will look like, so the work goes to a general-purpose system. That works for most things, most of the time.

But every so often a workload grows so large, and coheres into a shape so stable, that the economics of the workload can be improved by specialization. The work can migrate into hardware built in its own image: into custom chips and, increasingly, entire custom systems built around them. A machine designed for one job, and only that one job, can do that job faster and cheaper than a machine that has to be ready for everything.

That’s what we’ve seen through the history of computing. Graphics migrated from the CPU to a new class of accelerator, the GPU: the “graphics processing unit.” The same turn happened with networking: the internet’s trillions of daily packets long ago stopped moving through CPUs and started moving through custom switch chips from companies like Broadcom.

In short: the more important a workload, and the more stable its shape, the more sense it makes to build hardware in the image of the workload.

We are witnessing this now for perhaps the most consequential workload of our lifetime: AI inference.

Demand for inference is going vertical

When people talk about AI compute, they typically mean training: training has been where all the drama (and dollars) lived these past few years. But training, in accounting terms, is just capex. You pay for it once per model, the way a studio pays to make a film: an enormous cost, typically, but a fixed and amortizable one.

Inference, by contrast, is opex: a cost that scales with usage. Every time you ask ChatGPT a question or make an API call or have an agent write some code for you, the company that’s serving the model incurs some cost. To generate a single token, a model needs to read its weights—hundreds of gigabytes of parameters—out of memory, run them through the matrix arithmetic, produce the token, and then do it all over again for the next one. Every token, in other words, has a price in silicon time and electricity. Inference is the COGS of intelligence.

And for the companies serving models at scale, the last few years have seen the demand for inference, and the cost of inference, go sharply vertical. In May 2026, Google announced that it was processing 3.2 quadrillion tokens per month across its products, roughly 300 times what it had been processing two years earlier. We see the same story with OpenAI, which now has roughly a billion monthly active users and a rapidly growing coding agent platform in Codex, and with Anthropic, which has seen enormous success with its Claude Code product—not to mention all the companies building on top of these platforms. AI inference is quickly becoming the largest workload computers have ever run; and we believe that it’s only going to grow much further.

The machine serving nearly all of that demand today is the GPU. The reason for that is simple: the GPU is flexible. The GPU’s flexibility—the fact that it can support all sorts of demanding workloads—is the reason the AI era exists at all: when nobody knew which architectures would win, a processor that could run anything was exactly the right tool.

But if you hold the workload up against the machine, the mismatch is hard to unsee. Generating a token is less about arithmetic and more about memory: to produce each token, the chip must stream the model’s entire weights, plus a growing cache of everything already in the context window, and performing only a couple of simple operations on each byte it reads.

The standard patch for this is what’s called “batching”: serving many users’ requests at once, so that each expensive read of the weights from memory is amortized across dozens or hundreds of tokens instead of just one. Batching works well; but it’s not a panacea. It buys throughput, but it does so by selling latency and making every inference workload slower. That’s a trade that the fastest-growing workloads—coding agents in tight loops, long-context reasoning—can’t really afford: customers want things done fast.

In short: GPUs are built for everything; inference is now big enough, and differentiated enough, that it requires custom hardware. And this is particularly true for AI, which is so compute- and energy-intensive. Data centers are now gated by watts rather than dollars, and every watt that a general-purpose chip spends on flexibility that the inference workload never touches is a watt not producing tokens. Tokens per watt is the real currency of inference.

Over the last few years, the players with the most money at stake have more or less figured this out. For neural networks, this transformation started in the early 2010s, with Google’s development of the TPU, the “tensor processing unit,” and the entire infrastructure around it—custom interconnect, pods, and cooling. Other hyperscalers, like Amazon, Meta, and Microsoft, have followed; and now the frontier labs, like OpenAI, are following as well. This isn’t a mysterious development. It’s just the oldest pattern in computing reasserting itself: when a workload gets big enough and stable enough, it earns specialized hardware of its own. The machines that win a stable and huge market like AI inference are the ones designed as a whole, from the transistor to the data center floor, and that nail production scale to meet the demands of this new work.

The Etched bet

In 2022, months before ChatGPT and years before Cursor, Claude Code, and Codex, Gavin Uberti, Robert Wachen, and Chris Zhu dropped out of Harvard to make a bet that looked, at the time, nearly irresponsible: they would build an entire inference system—new chips, boards, interconnects, racks—from scratch.

Their thesis was simple. Inference would be the largest market in AI, and the general-purpose GPU wouldn’t be the endgame of that market. Hyperscalers had realized that already. But their custom silicon programs were for their workloads: you can rent a TPU, but you can’t buy one, or rack it in your own data center, or build a business on it. For everyone who wasn’t a hyperscaler—for AI labs without silicon teams, inference clouds, sovereign AI programs, enterprises building their own fleets—there was a huge opportunity for chips designed specifically for inference. Gavin, Robert, and Chris set out to build exactly that.

What began as a radical, contrarian wager has matured into a strong bet on the defining challenge of our era.

Over the last few years, Etched has hired more than 400 engineers from Nvidia, Google’s TPU group, Broadcom, Apple, SK Hynix, and TSMC. A striking share of them work not on chip design but on production: supply chains, rack assembly, burn-in, and logistics. In a supply-constrained market, the scarce skill is not designing a fast chip; it is shipping complete, working systems by the thousands. As Gavin, Robert, and Chris like to put it: production is the product.

Under the hood of that product are two core ideas. The first is low-voltage inference: running the math blocks at roughly half the voltage of a typical AI chip, packing several times more usable compute into the same power envelope—more tokens per watt, the currency that matters. The second is cluster-scale memory: an ultra-low-latency interconnect that pools memory across chips, so that during decode, an entire rack behaves like one enormous machine. Both of these technologies are not possible by only designing a chip: they require technical breakthroughs across the entire system across memory, power delivery, cooling and more—precisely the bottlenecks specialization predicts, in precisely the market where every multiple of tokens per watt converts directly into revenue.

But as impressed as we are with the elegance of Etched’s first product, what’s blown us away is their tempo of execution. Etched’s first production chip taped out on TSMC’s N4P process and worked on the first attempt—a“first-pass silicon,” in industry shorthand—rare even for industry incumbents. As we sat in their 2 megawatt lab in their office, we heard stories again and again of the team defying the odds to achieve unprecedented timelines, flying across the world to unblock vendors and running company-wide day and night shifts to bring up their chip in under two months. (The industry norm is between six and nine months.)

The result: those first racks ship to customers this summer; a 10 megawatt site is standing up, with customers already running workloads on Etched hardware remotely. Despite the origin story, what Etched has built is not a transformer-only machine, but a full-stack inference system to serve the most demanding workloads, whether many-trillion parameter MoEs, state-space models, or long-context agents.

History is unsentimental on this point: accelerated hardware will be built for the most important workloads, but very few can build a machine capable of production at scale. It’s not enough to have a great architecture and demo rack; you have to ship at scale, with the speed and quality that the most demanding customers in the world require amid fierce competition. In a world where production is the product and the machine is the moat, we’re thrilled to partner with Etched.

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

Renting an apartment is a strange transaction. It is the biggest bill most renter households pay each month. Yet the deal itself gets struck in a hurry: a couple of weeks of searching, a tour squeezed into a lunch break, an application that wants your pay stubs, your credit score, and possibly a guarantor. Housing is one of the only major purchases where the product screens you back. For rent, about 46 million American households (35% of the country) do this dance, per the Census Bureau.

Welcome to the latest edition of our chartpost series, where we use one company’s data to light up a corner of the economy. This time the corner is literally nationwide: the US rental market, seen through EliseAI, which builds the AI agents that automate communication and operational workflows for apartment buildings. When you text a building “is the one-bedroom still available?” at 11:48pm and someone helpful writes back immediately, that is likely Elise.

Here is why that vantage point matters right now. Census data shows that measured from the supply side, the rental market finally looked like it would see some relief. Builders completed more than 600,000 apartments in 2024, the most in a single year since 1986, and kept a heavy pace through 2025. Vacancy sat at record highs. National rents have been flat to slightly falling for three straight years. Balance, at last.

But that progress is occurring against a backdrop of a much larger structural shortage. Estimates still place the national apartment deficit at 5 million apartment units. Half of all renters remain cost-burdened, meaning they spend more than 30% of their income on housing. New supply matters, but it has not yet translated into broad affordability for renters.

To understand how that tension plays out in the apartment search itself, we need a view beyond supply, vacancy, and rent levels. For the first time at this scale, we can now examine the demand side through data generated by real renter interactions. EliseAI handles millions of conversations with prospective and current renters each year, throughout the country and at every hour of the day. The analysis that follows draws on aggregate data from those conversations to identify patterns in what renters say, what they do, and how their behavior is changing with AI.

Classify every opening message renters send and you can measure the share that leads with money: the rent, the fees, “any specials?” Call it the rent-anxiety index. It is climbing.

Two details stand out. The low point in the index was August 2024, in every market, and the climb since has been relentless. And San Francisco, of all places, entered this window among the least price-anxious big markets in the country, then added 15 points in fourteen months to surpass New York in late 2025.

So where does anxiety bite hardest? The cleanest way to see the country is to cross the market’s two sides: what landlords advertise, meaning concessions (industry speak for a move-in deal: a free month, waived fees, occasionally a gift card), against what renters actually ask about. Do that for the 25 biggest metros and America splits into four rental markets:

Squeezed, steady, cruising. Then look at the fourth corner, the one reserved for anxious renters swimming in deals. It is largely empty. No large American rental market is both generous and stressed. Score that same gap city by city, advertised relief minus renter asks, and you get a genuinely new scorecard: where renter stress goes most unanswered.

New York renters carry about 19 points of unrelieved stress. Denver landlords, meanwhile, are offering 38 points more deal than renters request. So, within the U.S., major metros present opposing problems and differing rental landscapes (duh).

This is also where the data starts overturning received wisdom. Did someone say “narrative violation”? Received wisdom says renters ask for deals where deals abound. Check the conversations against Zillow’s count of listings advertising a concession, and the truth runs backwards.

Renters ask about deals more in the places where the fewest are posted. Said differently, it makes a bit more sense: listings show what landlords advertise; conversations reveal what renters want but cannot find.

The stress map isn’t static; the metros where anxiety rose fastest are not the ones you would guess, either.

Boston. Baltimore. Minneapolis. Detroit. The surge lives in the Northeast and Midwest, while Denver and Southern California, currently blanketed in move-in deals, barely moved. This isn’t a story about cheap buildings. A quick decoder ring, since the housing industry loves letter grades: Class A means the newer glassy buildings with the rooftop and the package room, Class B means older but well kept, Class C means the oldest and cheapest stock, often called workforce housing. Here is the anxiety index for each grade:

The three lines sit on top of each other for two full years, never more than about two points apart. Luxury renters and workforce renters are both opening with the same nervous questions. So, are renters right to be anxious? Let’s examine what it takes to actually land an apartment.

Nationally, the cheapest homes draw the “longest lines,” and the renters with the least financial slack are the ones standing in them. But there’s a twist: crowded is not the same as slow.

Once renters engage on a Class C unit, it signs faster than luxury: 10.8 days to a lease, versus 12.3. The “search tax” on cheap apartments is competition, not pace. As an aside, across this entire dataset, how many people chase a unit and how long the process takes turn out to be almost completely unrelated.

Next, a ritual every renter knows: you send the inquiry, and nothing comes back. To measure the nothing, EliseAI ran mystery shops, sending realistic renter inquiries to leasing offices and tracking whether, and how quickly, anyone responded. The results are bleak.

Even Class A buildings answered only about half of the mystery shoppers, and when replies did come, the wait grew as the rent shrank. Again, renting is a rare market where the customer does the chasing.

At the affordable end specifically, renters face a double bind: longer lines for units that move faster, and slower, or nonexistent, replies. Those with the least financial slack must chase hardest and decide fastest, often with the least information.

But the reply is only the first gate. The biggest leaks come next: getting renters to show up for a tour and then begin an application. Once the paperwork starts, most finish it. In renting, the contest is attendance. Much of the attendance rate hinges on how far ahead the tour was booked.

Book a tour for today and you’re more likely to show up. Book it for next week and no-show rates more than double. In a fast-moving market, this is a problem for both renter and landlord. The longer the wait, the more time there is for plans to change, the apartment to be leased, or another option to move ahead. But, this is also the most fixable number in the piece. Instant scheduling, reminders, and self-guided or AI-guided tour options all help renters get closer to signing a lease.

The clock doesn’t do anyone any favors, either. Renter demand has two rush hours, and both are inconvenient. Tours cram into lunchtime: 52% happen between noon and 3pm, when the leasing office is at its busiest. Messages pile up after work: inquiries roll deep into the evening, after the leasing office has gone home. And that evening traffic is not just idle scrolling. Among renters who eventually complete an application, the most serious cohort in the funnel, a third made first contact outside 9-to-6. The leasing office may close at six; renter intent does not. The conversation increasingly cannot, either.

Everything we’ve covered so far is the machinery of why we’re all so anxious about rent. Now we get to look at some sharper cuts: by city, by building grade, by calendar.

City by city, the texture gets richer and San Francisco earns a paragraph of its own, because the crowding there is something else.

Behind every signed San Francisco lease stand 49 first inquiries. So for all of our SF readers, you and your 48 rivals almost double the national average. And once you’ve booked a tour, where you stumble depends on the coast.

San Francisco’s hurdle is physical: only 37% of booked tours actually happen there, versus 53% nationally. New York’s hurdle is bureaucratic: tours happen at the national rate, 51% attend, and then the paperwork strikes. Just 15% of booked tours ever become a started application, half the national 30%. From there the two cities move in lockstep and finish in the same place: 8% of booked tours end in a signed lease in both markets, versus 16% nationally. Two different leaks, two different fixes, and the same narrow finish. The one stage that behaves the same everywhere is the form itself: wherever you rent, about four-in-five renters who start an application finish it. Most of the chaos and mess is upstream; getting people to the tour and onto the application at all. The funnel differences show up in what each city asks about.

The two cities open on price at identical rates; the split arrives with the second question. New Yorkers sweat the application: income documentation, credit, guarantors. San Franciscans ask about parking and square footage. We’re all anxious, just about different things. New York’s paperwork fixation turns out to be one instance of a general rule.

In tight markets, renters do not just ask what an apartment costs. They ask whether they will qualify for it. This may be the purest expression of rent anxiety in the whole dataset: the fear isn’t just price. It is the audition (and the rejection). Question intensity isn’t limited to just price. In fact, inquisitiveness is a market personality trait.

The metros that grill hardest on price also grill hardest on amenities. Interestingly, this isn’t just a graph of income. One interrogation, many subjects, and the intense corner of the chart is almost entirely non-Sun Belt.

Amenity questions carry their own surprising signal: the metros asking most about gyms, pools, and rooftops are the metros stacking the most inquiries behind each signed lease. Asking about the pool is not idle browsing. You could read it as value-checking, making sure an expensive apartment earns its rent, but the pattern tracks competition, not price: in crowded markets, renters ask everything up front because they may not get a second conversation.

Before moving in, renters ask about the gym, the pool, and the rooftop. But, once they move in, conversations become about what goes wrong.

EliseAI powers maintenance workflows for apartment buildings (fielding requests like work orders, in industry speak), which makes this dataset a running census of what actually breaks in American apartments. When shit goes wrong, well … shit goes wrong. Plumbing, together with appliances and HVAC, makes up nearly half of all maintenance.

Breaking things even has class personalities. Luxury buildings log more appliance, electrical, and common-area tickets. Budget buildings log more pipes and pest control.

One more cut before the finale: loyalty. Once renters get in the door, who stays?

Notice that the anxious Northeast and Midwest metros cluster near the top of the loyalty table. Renewal rates also climb down the price ladder, from 57% in Class A to 63% in Class C. The renters who fought hardest to get an apartment (and a deal) hold on to it most. Anxiety and loyalty are partially two sides of the same coin.

So the market is anxious, crowded, leaky, awake past midnight, and, once settled, loyal. That’s one story. Here’s the other: how AI, quietly and quickly, started changing that entire search.

The data suggests that this is one of the clearest examples of the general public embracing AI in a major life decision, and doing it at scale. EliseAI now powers renter interactions for roughly one in six U.S. rental apartments and has engaged tens of millions of unique renters through those conversations over the past few years.

As a16z GP, Alex Immerman put it, “Finding a home is high-stakes, deeply personal, and often stressful. The strongest evidence of AI adoption is what people actually trust AI to help them do. This data shows millions of renters using AI to ask questions, compare options, and navigate the process of finding a home. The richness of these increasingly multi-channel conversations shows how quickly comfort and trust are growing. It’s exciting to see AI meet people where they are and deliver genuine value in the life decisions that matter most.”

That growing comfort shows up in the shape of renter conversations. AI is making them longer, richer, and available across channels—including voice. In early 2024, the median journey from first message to signed lease ran about 9 days and 3 messages. By late 2025: 14 to 15 days, and 7 messages. More conversations give renters more chances to ask, compare, and clarify before committing. Renting an apartment is becoming less like a transactional purchase and more like a correspondence, with more questions and more comparison along the way.

Those extra conversations matter: the talkers are the signers. Most inquiries stop after a single message, and on sheer volume still produce a quarter of the leases. But the conversations that reach five or more messages make up 14% of traffic and 39% of its leases. Depth is how the crowd gets sorted, and someone, or some system, has to be available to do the sorting.

The conversation is also spreading out. Two years ago, roughly a third of signing renters had used more than one channel along the way. Now it is 61%, spanning email, text, webchat, and voice, and journeys touching three or more channels have doubled. AI is allowing your lease to move the way the rest of your life does: across every app on your phone, with the thread still intact.

So much for the uncanny valley. In mid-2023, the median AI-handled call lasted 46 seconds, roughly the time it takes to recite office hours. By late 2025, AI calls run as long as human ones, about a minute and forty seconds each. The AI that greeted you can now actually carry the conversation.

None of this removes people from the loop. By handling the volume of routine questions, scheduling, and follow-up, AI gives leasing teams more capacity to step in when a conversation needs a person. What happens next is the sharpest conversion lever in the dataset: answer that handoff immediately, and 26% of those prospects still sign a lease. Let it sit for three days, and you’re left with less than a third of that. The clock decides who signs. Great property managers matter, and AI helps make sure they are available when it matters most.

Which brings us back to where every renter’s story begins: waiting to hear back.

As renting has become more stressful, renters have gravitated toward AI as the fastest, lowest-friction way to get at correct answers.

For a century, the rental market has been measured from the supply side: what got built, what got listed, what got charged. The demand side was 46 million households talking, asking, deciding, one conversation at a time. What emerges from that side of the ledger is a renter more anxious than the vacancy numbers suggest, and considerably more comfortable turning to AI than anyone might have guessed.

Headline numbers suggest rental market pressures are easing. The conversations say renters are still waiting to feel it. At least now, someone is there to answer.

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

Read more about Atoms from Travis himself.

We’re backing Travis Kalanick’s “new” company, Atoms, and I’m joining the board. I say “new” company a bit tongue-in-cheek, as Travis has actually been working on Atoms for 8 years. It is the realization of Travis’s multi-decade long vision and ambition to digitize the physical world.

Some companies are inevitable. The moment we invented the computer, it became inevitable that computers would take over the entire world of bits. Computers inevitably did the job of transforming information, sending information, and storing information, for every application in the world, in order to save us time and make us more productive.

Similarly, it is inevitable that within a generation, robots are going to do most of the menial work in the world of atoms: transforming atoms, moving atoms, and storing atoms are inevitably robot tasks. And while our imagination may be captivated by humanoid robots, the specialized ones are far better suited to most of those jobs. “Industrial AI” as a category is the inevitable application of specialized robots to atoms-heavy industries.

It takes a rare kind of entrepreneur to change these old-school, heavy parts of our economy. They need a gritty work ethic, drive, and range that spans across domains, from software architecture to mechanical engineering. Travis is that guy.

Atoms makes computers for the physical world

We have a mission as technologists to give everybody the opportunity to be productive, especially here in America. And Travis built one of the great companies that gave people the opportunity to be productive. If you had a phone and a car, you could solve a problem for someone else right now. What an achievement.

I think the most valuable thing someone could do with AI and robotics is to repeat the same thing Uber did for transportation, or that computers did for the digital world: to make everything and everyone more productive. There are so many companies out there (and we’ve backed many of them!) working on various digital forms of this. But most of the jobs we do are still in the physical world, whereas the frontier of productivity means physical productivity: making things, moving things, storing things.

Travis has been quietly working on a company for the last decade called “City Storage Systems” (a.k.a. “CloudKitchens”). The point was to discover all of the primitives and patterns for how people make, store, and move things for other people, and learn how to build a computer for the physical world. He started with food, because food is something everybody needs, it’s something we prepare locally, and if you can manufacture and deliver a healthy, freshly cooked meal cost-competitively with the grocery store, then that’s going to win a lot of market share. But you can only achieve it with some serious automation of the physical world, and the atoms-heavy industries that inhabit it.

Atoms will make us more productive in the physical world in the same way that digital computers did for bits. He has three of them so far: Atoms Food, Atoms Mining, and Atoms Transport. These are all trillion dollar industries, so he’s got some work to do, and we’re excited to help him win.

Reality is more malleable than you think

The lesson of Uber was, if people really want the product, then their everyday use of the product is undeniable legitimacy for your case to operate. And every battle, ultimately, is about your legitimacy to exist.

The backdrop to company-building today, unfortunately, is that a lot of people out there think AI shouldn’t exist, or shouldn’t be built by private companies, or other stances like that that are plain wrong. As a result, fast-growing companies that do things in the physical world will have to play on Hard Mode, and secure their social license to operate. This means, becoming essential to people, quickly. Uber won its battles, city after city, because they became essential to people so quickly that their social license with everyday citizens overcame the political and regulatory pressures to stop it.

This lesson is going to be applied over and over again as AI and Robotics get deployed into the physical world. The future belongs to those products that become legitimate by bringing some hustle to industries that have waited for technology to happen to them.

This is the art and science of building “self-creating realities”. The ability to call your shots here, and execute on them, is like using the Force. It’s the critical founder skill set in this new world. We’re about to invent the most amazing personal and professional productivity tools ever invented; and a small minority of people who hold incumbent positions of power are going to try and stop them. The rate-limiting step for founders is, can you make something that industry (and, eventually, end consumers) want so much that it incepts the future we all want?

People will always be the long pole in the tent

Back when Travis was running Uber, we were still in an “old media” dominated landscape, and most of the old media people weren’t really interested in talking to Uber drivers unless they had something bad to say. So, now, we’re very excited to get to work with Travis; not only to help build and fund Atoms, but also to help tell the story of people thriving as AI and robotics give us more and more ways to be individually productive.

There is a very bad storyline in tech right now that argues AI is going to take away everyone’s job and purpose and livelihoods. Therefore, for their own good, it must be taken away from everyone. (This is coming from many folks outside of tech that you’d absolutely expect, but also, shockingly, from some powerful people inside our world as well, which we obviously don’t agree with.)

We need to tell a better version of the optimistic AI story, and it can’t just be “technologists hyping it up to other technologists”, we’re past that point. The story is, now that intelligence runs on tap, what do people do with it? And the answer is, a lot! Human-driven work is more in-demand than we’ve ever been, when we have purpose and initiative. That’s why productivity is the most valuable activity we can cultivate, as individuals and as a society. People who are already being productive naturally and insatiably apply everything available, to make things better for others and for themselves. They will always have bottlenecks to work out, and judgement to apply, to new and interesting problems. There will always be things for people to do, for people who take initiative.

Limitless ambition

My partner David George wrote last month: “In late stage venture, founders are the asset class. There are certain people who understand new technology as it unfolds, who know where to deploy it into their own opportunity set, and can keep investing new capital attractively, forever.” Atoms is clearly one of those opportunities. Travis has drawn a very broad scope (move, transform, and store physical matter! Spanning from restaurants to mining!), but of course, this isn’t his first rodeo. It’s not unfocused. The discipline it took to run a company for 8 years in stealth (meaning, entirely outbound sales and recruiting!) is a remarkable test of true focus and PMF, while keeping the breadth of opportunity effectively limitless.

Limitless ambition, plus the ability to “use the force” to make reality auto-create itself, is a rare and special combination. It’s how we build real-world services, take big risks, and move the Overton Window of what’s possible. On behalf of a16z, founders everywhere, and everyone out there with initiative: Travis, welcome back to the game. Let’s cook.

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, ATOMS is announcing a $1.7B equity investment with lead investor A16Z and Ben Horowitz joining the board. As part of this investment we have merged our various businesses into one Atoms equity structure. I’ve known Marc and Ben for a long time, and we got oh-so-close to partnering up at Uber in 2011. I blame Marc, Ben blames himself but let’s just say it was on all of us and in 2017 Uber suffered the consequences of not having Marc on the board - iykyk.

On many levels, this round is a bit of unfinished business. Fuel to complete the bits-to-atoms story arc we started at Uber, continued at CloudKitchens and will now finish at Atoms.

Most importantly, it’s about people. As a young unproven entrepreneur, you work with whomever is willing to work with you. But once you’ve succeeded, you get to work with the people you choose to work with. I’m a sucker for excellence and working with Ben and Marc closes that open loop from 2011. I’m as passionate as I was then about building. After more than a decade, I’ve only gotten better, and so has A16Z. LFG!

With the ATOMS launch a few months ago, there was a huge amount of interest after many years of ninja level stealth. My vision letter outlined Physical AI in industry, but it cruised at a fairly high altitude. I figured this moment would be a good time to connect the dots between our bold vision for the future with the actual business we are building, and bring to life what A16Z and other great investors see in it.

Bits to Atoms

16 years ago, I started a journey to digitize the physical world. Understand, predict and control the physical world with software. Building “atoms-based” computers where CPU is manufacturing, storage is real estate, and network is transportation.

Uber was first. Network for the physical world. Digitized transportation for the masses.Cloud Kitchens was next. A food computer with a focus on CPU and Storage (manufacturing and real estate).

But what about an OEM that builds atoms-based computers for all the major industrial sectors? We just launched that company - ATOMS. The vision is about the coming industrial revolution where large industrial economic sectors get completely digitized.

Look around you. Everything is grown or mined - manufactured and moved. That is industry. The large-scale, organized production of goods, typically involving machinery, factories, and coordinated labor and capital. Mining, construction, heavy transport, food production are just a few examples of atoms-heavy industries awaiting massive digital transformation. Multiple trillion dollar industries rapidly transforming over the next decade powering an engine of prosperity unparalleled in human history.

The impetus for this transformation is Industrial AI - Systems of software, sensors, robotics and AI used to automate the operations of entire industrial sectors. Industrial AI is getting ready to power the next Industrial Revolution.

The Age of Atoms

The better we are at finding valuable unknown truths, the more change we will create in ever-shortening amounts of time. Changemakers, the builders of tomorrow’s progress machines, will inevitably go up against the final boss, Nature and its fierce resistance to change. Nature is going to throw everything it has at the builders in this new industrial age and it’s going to take humanity’s strongest to stay the course and get these complex systems and industries over the finish line.

Vision.

Grit.

Systems thinking.

Creativity.

Resilience.

Empathy.

Stubbornness.

Polymathic problem solving.

Regulatory change.

Discipline.

Philosophy.

Who are these builders, these renaissance industrialists?

At Atoms, we are going to give everything we’ve got to be one of them.

I want to thank A16Z, and our many other capital partners for believing in us and making the bet on the progress machines we build to power this next industrial revolution. It’s time to close the loop on some unfinished business.

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 are excited to announce the fellows who will be part of the inaugural a16z Forward Deployed Engineer Fellowship, an 8-week cohort of FDEs and Applied AI leaders at the forefront of deploying AI into real-world enterprises.

We received thousands of applications from talented people around the world. Today, we’re thrilled to share more about the inaugural cohort of 65 fellows who are leading FDE at companies like OpenAI, Mistral, Cognition, HappyRobot, Poetic, and Decagon.

Many are Palantir alumni who are now founding members of companies pioneering the FDE playbook like Barry McCardel from Hex, Kevin Bai at Anthropic, Isa Gomez from OpenAI, Senta Knuth from ElevenLabs, and Vinoo Ganesh from Kepler.

Some are FDE architects behind top high-growth startups like Luv Kothari from Snowflake, Perry Ha from Decagon, Bo Sun at Notion, Leo Mehr at Ramp, Niko Grupen at Harvey, and Anitej Biradir from ElevenLabs.

Others are the next generation of FDE builders already deploying live like Priya Khandelwal from Nixo, Tom Kimbrell from Stuut, Lucas Kiefer from Strala, Kian Fattahy from Saronic, Chelsey Fang from Pylon, Amy Jin from Decagon, Rohan Ramani from Netic, Jakub Hojsan from Exa, and Mariano Pennini from Spade.

Congratulations and welcome, a16z Forward Deployed Engineer Fellows!

Sign up here to stay in the loop about future fellowships and a16z opportunities.

Sign-up

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

This week, Applied Intuition is launching Dana, an agentic platform for developing physical AI applications. Applied Intuition began by building the tools engineers needed to develop autonomous systems, then the OS that underpins them all, and finally moved into the intelligence running on the machines themselves. As co-founder and CEO Qasar Younis describes the company today, the destination was always much bigger than tooling for self-driving vehicles. The mission is intelligence on a billion machines.

The unfashionable layer

In 2017, the autonomy industry shared a collective prediction: one of a small number of vertically-integrated companies would solve self-driving cars, operate the vehicles, and capture the entire market. As a result, capital overwhelmingly flowed into full-stack robotaxi programs. Each program hired its own engineers, assembled its own fleet, and rebuilt roughly the same internal development infrastructure.

Engineers who moved between major autonomy programs encountered a funny ritual. They left behind the tools they had used to develop and test autonomous systems, arrived at the next company, and began building those tools again. The industry had plenty of conviction about autonomy but very little shared best-practices for producing it.

Applied Intuition proposed to make those best-practices a company.

This was a minority view, and it wasn’t a conventional way to participate in an increasingly fashionable market. The large Level 4 autonomy developers did not think they needed an outside tooling supplier. Applied Intuition tried to sell to them anyway. Prominent self-driving programs like Cruise said no. Maybe it was an understandable position at the time: they weren’t going to wait for a startup like Applied Intuition to build tools that their own engineers already intended to build internally.

Many of the era’s prominent full-stack programs no longer exist in the form they did then. Cruise, for example, was acquired by General Motors and subsequently sunsetted after a safety incident. Meanwhile, Applied Intuition has outlasted (and outperformed) the majority of them.

Applied Intuition’s founding deck contained a slide called “Ideas We Believe To Be True.” One of those ideas was that any sole autonomous vehicle maker would remain a single-digit share of the global car market because the structure of the automobile industry tends toward distribution across many manufacturers.

If one vertically-integrated company took the vehicle market, the rational strategy was to be (and invest in) that company. But if it wasn’t a winner-take-all market, then the other ninety-odd percent of cars would acquire intelligence through the existing automotive industry. Those manufacturers could not each reproduce every layer of the modern software stack. An independent supplier would have to provide common infrastructure.

Qasar and Applied Intuition’s CTO and other co-founder Peter Ludwig were wagering on diffusion over concentration. The autonomous future would arrive not only via new entrants replacing incumbents, but because incumbents could become technically different companies. Applied Intuition would give them the means to do it.

Qasar and Peter brought together experience from General Motors, Google, and Y Combinator with technical depth and an unusual tolerance for the slow, exacting work of building safety-critical systems. After the large Level 4 companies initially declined to become customers, Applied Intuition sold into smaller Bay Area autonomy teams, including Voyage and its cohort. The products they developed followed the demands of their customers’ work: first a simulator for planning, then a simulator for perception, and finally infrastructure for managing data and executing millions of these simulations.

Deterministic simulation that correlates with physical reality can sound like an unglamorous product category. In practice, it gave the industry a shared touchstone for determining whether an autonomous system worked, without needing to put a vehicle on the road and testing in prod. A simulator turns a road event into a repeatable test, which can then be implemented in reality.

Market structure has since reflected Applied’s “Ideas We Believe…” slide in favor of the monolithic manufacturer thesis. This really wasn’t that obvious in advance, but in retrospect it makes sense. The automobile is not a pure software product. It is a regulated, capital-intensive physical system that’s sold through entrenched distribution networks. Software can reorganize where value accrues and which capabilities matter. But it cannot, by itself, flatten that industrial complexity into a single manufacturer.

On a clear day you can see General Motors

Around 2018 and 2019, General Motors issued a formal request for development tooling. Twenty-eight companies bid, including NVIDIA and Ansys. Applied Intuition was still a small startup and GM had procurement procedures and myriad options all vying to be chosen. Applied’s products performed better against the specification, and that’s why their pitch won.

The GM award converted a wedge among Silicon Valley autonomy startups into legitimacy with traditional manufacturers. The playbook became repeatable: begin with technically aggressive smaller companies, then use the resulting product and evidence to serve incumbents operating at industrial scale.

Roughly eighteen months after founding, Applied Intuition entered the defense industry. The company hired people who understood the field from the inside, and then applied the automotive pattern of simulation, data, integration, and validation to defense systems. Construction and mining followed, then commercial trucking.

A generic software company often enters a new vertical by changing some of the nouns in its sales deck. Applied entered by hiring teams native to each domain and rebuilding the product around the constraints of the customer, procurement system, and safety cases. Meanwhile, the product moved up a technical ladder.

The lower layer consisted of simulation and data infrastructure: generating scenarios, collecting machine data, reproducing events, and evaluating behavior. Above it came operating systems for physical machines: scheduling, middleware, memory management, communications, and functional safety. These were operating systems, responsible for the machine rather than a collection of applications displayed on it. Above that came the intelligence itself: world models and planning systems deployed across machines operating on land, in the air, and at sea.

By 2024 and 2025, the result could be measured externally. Applied Intuition raised a round at a $15 billion valuation. By 2025, eighteen of the twenty leading non-Chinese automakers were customers. Level 4 trucks using Applied Intuition technology were operating without drivers in Japan. The company had deployments with the U.S. Army and an operating system aboard U.S. Navy warships, along with production work in mining, construction, agriculture, and trucking. And in what is an anomaly for many Silicon Valley startups, Applied Intuition has largely preserved the primary capital raised from investors (close to $1 billion) while nearly doubling revenue at scale for multiple years in a row.

By then the original problem had changed. In 2017, the industry wondered when autonomous intelligence would become capable enough. By the middle of the 2020s, thanks to the increasing sophistication of transformer models, model capability was arriving faster than large organizations could deploy it.

This is why tool companies can become platform companies. A tool begins by solving a bounded task. If it succeeds, it becomes the common interface through which many tasks are performed. It accumulates integrations, test cases, workflows, and organizational memory, and eventually it stops being an accessory to the production system and becomes the environment in which production occurs.

The web browser offers a historical corollary. It didn’t “create” the underlying internet, but it did make the net usable by the vast majority of people who weren’t nerdy hobbyists, and trillions of dollars of economic activity moved toward what that access allowed them to build and do. Applied Intuition’s progression from simulator to data infrastructure to operating system followed the work its customers were already doing. As more of that work moved into Applied’s products, the software became part of how those companies built machines rather than a tool used for one stage of development.

Dana

Today Applied Intuition launches Dana.

For nine years, Applied Intuition has built the technology used to develop intelligent machines. Dana puts an agentic interface over that accumulated system. An engineer can start with a requirement, connect it to the relevant code, run the change through simulation and defined evaluations, move it onto a test bench or hardware-in-the-loop system, and eventually stage it for a physical machine. Much of that work previously required engineers to pass results manually between specialized tools. Dana can coordinate the path while retaining a record of how the system changed and why.

This arrives as the technical method behind autonomy is changing. For most of the self-driving industry’s history, advancements came in the form of imitation learning: gather enough recorded human driving and train the model to copy it. The frontier has moved toward end-to-end reinforcement learning in a closed development loop. The system encounters a problem, the team finds or generates more examples of that situation, the model trains again, and the same scenario is rerun to see whether its behavior improved. Applied already has the simulation, synthetic data, evaluation, and deployment systems required to run that loop. Dana gives agents a role in operating it.

The physical economy has historically remained resistant to software for reasons that have seemed intractable. These industries have long hardware cycles and lots of physical world considerations. These are mundane engineering problems until the machine weighs several tons and is moving near people, and someone gets hurt. At that point, every link between a spec and a deployment decision has to survive scrutiny.

Autonomous machines face an unusual burden of proof. A manufacturer may believe that a model performs well and still be unable to ship it until regulators can understand how that conclusion was reached. Dana was built around that. Its usefulness depends as much on traceability and evaluation as on the speed with which it can generate code or run a workflow.

The economics are changing at the same time. Intelligence is moving down a steep price curve. The autonomy layer itself will trend toward abundance and, in many settings, toward a negligible marginal price.

One robot maker can sell only the machines it manufactures. Dana can be used to upgrade equipment already expected to remain in service for another twenty years, and to develop new machines whose shape is no longer determined by the need to fit a human driver inside them. In other words, Dana is one way we can get a billion intelligent machines.

For years, autonomous systems were limited by the models themselves. That constraint has eased. The remaining work sits inside the manufacturers and operators trying to deploy them, in the form of adapting models to real hardware and supporting the system once it is operating far from a research lab. The market is ready to use AI in the physical world, but most companies cannot afford to assemble a thousand-person autonomy organization to do it.

Applied Intuition spent nine years learning how intelligence survives contact with hardware and the physical world. Dana puts that knowledge in the hands of anyone trying to make a machine move on its own.

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 Connor Love who is joining a16z as a General Partner. Read more about it here. - Danco

When deploying physical hardware or dealing with a diverse set of stakeholders, business models become uniquely unforgiving. In a way, this is a good definition for what an “American Dynamism” company actually is. As a result, hiring has higher stakes for engineering and non-engineering roles alike.

Strong engineering is table stakes. Drones buzz, rockets boom, boats zoom, factories clang – the products grab attention like something out of a Dr. Seuss book. But the engineering feats that take center stage are just the beginning. Whether a great product ever becomes a great business is decided elsewhere, in the less-obvious functions like Manufacturing, Deployments, Supply Chain, Finance, Sales, Policy, and Marketing. Get these hires wrong and a remarkable product stalls out. Get them right and the rest of the company compounds. So, I spent the last few weeks talking to the most seasoned founders and operators about how they built these teams.

The core takeaway is probably obvious to most, but how to execute on it is less so. The best teams look for hires that have an instinct for what “great” looks like, developed through the crucible of a more established organization. Critically, however, they must not be so fanatical about a very-specific way of doing things — that is, that they can still adapt and fit into a new culture.

Below is how to find the right people for each critical function, and how to know it when you have.

Manufacturing

“You have to be creative here in the US. We are reindustrializing, but you can’t just go to a phone tree of power electronics manufacturing engineers. When we were building the 4680 facility in Texas [at Tesla], there weren’t a lot of battery operations in the United States, so you have to look for analogs. We were hiring people out of high-speed bottling plants, out of syringe manufacturing facilities where they’re making billions of units. If you can get that creative hat going, you find there is immense depth of talent in the US and people are excited to work in new industries.” - Drew Baglino, SVP Powertrain & Energy Engineering, Tesla; Founder & CEO, Heron Power

The conventional wisdom is that you hire manufacturing talent from other hardware companies. This is partly true but mostly limiting. When Drew Baglino and his team at Tesla launched the Texas 4680 battery facility — a 50 gigawatt-hour plant built in eleven months inside a former JC Penney warehouse — the hardest problem was staffing it. There were not enough battery manufacturing operators in Texas, let alone the entire United States, which forced a deeper question: where do the underlying skills they were looking for actually live, independent of domain?

While it’s true that the discipline of building millions of things on a single line has atrophied in the United States, there are some industries that were never offshored. Consumer packaged goods, for example, are difficult to offshore for a number of reasons I won’t get into here. The result is a pool of American talent that has made millions of a thing to a particularly demanding spec.

But the unfortunate reality for many manufacturing projects is that decades of offshoring really has created a shortage of manufacturing talent in the United States. In order to make it viable again, companies like Hadrian have had to get creative and use AI to fill the gaps. That doesn’t mean “replace humans” – in reality, it looks like Hadrian’s factories where building out software and automation made it much easier to hire the workers that were available, rather than none at all. This March, Hadrian’s Factory 4 opened in Cherokee, Alabama, and is expected to create 1,000 high-paying manufacturing jobs when fully operational.

Where to find: You might think hiring someone who worked in automotive or defense would make more sense, but the true skillset that Tesla was looking for was a relentless pursuit of volume and scale akin to Asian electronics conglomerates. Candidly, this was a type of scale and precision that was found in bottling plants, not missile factories. The domain expertise can be taught, the process discipline is hard to find.

How to recognize: The best manufacturing hires also recognize the value of co-locating engineering and manufacturing. Designing something that must be incredibly performant and made millions of times is challenging and requires tight departmental integration to break through local minima. The right question to sniff out aptitude here is to ask them to describe a project they completed with half the time, half the budget, and half the people it needed. They don’t necessarily need to have achieved this, but they should be able to walk you through how it might’ve been done, or where things went wrong.

Deployments

“The deployment hire isn’t someone who just manages a process. It’s someone who owns the outcome, whatever it takes to get there. The best Deployments folks mix engineering, operations, and raw scrappy horsepower”- Justin Lopas, Co-founder & COO, Base Power

The job is defined by things going wrong in ways nobody anticipated, but finding a way through anyway — all in front of customers. Most founders hire this function too late, after the first deployment goes badly. At that point, you have both a customer problem and a hiring problem, which is a lot to get hit with at once.

The actual right time to hire for this function is when you can see your first ten deployments on the horizon, and the best structure to start with is to pair hustlers with a technical expert who has seen your specific product fail in the field before. The hustlers will move fast and figure things out, while the expert provides the hard-won knowledge to leverage. High-volume deployments should also be conceptually thought of as an extension of the manufacturing line, and any deployment blockers should be iterated on the product side, too. The best part is no part, as they say.

In defense, the function is lower volume and takes more of the form of a forward-deployed engineering (FDE) team. Being in the field with customers establishes trust, can help guarantee the success of missions, and is often the best chance to understand real operational requirements that never survive the trip up the bureaucratic chain.

Where to find: Finding good deployment people is hard, mostly because there isn’t a large pool of folks who’ve done it before at a high-growth company — which is to say, have built the plane while flying it. Dana Paz is one such person; she joined Base Power from Anduril’s manufacturing team, admittedly knowing very little about energy. But she was a fast learner, could think on her feet, and build a motivated team quickly.

Her best hires came from even farther outside the sector: hungry new graduate engineers and SpaceX production coordinators were more down the fairway, but she also found veterans were particularly well suited for this role. Veterans who had operated in high-stakes, low-resource environments translated exceptionally well. Connecting a home battery in Austin is nothing compared to rucking through Appalachia, developing HUMINT relationships, or any type of “thinking on your feet” familiar to them.

How to recognize: Beyond pure grit, the interview signal to look for is curiosity that extends beyond the immediate task, the candidate who asks why, who thinks about the next step before being asked. This is a quality you take for granted hiring engineers, but is less of a given in operational roles.

Supply Chain

“The supply chain risk that stops production is almost never the one you’re watching. We lost two weeks of production because of a coating on a washer inside a motor assembly — a tier-five supplier in India that we didn’t even know existed. The tier-four didn’t check the tier-five and suddenly you’re on a plane to get a mom-and-pop coating operation back online. You have to hire people who can think many layers deep before something breaks, not after.” - RJ Scaringe, Rivian and Mind Robotics CEO

Supply chain is a lot like a game of Settlers of Catan. You are building roads, trading resources, managing scarce inputs, and trying to win against forces you can barely control. Your primary constraint isn’t really even finding suppliers, but convincing them to prioritize a company that isn’t spending enough money for them to care — which makes this a sales problem as much as an operational one. The person you need to hire for supply chain is equal parts negotiator, technical translator, and program manager.

That last point is worth emphasizing. In the early days, you will almost certainly be managing an unsustainable number of suppliers. The right supply chain hire is someone who can operate in that chaos without flinching, and who also knows when to make the strategic call to bring something in-house. Rivian started R1 with over 600 suppliers that required overwhelming coordination. On R2, they deliberately consolidated, accepting more concentration risk in exchange for less complexity. That judgment, when to absorb supplier breadth and when to insource, is one of the highest-leverage decisions a supply chain leader makes.

Where to find: The person you are looking for is a specific combination that does not map neatly to any single job title. Technical enough to understand how components actually function in the product. A negotiator’s instincts: careful, persistent, slightly-lawyerly disposition that can hold a supplier to a commitment made before the order was placed. And a program manager at heart, capable of coordinating across engineering and other internal teams with speed. Jesse Stevens, Base Power’s supply chain leader, came out of Tesla by way of Duke Energy, a pairing of a fast paced culture and seeing a supply chain work at scale that maps almost exactly onto what a commercial hardware company requires. The direct domain experience in energy is a cherry on top.

How to recognize: Scenario-based interviews work best in finding the right person. Give candidates real situations like a sudden tariff shift, a tier-one plant that burns down, a supplier that can’t ramp. Ask what they do, how they think, what dashboards they would build, what metrics they would track, etc. The right person will also be already thinking about whatever is happening in the world that day, too, not because they prepared for the interview, but because they find these things genuinely interesting. That kind of obsession (or paranoia) is hard to fake and easy to spot.

Finance

“Your CFO can’t just be a scorekeeper. Fifty percent of the value we create over the next decade is going to be capital optimization - so the person running finance has to understand the factory floor as well as the cap table, and they have to be a true partner to the founder to make the calls that make building this big possible.” — Chris Power, Founder & CEO, Hadrian

Before joining, Hadrian’s CFO, West Owens, spent decades as a financial operator in energy, one of the most volatile industries of the time. As CFO of TerraWatt Infrastructure and VP of Structured Finance at SolarCity, Owens built the kind of experience managing capital through genuine adversity, not just growth, that many finance leaders simply don’t have. It’s also experience junior talent wants to learn from.

Subscribe now

Finance is categorically harder for hardware companies because there is no wiggle room. Software companies can adjust prices, cut headcount, move people across orgs, and rationalize spending in interesting ways. But when you bid a fixed-price hardware contract today, you are locking in your cost assumptions for five to ten years, and it’s hard to get back on track if you don’t have a clear handle on your finances from the beginning. This is also why finance ends up being the glue across the org: the finance leader is joined at the hip with manufacturing and supply chain, has to speak their language, and only finds the right solution when the three of them work it out together.

The right time to bring on a dedicated finance function is when the team can’t track what it is actually spending. For a hard tech company with a complicated supply chain, confusing contract timelines, and purchase orders charging on irregular schedules, that moment comes earlier than most founders expect – and the first person you bring on may not be the one you eventually call CFO. RJ Scaringe, Rivian and Mind Robotics CEO, frames early finance hires as a choice between architectural CFO-types, who think in big structural moves and strategic capital allocation, and the operational controller-type, who drives consistent execution month over month. That first hire is more likely the latter and could come from a variety of places: straight off a trading floor, private equity, operating within a company under a strong finance leader. What matters is that they bring the controls and audit rigor of a large company without the bureaucracy.

An architectural CFO looks beyond the P&L. At scale these companies weigh every source of capital – debt, non-dilutive funding, project-level financing. But early on the menu is thinner: the first dollars in are usually just venture dollars, and they stay that way until an F2 or F3 (read: the second or third factory, once the first has proven the line can be built and repeated) is running. Early factories double as prototypes, before the financial profile is legible enough to attract anything else. The job in that window is less about exotic capital structure than earning the right to it – finding the metric that justifies the massive upfront investment, whether unit economics, production volume, or the path the P&L takes, and being able to say plainly why an investor should be excited. That takes charisma and executive presence to explain the art of what’s possible, to investors and employees alike. It maps onto all five dimensions Brian Roberts, GP at a16z and former CFO at Lyft and Splunk, identifies as key for CFOs: leadership presence, long-game capital allocation, founder-like intensity, the ability to distill complexity, and standing as a true peer to the CEO. Eventually all hardware companies need to graduate to this strategic CFO archetype. The primary job is to give everyone confidence that the company will not run out of money. When that confidence exists, board conversations shift from defensive survival to offensive strategy.

Real estate deserves specific mention, too. Decisions around when to move, how much space, and how physical footprint maps to manufacturing capacity, are financial decisions as much as operational ones.

Where to find: Some of the best American Dynamism finance hires grew restless in banking and wanted to work on something meaningful — but banking is one source, not the source. Claire Rauh McDonough, Rivian’s current CFO, came from traditional banking with deep capital markets experience and little operational background, but earlier finance hires had been more execution-oriented. The first hire (not CFO) will likely be that operational archetype and could be in traditional finance or already inside an operating company.

How to recognize: Ask every finance candidate, especially those coming from banking, “what gets you out of bed in the morning?” Finance is the function most likely to attract candidates who have been trained, explicitly or implicitly, to optimize for title and compensation. That conditioning is hard to undo and easy to miss in a standard interview, even more so because they’re often well-practiced interviewers. The answers you’re looking for are specific and personal, and tend to come from the person who finds the manufacturing line as interesting as the cap table.

Sales

“The fastest path to a defense contract doesn’t always start with the most senior relationship in the building. It starts with identifying the right program office at the right moment in the budget cycle, and being in the right conversations six to twelve months before the RFP drops.” - Bryon Hargis, Founder & CEO, Castelion

It’s natural to want to look at what Palantir and Anduril did across government and large enterprise sales as a playbook, but the honest truth is that their effectiveness in breaking down barriers has created a much more legible path for companies today. And while government procurement and utilities sales cycles remain far from perfect — something we at a16z spend lots of time trying to improve — the Pentagon is much more forward-leaning and excited about adopting new technology, and other highly regulated buyers are on their way there, too. In practice this means educating customers is less of the job, so sales teams can stay leaner longer (and you don’t need to sue the government to win a contract). Yet, despite this narrative shift, you have to keep in perspective that the primes are still taking home the lion’s share of DoW revenue and relationships with buyers are still what matters most.

The first thing Scott Sanders will tell you about government sales is that you cannot hire your way out of early stage sales pain. Sanders was one of Anduril’s earliest sales and business development hires, yet he knows the starting point must be the same for every company: Founders need to hold the function until the motion is working and they understand why. Not because founders are always the best at it, but because you cannot evaluate whether someone else is doing it well until you have done it yourself. In defense, the function can be broken down into four activities that need coverage: relationships with budget holders, relationships with the acquisition arm, relationships with end users, and technical writing for proposals. Sometimes you find a unicorn like Sanders who can do all of the first three, but more often it ends up being several people, each stronger in one area than the others. Building relationships with budget appropriators, mainly Congress, also matters but is something you can hire government affairs firms to help with.

The most common misstep is often confusing the budget holder, the acquisition arm, and end users. As Vib Altekar, Co-Founder and CTO of Saronic, puts it, “People think getting in front of individuals with big titles is the win. But they might have zero budgetary authority. They can tell you the problem, but they can’t buy the solution.” Being able to distinguish these three categories is critical to successfully navigating government sales. That said, the end user relationship can often be underrated. This hire is really a product liaison, working on translating what the customer actually needs into something engineering can build, and fighting internally when the company is doing the wrong thing by the customer.

Technical writing can be outsourced early and brought in-house when volume demands it, but it scales into more than writing. Mature defense sales orgs end up with dedicated proposal managers who know each customer’s house style, and contracting specialists who live inside a single organization’s process, because SOCOM, NAVSEA, and the Air Force each run their own. Saronic’s Navy contracts run through a younger hire who came straight out of NAVSEA. The knowledge of how one shop actually moves paper is specific enough to be worth a dedicated seat.

Most of this advice applies to sales generally across regulated industries, not just government. In energy for example you are often working with long sale cycles and lots of regulation. Understanding these industries well enough to know who the budget holders, end users, and other players are is necessary to landing these sales.

Where to find: Context and relationships in these spaces are not just helpful. They are compounding. In most regulated markets you want to look for someone with strong sales experience who can learn and traverse these industries or raw hustle and genuine curiosity about how they work.

In government sales the profiles that work tend to have either end user experience (from a military or intelligence community) or have been part of the acquisition process (from either government, a prime, or another defense startup). If they come from a prime, make sure they actually won contracts rather than inherited them. If they come from another startup, make sure it was one with real contract wins.

The government affairs component is its own hire and often ties into Policy (more to come later), and the best ones are almost evangelical about their cause. Not people who rotate between clients and causes, but people who have decided that this mission specifically is worth their time and credibility.

How to recognize: The evaluation criteria that work across the first three are the same. Ask candidates to walk through their decision-making process on an ambiguous problem with no right answer. The government sales hire worth making has tenacity, gets kicked ninety-nine times and shows up for the hundredth, has technical competence sufficient to not get caught bluffing by a program office that has seen every variety of overstatement, and the listening instinct to understand what a customer actually needs. Context accelerates. Problem-solving is the requirement.

Policy

“The hardest thing to find in a policy hire is someone who can speak both languages fluently. Technical enough to understand what the product actually does. Washington enough to know how to present it without scaring people. Our first policy lesson at Radiant was that you can be completely right about the technology and still lose the room if you don’t know how to translate it.” - Tori Shivanandan, President & COO, Radiant

If you are a lawyer reading this, please know that I see you, I respect you, and I am a little afraid of you. The policy function is genuinely nuanced, and I recognize the moment I say something definitive about it, someone with a JD is going to tell me exactly why I am wrong. And they will probably be partly right.

The function is often divided under either Legal or go-to-market – or both. Where it lands depends on a simple question: Will a policy failure kill the company? If the answer is yes, often the case for energy and public safety companies, policy work will sit under Legal. If the answer is no, as it is for many defense companies where the primary value of policy is tactical partnerships and access, the function can sit in go-to-market.

Saronic has both. Under GTM, the role is more government relations and looks like lobbying, budget, and relationship work. Under Legal, the work is slower and more technical, focused most often on shaping the legislative landscape so a company is able to operate freely. That might mean securing DHS waivers so an autonomous cargo ship can sail from Puerto Rico to Houston or working with Office of Science and Technology Policy (OSTP) to shape technology policy for the industrial base.

The credibility and reputation of these hires matters most. Three of Radiant’s five executives have PhDs, not because the policy work is purely technical, but because the intellectual firepower required to gain the respect of government experts and regulators demands it. The team is also deliberately structured as translators first: Rita Baranwal’s, Radiant’s Chief Nuclear Officer, earliest lesson at Radiant was not about the regulations themselves, she knows those well, but about how to present the technology without shoving it in peoples faces. Being right about the engineering is not enough. You have to know how to say it and how to get people to believe in the future you are building.

Where to find them: The two functions draw from different pools. Government relations hires tend to come out of the major lobbying firms or from Capitol Hill. The legal and regulatory side follow classic legal paths: big law, maybe a tour through the DOJ or the State Department, often private industry experience.

The right evaluation question: Unlike other roles, context and relationships are not a nice-to-have. They are the job. A policy hire without government experience in the relevant domain, a real network, or the cultural fluency to speak the language of the institutions they are navigating is not set up to succeed. The best lawyers also have won cases where the cards were stacked against them. Asking about underdog scenarios can tell you a lot.

Marketing

“Defense companies fail at communications when they assume complexity is an excuse for vagueness. Our job is the opposite: take a business most people find opaque and make the argument so clear and interesting that the right customer, the right hire, and the right regulator all arrive at the same conclusion on their own.” - Shannon Prior, Head of Communications, Anduril Industries

All startups have to cater their messaging to future customers, talent, and investors. But American Dynamism companies have to tell the stories of complicated businesses clearly. And to a wider variety of talent and regulators, too. The companies that build this muscle early build a flywheel that makes every other hire easier.

The shape of this function can vary from “growth” to “marketing” to “comms” in the early days, but the unsurprising throughline is that creativity is a must. Hiring from either outside the industry or outside of traditional marketing backgrounds can be a hack for this because those hires will, by design, have to get creative.

High-slope, early marketing hires don’t always even need to come from traditional marketing. Cole Jones, who leads Growth (and marketing) at Base Power, is an example of this. He joined Base as one of the first employees from SpaceX where he had gone from an ops engineer to a senior product growth manager, but never a traditional marketing role. This jack-of-all-trades kind of early growth hire has positive side effects. Hires internalize and project a message that isn’t just the mission but, as cheesy as it sounds, the identity of the company. Marketing that’s been developed truly with a company is more authentic and resonates with audiences more.

Where to find: When you hire from traditional marketing backgrounds, the industry to hire from is likely not the one you’re building in. You don’t want to produce communications that sound like every other company in your industry because the messaging becomes invisible to the audience. At Anduril, Shannon Prior, VP & Head of Global Communications, came from Hulu and Mozilla and Jeff Miller, VP of Marketing, came from PepsiCo and Snap. Turns out consumer companies spike in Marketing for a reason!

Anduril’s “Don’t Work at Anduril” did not go viral just because it was clever. It was honest: a blunt message with no attempt to smooth over the hard parts. The people it repelled were exactly those Anduril didn’t want. Base Power’s “Don’t tell your grandkids all you did was B2B SaaS” meets the target candidate where they are, acknowledges the comfortable alternative, and makes the case without overselling it. Both campaigns need a foil and lead with the hard parts instead of hiding them.

How to recognize: Ask what company in the space gets its marketing completely wrong. A good answer comes fast, names names, and gets specific about which audiences are being missed and why. A great answer follows up with what they would do differently as CMO. What you are screening for is immediate conviction, concrete tactical thinking, and a clear point of view. A candidate who hedges, speaks in generalities, or can’t think of anyone is telling you something important.

These foundational hires compound. A great manufacturing leader shapes what engineering gets to bring to reality. A great deployment lead closes the feedback loop that the product team might otherwise miss. A great policy leader has a voice in the room when requirements are written that shape company roadmaps. A great marketing hire makes every engineering role easier to fill.

None of these functions are glamorous. They don’t get the stage time that a successful launch or a breakthrough prototype does. But they are the difference between a company that builds something remarkable and one that actually becomes something. In hindsight, the product was almost the easy part.

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 endpoint security market is entering its third generation. The first was antivirus: static signatures, easily evaded, built for a world of known malware. The second was behavioral detection and response such as SentinelOne, CrowdStrike, and the EDR category they created, which watched process behavior instead of file hashes and became the standard enterprise defense for over a decade. In both of those generations, the assumption was that the endpoint followed a familiar pattern: A human clicks, a program runs, and security tooling watches for the process to misbehave.

That assumption no longer holds. AI agents now run directly on the endpoint with the same privileges as the user they serve, reading files, executing code, calling APIs, and moving data between apps. Microsoft embedding Copilot into Windows made this the default configuration for the modern enterprise. The interactions on the endpoint have fundamentally changed and every control built on the old assumption (EDR, DLP, Zero Trust) breaks the moment it can no longer tell whether an action came from the person at the keyboard or the agent acting in their name.

This is not a gap incumbents can patch. They built durable platforms for the behavioral-detection era, and reworking that architecture to attribute every action and intent to both humans and synthetic agents would mean rebuilding from scratch. Most will add basic AI discovery as a feature, but this new problem requires a complete rethink of what security means in the agentic age.

Neo is built for this new generation from the ground up. Rather than attempting to retrofit yesterday’s approach, Neo is pioneering agentic software control for a world where AI agents act continuously, often outnumbering the humans they serve. Its core bet is that security has to anchor back to a few simple questions at the moment of execution: is this action coming from a real, verified person or is it coming from an agent, is that agent configured with the correct guardrails, and what happens when malicious or negligent activity occurs. Answering that in real time, and enforcing it autonomously, is what the tools out there today were never built to do.

Neo’s founders are as strong as they come in security. Nick Warner scaled SentinelOne through its IPO in multiple executive leadership roles. Shlomi Salem spent over 11 years leading threat research at SentinelOne. Eran Shirazi brings deep technical grounding and prior experience co-founding and serving as CTO of EasySend. Together, they combine the commercial instincts, security depth, and credibility with top-tier security talent that this category demands.

The shift from antivirus to EDR created category-defining companies. We believe this new shift is at least just as large, and arguably more urgent. We’re proud to have led Neo’s Seed round and to partner with Nick, Shlomi, Eran, and the rest of the Neo team as they build the security layer the agentic endpoint requires.

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 we started thinking about people who could help build the next decade of American Dynamism, and talking to founders, one name kept surfacing. Founders building the hardest things in defense, space, and manufacturing would mention Connor Love, and they would tell us the same thing over and over: he understood their business better than any other investor on their cap table. Not the version that fit neatly into a deck, but the one they were actually building.

That kind of reputation is hard to earn and impossible to fake. It comes from doing the work. Long before a deal, Connor would get on calls with founders every few weeks to stay current, offer perspective, and build a real relationship, and to make sure he was always on top of how a company and its market were evolving. He shows up because the mission is personal to him.

Connor is a West Point graduate and a former active-duty Captain in the U.S. Army. In 2019 he was set to start business school when he chose instead to deploy to Northern Iraq, working alongside the joint military community and our Kurdish partners. He has led soldiers, advised senior commanders, and seen up close what motivated, purpose-driven teams can do under real pressure. He started in venture six years ago with that experience behind him, and it shapes how he sees companies and the people building them. He’s spent the years since becoming one of the sharpest American Dynamism investors of his generation, backing founders in defense, space, manufacturing, and autonomous systems.

What convinced us wasn’t the track record, strong as it is. It was that Connor deeply believes the same thing we do: some of the most important companies of this decade will be the ones serving the national interest, and that serving the country and generating extraordinary returns are not in tension. They reinforce each other. He wants to help build a firm that backs those companies for the long haul, and he’s joining Andreessen Horowitz to do exactly that.

The stakes here are only getting higher. The next decade will be defined by whether America can build fast enough to stay ahead, in defense, in space, in manufacturing, and in the autonomous systems that increasingly tie them all together. We believe American Dynamism is still early, and that many of the companies that will matter most are being started right now, often in areas the rest of the market has been slow to take seriously. Backing them takes conviction and patience, a willingness to stay in through the long and unglamorous middle where real companies are built.

On behalf of the entire American Dynamism team, I’m thrilled to welcome Connor Love as our newest General Partner. He’s written his own note on joining. You can read it on X.

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

Software’s Selective Sell-Off

The trials and travails of publicly traded software companies mostly continues apace, but as we touched on before, the current state of affairs is not an undifferentiated ‘bloodbath.’

It’s also the case that the software sell-off isn’t so much a reflection of present or recent performance, but rather the market’s collective doubt as to whether software cos can keep it up, over the longer-term:

Multiples on next-twelve-months free cash flow are at or below 2014 levels. In other words, investors are assigning their lowest premium on software’s cold-hard-cash generation in over a decade.

While the “market” never really speaks with one mind, the simplest explanation for the discount is that the market does not think the next twelve months will reflect the next twelve months after that.1 It’s an AI disruption story, of course. Performance is fine now, but how long can it go? After all, print media stocks traded down well before their doom was reflected in earnings, and investors were surely right to be early (so the story goes).

The point being that even if the evidence for the death of SaaS is a bit thin right now, investors are paid to have a view of the future, and the future is (in their view), a bit gloomy.

Gloom is not the outlook for all software companies, though. Again, differentiation and discernment is increasingly the name of the game. You can see it a few different ways.

First consider the gap between the median, top- and bottom-quartiles of the IGV’s performance:

Over the last 30 trading days, both the top quartile and median IGV stock have outperformed the ETF as a whole—it’s the bottom quartile (which include some of the largest companies) that’s pulling down overall performance.

Compare that with a yearly lookback, and it’s mostly a different picture:

Over the past year, the median tracks the overall ETF much more closely—it was only around the turn of the calendar year that performance quartiles began to disperse (with now a ~50pp spread between the top and bottom).

Returning to the 30-day view, it’s also clear that fundamental performance is not strictly speaking the driver of recent stock performance:

From a revenue growth perspective, there is basically no correlation between growth and recent performance (although, to be fair, a lot of that is explained by changes in growth, rather than levels). When broken out by performance quartile, the same pattern emerges: there are lots of software companies clustered in the high-teens low-twenties of revenue growth, but performance is stacked almost vertically, with some of the fastest topline growers (and largest companies in the ETF) in the bottom quartile of performance.

Rather than growth, what matters more is perceived longer-term durability. Here’s the same chart broken out by sector:

While it’s not so easy to cleanly categorize every company, the top and second quartile performers tend to be dominated by (1) cyber/observability; and (2) vertical SaaS. The bottom two quartiles, by contrast, are some combination of horizontal SaaS, cloud/infrastructure, and the “other” category, which includes marketplaces, ad tech, and point solutions (and whatever MicroStrategy happens to be).

The point here is pretty clear. Rightly or wrongly, the market has perceived a clear difference between software companies that seem to have some degree of AI defensibility and/or tailwind (and those that don’t).

• For cyber and observability, AI is expected to increase buyer urgency (and incumbents have a trust premium that AI upstarts will find harder to disrupt).

• For vertical SaaS, specialized information around workflows, data and customer relationships is considered a barrier-to-entry for AI challengers (while the same cannot be said of horizontal platforms).

For everyone else, though, the message is pretty clear: software alone is not a moat.

Years of sticky and accumulating ARR, huge feature-sets, and widespread adoption are just that—yesterday’s news. It very well may be that a few quarters of steady and/or improving growth will flip the script, but for now, the market is unimpressed and unconvinced.

AI Demand Growth > AI Spend Growth

You may have heard recently that there’s a perceived shift from token-maxxing to token-optimizing, and it’s a big deal.

The gist of it is as follows: while frontier models are increasingly expensive to train, the marginal benefits of those frontier improvements may not be enough to persuade customers to absorb the additional cost. Rather than maxxing out on the latest and greatest tokens, customers may instead downshift where they can to cheaper, less-powerful models (including open-weight models).

For some people, they see that shift as calling into question the sustainability of frontier model development—if customers won’t support the cost of the latest and greatest, then how can the labs? Fair enough, although the objection is somewhat ironic, if only because the supposedly rapid obsolescence of non-frontier models was also supposed to threaten the sustainability of frontier model development . . . and now that non-frontier models are increasingly not-obsolete, that’s still apparently a bad thing. OK, alright.

Without getting too far into the merits of that debate, we’re only here to observe that it is indeed true that the cost of intelligence is falling rapidly, and that it appears to have positive implications for AI demand. In other words, the Jevons-ish dynamic continues to hold: the cheaper AI gets, the more people/firms want to use it (for more things).

Increasing the surface area for demand is exactly what you want to see.

Let’s start with some precedent from another innovation that started concentrated and expensive, but became diffuse and cheap: computers. In comparison to the previous tech-leap forward driven by PCs, when it comes to AI, the cost of intelligence is dropping much more quickly:

It took PCs almost two decades to achieve the affordability gains that AI has achieved in ~3 years.

That’s a remarkable trajectory, and as with computers, it seems like falling costs are helping push AI demand up-and-to-the-right (with a lot more runway to go).

On the consumer side, according to data from PNC Bank, paid penetration is still tiny, but definitely growing:

Both the share of households paying for AI, and their average monthly spend, continue to climb—monthly spend has increased even more steeply, rising ~25% since the beginning of the year.

They’re obviously not perfect comparisons, but for some perspective, in 1997 only ~45% of adults aged 35-54 reported owning a PC (decades after computers had been released for commercial use, mostly on the enterprise side). Currently, the share of households with a computer is ~90%, and closer to 97%, if you include smartphones—the point being that mass market adoption can take some time, and it scales with both utility and cost.

Subscribe now

There are, of course, no free versions of PCs (as there are with LLMs), but at only 2.2% paid household penetration (and with bang-for-the-buck increasing exponentially), it seems reasonable to assume that wider consumer adoption still has a ways to go—even as the average price point rises, albeit not as quickly as intelligence.

There’s other data, as well, that shows demand for AI rising in tandem with cost-efficiency.

According to YipitData’s analysis of OpenRouter data (which measures only a subset of overall token consumption), frontier token usage continues to grow, at the same time as open-weight tokens continue to rapidly gain share:

The “Asia-based provider” share of total tokens, which is a proxy for open-weight alternatives, has grown to ~60%, a three-fold increase from the beginning of the year. OpenRouter’s sample is probably at least somewhat biased towards open-weight users, but that’s certainly consistent with the price-differentiation story, and a Jevons-ish one, as well.

Similarly, according to OpenRouter, while both tokens-per-user and spend-per-user are growing quickly, the former is growing much more quickly than the latter, since the beginning of the year:

Again, with the caveat that OpenRouter only sees what it can see, this pattern is pretty consistent with a Jevons setup (where demand grows concurrently with the rapid cost-efficiency gains of AI).

For now, at least, it seems pretty clear that as intelligence gets cheaper, it pushes the demand curve even further up-and-to-the-right. Cheaper tokens from sub-frontier models have certainly gained share, but the net-effect has been to drive overall spending exponentially upwards. That’s hardly a doomer story.

Doomers might still claim that open-weight models are cannibalizing the frontier, but the alternative scenario where efficiency gains had no apparent effect on demand and/or were driving aggregate spending down, would actually be much closer to a doomsday scenario. Jevons, by contrast, is exactly what the bulls are hoping for.

Entry-level Tailwinds

On the subject of AI demand, while AI is supposedly not useful enough to generate meaningful ROI, but also so useful as to eliminate all the jobs, we’re pleased to report that there continues to be little to no evidence that AI is, in fact, having much if any job-killing effect.

Indeed, even the one soft-spot in the labor market (i.e. entry-level hiring) has some recent wind in its sails, and if anything, AI appears to be helping, far more than hurting.

First, as per Revelio’s data, summer internships for 2026 are tracking much higher than years prior:

Internships aren’t the same thing as jobs, but they’re at least some semblance of a demand signal for the young and restless, and 2026 cycle is way ahead of ‘25 and ‘24 (if not quite at ‘23).

Even better than that, pay growth for teens and young adults also appears to be on the rise:

According to ADP’s data, wage gains for young workers (16-24) have rebounded off their 2025 floor. Wage gains are almost certainly a demand signal, and if wages are going up, one can reasonably surmise that entry-level employment prospects are going up, as well.

With respect to AI’s impact (if any), there the story is even better. If anything, AI appears to be accelerating entry-level hiring in a big way:

Per an analysis of Revelio’s job data and Ramp’s spending data, “High-Intensity AI Adoption” corresponds to a ~6pp increase in entry-level headcount, two years after adoption. “Low-Intensity” AI-Adoption, by contrast, corresponds to a ~.50 pp decline.

Now, there are few ways to interpret this, from “AI is creating entry-level hiring,” to “AI adopters are the growth companies and therefore of course they’re the ones doing the hiring,” to “Ramp’s data may not be a representative sample of the economy writ-large, and therefore it’s hard to conclude anything at all.” That’s all well and good, but one interpretation that almost certainly does not follow is that “AI is killing entry-level jobs.”

Perhaps one day (although, again, there is every reason to think otherwise), but not today.

In general, while much was made of post-ZIRP decline of “AI exposed” jobs, there has been comparatively little discussion of the fact that “AI exposed” jobs have recently been leading the comeback charge:

As per the Indeed Hiring Lab, since May 2025, when it comes to job openings, the greater the AI exposure, the greater the recovery. That’s especially true for software engineers, where job postings have increased by ~15%, in comparison to an overall ~7% decline (measured from the release of Claude Code).

In the bigger scheme of things, the reality is that it’s too soon to tell, but to the extent that anyone is inclined to declare that ‘AI is causing human obsolescence,’ the data simply is not on their side. If anything, it’s quite the opposite—AI-exposure seems to positively correlate with job-growth.

Now, to be fair, it’s unclear how meaningful the “AI exposure” category actually is because there’s very little agreement on what or who is exposed (and by how much):

It turns out that the higher the average “AI Exposure” for any given occupation, the more disagreement there is with respect to the degree of exposure.

Unsurprisingly, when it comes to predicting the future impact of a novel technology, reasonable minds can disagree.

There are some exceptions, of course. Everyone seems to agree that proofreaders, insurance underwriters, statisticians and, funnily enough, economists, are all very AI-exposed. If AI does eventually disintermediate human economists, will it cop to the crime or will there be no economists left to tell the tale?

Data Centers Make Energy Cheaper

Earlier this week, New York State’s Governor Hochul announced that she was imposing a 1 year moratorium on data center development. Among other things, the stated goal of the moratorium is to “protect [electricity] ratepayers . . . as data center development threatens to hike up utility bills.” Without intending to sell her short, the Governor offered a bit more color about her reasoning on the Odd Lots podcast, so check it out.

That said, it’s still a curious pronouncement.

While it’s certainly true that data centers are increasing demand for electricity, the evidence suggests that, if anything, data centers are actually helping reduce the cost to ratepayers:

Based on research from the Berkeley National Laboratory and The Brattle Group, at the state level, price-increases for electricity are negatively correlated with increased demand. Consume more, pay less—strange but true.

Indeed, over the past 6 years, the states with some of the steepest load growth (and the most data center development), like Texas and Virginia, have seen little-to-no price increases, at all. On the other hand, states with the steepest price increases (and very few new data centers), like CA and NY, have experienced a decrease in load-growth—California has reduced electricity demand by ~5% while increasing prices by more than anywhere in the country, and ~33% more than the next closest offender.

Not only that, but the inverse relationship between energy demand and energy prices appears to hold in Europe, as well:

As per the EIA, with the exception of Ireland, the EU countries with the largest increases in prices, also had the largest decreases in demand for electricity.

Again, consume more, pay less. It’s counter-intuitive, right?

We all know that all things being equal, more demand should increase prices, not the other way around (as per Gov. Hochul’s declaration). But when it comes to the grid, the story flips. The reason that demand is inversely related to cost is that the grid tends to benefit from economies of scale: the more electricity demand there is, the more distributed fixed infrastructure costs become, leading to an overall drop in prices for ratepayers.

In other words, rather than “protect rate payers,” Hochul’s Data Center moratorium may well have the opposite effect: electricity prices could rise much more so than if data centers were there to help carry the fixed costs of the grid. Perhaps that relationship will not continue to hold, but that’s been the story to date.

Energy costs notwithstanding, there is also the broader economic impact to consider. The reality is that without AI-related infrastructure investment, there is very little investment growth, at all:

Tech-related investments are the only growing categories of investments, so a moratorium hits right where it counts the most.

Surely the Governor has her reasons, but a policy that may both (a) increase electricity prices (as part of an effort to “protect ratepayers”); and (b) deprive NY State of the most important investment tailwind in the game, is certainly hard to figure.

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

AI models have required rethinking the GPU stack. And now agents are doing the same thing for CPUs.

Ironically, the direction compute has taken over the last decade has made it less suitable for AI. The driving forces have been focused on human developers pushing platforms to offer higher levels of abstraction, much more opinionated approaches to state management, and heavily focused on devX.

But it’s increasingly clear that AI agents just want a computer. By that I mean a complete OS install, fully stateful, and compute that can run locally, or in the cloud. And agents need a lot of compute. In fact the recent upsurge in agents has resulted in a massive CPU shortage alongside the model-induced GPU shortage.

So similar to GPUs, now it is time to rethink the execution layer for agents. And we’re extremely excited to be investing in Runta which is tackling this problem head on. In short, they’re rebuilding the execution layer to provide the abstraction agents want (a computer), but be extremely efficient, run locally or in the cloud, and yet provide security and policy interfaces allowing agents to be run safely within the enterprise. Or anywhere.

We’ve known Guanlan, the CEO of Runta, for nearly a decade. He’s a legend within the systems community. He was a tech lead on Cloudflare’s early edge team that built out the edge proxy, caching and WAF. And then we worked together at Kong where he led the team that built the core proxy. He’s put together an ultra-strong team of systems developers to rebuild compute for agents. To be clear, this isn’t yet another sandbox cloud, which only solves a small portion of the isolation problems. But core systems software re-imagining the execution layer the right way, purpose built for agents.

We’re leading the seed round. I’ve been in the industry long enough to remember the transition from client/server to datacenters. And from physical computers to virtual machines and then again to containers. This shift in compute is the largest and most important yet. From hosting software to hosting agents. And Runta is the right team to get us 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

To hear more from Tori, President & COO of Radiant, follow her at @torishiv

A few years ago, John Salvatier wrote an essay about building basement stairs. He expected a simple weekend project. Instead, he found that wood warps after milling, that a screw won’t enter a hole at a different angle than it first did and there’s no such thing as a straight board. His conclusion: “…that reality has a surprising amount of detail.” And, I couldn’t agree more.

I still consider myself an industry outsider, but in my years working in nuclear, there are ‘before’ times, and there are ‘after’ times. The Executive Orders invigorated the industry. After decades where nuclear barely moved, our industry is being asked to move faster than it ever has. And that transition is exactly the moment when the small details matter most.

We are building microreactors at Radiant, and I can tell you that the small details, the ones that often seem inconsequential — are the ones that prove your product and de-risk it for your customers. The thing is, you have to look for these details, question everything and test continually. Our mission is to be an energy company that powers the global economy everywhere on earth and amongst the stars, too. To get there, there are small steps, and there are big steps, and we can’t assume that if we can do A, doing B will be easy. Overconfidence is when reality teaches you that details matter.

Building a microreactor is new territory and the decisions we make at every step have consequences. This is why we challenged ourselves to go full power, full temperature with our commercial product in our testing at the Idaho National Lab DOME this summer…because our job is to find the details our customers need us to get right.

About those details…

A bolt is just a bolt, right?

My dad built homes. So did most of my friends’ dads. By the time I was a teenager, I was comfortable on a construction site and knew that nothing about building is as simple as it looks from the outside. I carried that into the formative years of my career at McMaster-Carr — a company solely obsessed with the details of industrial parts. When I look at our industry’s current moment, I see the same trap John Salvatier saw on his stairs, just at a much larger scale.

Consider a bolt. It’s a simple object — until you need the right one.

When I was at McMaster-Carr, we sold 91,130 distinct products called “a bolt.” Not one of them, on its own, is necessarily the right bolt to hold the lid on a nuclear pressure vessel. So, let’s design the right one for a microreactor together and see how long it stays simple.

The first thing is to choose: is it a standard bolt, or custom bolt? But, you don’t know yet. What strength does it need? Also unknown, that depends on temperature, because steel weakens as it heats, and our reactor is hot — upwards of 600 degrees Celsius. It’s not just the heat though. Hold metal under load at high temperature long enough and it doesn’t just sit there — it slowly stretches. The phenomenon is called creep, and it means the bolt you torqued perfectly on day one is a different length by day one thousand. Imagine that bolt is sealing your pressure vessel. How long does the lid need to stay tight? Ten years? Twenty? It’s on us to determine and these decisions have real world consequences.

To fight creep, you need to dig into metallurgy and reach for a high-nickel alloy. But steel will relax over time, so you test it by stretching to an extreme preload, high enough that even after years of relaxation it will still hold tight. But you can’t reach that preload with a wrench and a strong arm, you need a hydraulic tensioner. Do you see what’s happening? That bolt now means you have tooling.

And there’s more to consider. The threads, you know the swirly part on a bolt, aren’t cosmetic. Cut them too fine and they get destroyed under that preload. Run two threaded metal parts together at that load and they can gall — weld and tear against each other as they move. That means you need a lubricant. Which one? Real questions, real consequences.

We’re still talking about one bolt and there are even more complexities with this ‘simple’ part. We haven’t gotten to the reason this is a nuclear bolt at all: it sits next to an operating reactor core. Imagine 1 trillion neutrons per second acting on every square centimeter of your bolt. Engineers didn’t guess at what was happening – they measured it, starting with reactor materials testing back in the 1960s and 70s. Their work found neutrons flying through the microscopic crystalline grains that make up the high nickel alloy and knocking atoms out of place in a process called Neutron Displacement Damage. This lowers strength and causes physical swelling. And the displacement rate is slightly different on the nickel alloy versus the material around the bolt, which means you need more spacing between each.

That’s one bolt. More than a dozen consequential decisions, and I left some out. Then we move on to who we buy the part from. We need to be confident that they will produce our bolt exactly, repeatedly, and at scale. As Salvatier put it: “At every step and every level, there is an abundance of detail with material consequences.”

We have 56 bolts on the top of our reactor pressure vessel alone. And we’re still only talking about bolts.

The more difficult your mission, the more details are important to understand.

Look at the nuclear industry from the outside and it seems like we’ve already won — positive headlines, milestones that seem like progress to the general public, policy tailwinds and more. But this is exactly why we have to be vigilant because apparent momentum makes it easy to sit on your laurels and brush aside details. Not through laziness or disinterest, but because it seems like things are going well, that the milestones are stacking up and the hard part must be behind us. The details ahead get quietly dismissed because it seems like perhaps the future is inevitable- and we’ve mostly arrived, right?

Wrong. I ask you to consider, if one single bolt has so many material consequences, what about the parts or designs we haven’t built yet? The distance between our prototypes and products? This is why we test and validate everything and then turn around and test it again. We have to search out the details. All of them.

Regulatory approvals have moved quickly lately, but not because the regulator cut corners. They moved quickly because the systems being evaluated have not brought that much detail to surface. Fast approvals of simpler systems tell us almost nothing. Our regulators are aligning with an industry that is quickly advancing and I respect their technical expertise. I would vehemently defend against any claim that DOE or the NRC are skipping steps. They’re ensuring the rules don’t get in the way of us being safe. And, this is also why we are testing at full temperature and full power. We need to performance test — and so does our regulator.

If you’re trying to do something that has never been done before, you can’t assume the details will sort themselves out. The more difficult your mission, the more the details are critical to understand.

You find the details by running the thing.

This is exactly why we built our five-phase test campaign at the Idaho National Laboratory DOME facility. It’s also why we're so focused on full-power, full temperature testing. Testing a reactor isn’t a single day event. It’s a multi-month, heck, multi-year process, and it’s the same attention to detail that iterating on any good product requires.

Our five phases are designed to de-risk a commercial product – one that will be delivered to our customer by 2028. That’s 18 months from now. The first two steps are table stakes: ship, install, and fuel Kaleidos at Idaho National Laboratory’s DOME facility, and then go zero-power critical. That means the reactor is ready to turn on but has no meaningful power. The third step is bringing the reactor above 1 MW thermal to validate the high-temperature range and the reactor’s intent, making it a viable product. Fourth is full temperature and power, proving we can generate enough heat to one day become usable electricity end to end. Fifth is 150 continuous hours, our first measure of reliability.

Although the system is built for it, we don’t leap to full power, because each phase is designed to provoke specific behavior and measure it against what we predicted. Where they disagree, a detail was hiding. The hold points between phases exist so we stop and understand why before asking anything more of the machine. That pause is the structural refusal to assume the hard part is behind us.

The bolt you need for steps 1 and 2 is considerably different than the bolt you need at steps 3, 4, and 5.

Don't spend time justifying what you did. Look for the details you can’t yet see.

The momentum in this industry is real. The policy, demand, talent, and public goodwill are a huge tailwind, and I am deeply grateful to those who’ve made it possible. But this momentum can also hide the details. We still do not have a domestic HALEU supply that doesn’t rely on geopolitics we can’t control. We still have not put a power producing reactor all the way through the reformed regulatory process and our customers are still reactor-less.

So let me be clear about what winning actually looks like. The US military and investors both deeply care about winning and for them, it isn’t a ribbon-cutting. Winning is the hundredth reactor, and the thousandth, every bolt the right bolt, every pipe, sensor, and line of code built so reliably that nuclear becomes boring instead of a spectacle. It’s why we have booked the INL DOME for a full year. And this is why we know we will deliver our first reactor to our customer, get run time, and then get back on the wheel to improve, again, and again, and again.

We are closer than we’ve ever been — further along than the headlines suggest, and further along than most people realize. It truly is in the details and we are detail obsessed.

That’s the company Radiant is. Everyone has their own view and biases, but our judgement is that the reactors that actually get built — the ones that go to the datacenter, the hospital, the base and the disaster zone — will be built by the team that never stopped looking at the details.

Salvatier closed his essay with a line worth keeping close: if you wish to not get stuck, seek to perceive what you have not yet perceived. For us, that means: don’t take anything for granted. Find the details you cannot yet see.

If you wish to not get stuck, seek to build what you have not yet built.

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

This week the University of Chicago Law School announced a new policy regarding the use of AI in teaching law. The headlines read “AI strategy bans phones, laptops in class for first year students.” What does history say about this sort of action so soon in the evolution and understanding of a new technology?

Whether this is the right thing for a short time, a new long-term approach to knowledge tools, or a step backward is the subject of much debate. In the short term, it is probably fair to say the general concerns over the rise of AI in terms of jobs, replacing humans, and current limitations of the technology (to name just a few debate topics) mean this change is seen broadly positively. This is especially true because it is from a thought-leading intellectual powerhouse.

It might be a surprise to some, but almost 45 years ago Harvard Law School banned the use of the first truly portable computers. Students were bringing (lugging) the 25lb (11.5kg) Osborne 1—the computer I learned to program on—to exams to type and then print their answers. There were no caselaw databases, no internet (didn’t exist!), no modem, no battery even...just a loud grinding tractor fed dot matrix printer and the first version of WordStar on a 5” screen and extension cords. Other students used a typewriter, probably the one they got as a gift for graduating college.

Two students brought computers to the exam in Alan Dershowitz’s criminal law class. Mid-exam a note was passed to the Dean of Students. The brave computer user who sought no special permission to use a computer—not because they were a hero or rule breaker but because using a computer seemed as obvious as using a calculator in math class—instead of a typewriter was summoned to the Dean’s office. In a follow-up meeting the student was given permission to use a computer while a hearing was scheduled. The primary concern it seemed was the anxiety the computer created for other students with regard to fairness. The hearing was months off.

First the story was written about in the Harvard Law Record. Then the Wall Street Journal picked up the story, March 23, 1982, with the title of the front page story quoting one of the computer users “Will Computer Memories Replace Notes on the Shirt Cuff in Exams?” Then, Time Magazine covered the story:

This was huge news among a tiny set of very interested people, myself included. Computers were new and scary, but the world was optimistic. Time Magazine had yet to announce the computer as “Machine of the Year” (December 1982) but the world was abuzz with the potential of a computer in every home and on every desk. Atari sold twice as many computers as the next leading company, Radio Shack, both more than either Apple (279,000 units) or IBM (240,000). Literally almost no one had a computer with about 1.8M units sold collectively worldwide since 1977.

While the initial concerns were about a potential unfair advantage for students with computers—something that will later be known as the digital divide—some began to speculate about longer term concerns of computers providing a form of cheating such as pre-written text. Interestingly the idea of connecting the computer to a database of other text or case law resources was simply speculation at this point. It is worth noting that just a year later it would be routine to use a dial-up service to gain access to databases like legal cases as I did with my Osborne.

I was in high school using my Osborne when I first learned about this story. It appeared in the #2 issue of the magazine dedicated to Osborne users. Back then every computer had its own magazines. The story detailed with photos the whole Harvard experience. Below you can see the whole story. It is just too wild to read today. Deans and university presidents getting involved. Banning the use of a tool. Fear. Anxiety. It all seems too absurd over a computer.

Oh wait…this time is different.

In the end the school banned the use of computers and word processors from exams. The brave student told the Dean “I opened the door to 21st century technology and they slammed the door in my face.”

Seems crazy. I know.

Regardless, students increasingly used a computer in preparation of course notes, outlines, and assignments, though not in exams or take-home exams. In just a few years, the computer was ubiquitous in law school. By far the most interesting development at first was just how much students were typing compared to how recent graduates were practicing law where typing was the first thing you stopped doing as a lawyer.

Business school was the same way but instead of word processors it was spreadsheets. MBA students began using Lotus 1-2-3 in droves years before the banks and consultancies they would work for would make worksheets ubiquitous. The most tech savvy students were using VisiCalc on Apple ][ first of course.

My own college experience looked a lot like what I read about at Harvard Law. I brought my Osborne to school. I had the computer because my father had bought one to use for his wholesaler, even though he had no knowledge of computers whatsoever. It was a wild fluke that I ended up moving from the warehouse to the front office to figure out how to make use of it.

My freshman year (1983) had a required writing course. I just assumed I would use my computer as I had to write my admissions essay and even some papers in high school. I had the only computer in my dorm of nearly 100 students where Smith-Corona typewriters were the ubiquitous tool. To use a computer at Cornell meant slogging up the hill to the “terminal room” and using a connection to the IBM 4300 mainframe which was housed at the airport. When I set my Osborne up on my desk and dialed into the mainframe—no small feat—I was able to amaze my new classmates during orientation week.

I was so anxious having read the story about Harvard I actually asked my prof if I could use a computer before the first class even started. She was very excited to tell me that the writing program was embarking on an “experiment” (her words) to “test the quality of writing that students produce when using a word processor compared to a typewriter” and our section was deemed one of the sections that would use a word processor.

I kid you not. In my limited world view computers were already ubiquitous. I had computer books and magazines. I sold computer software I wrote the summer before arriving at school. I was in a computer club filled with NASA engineers working on the Space Shuttle. I was dialing up to FIDONet bulletin boards and downloading CP/M utilities. Literally everyone *I* knew used a computer.

But Cornell, like most every school, was skeptical. The first week of our writing class the dozen of us in class took a field trip from our 19th century, wood-paneled seminar room to a tiny and sterile closet of a room that housed two Wang “word processors.” These were computers from a then famous Boston-based company that did one thing and one thing only, which was word processing. They cost $10,000 and had monochrome portrait-oriented screens with elaborate keyboards featuring rows of dedicated function keys for word processing capabilities. There was what was called a “letter-quality” printer which was like a Selectric typewriter with no keys. To use one we had to buy our own 8” floppy disks at the bookstore.

It was then my turn to meet with the Dean. When I asked to use my own Osborne in my room I had to seek permission from the Dean who was overseeing this experiment. Would using a computer invalidate the “data” or give me an advantage? A demo was impractical but I was asked to bring a sample printout to see if the prof could read it. I was told I needed to achieve letter-quality output. I called home in a panic and using my magazine knowledge I secured a Royal electric typewriter that connected to the Osborne with a 2” wide ribbon cable from 47th Street Photo in NYC. It sounded like a strafing A-10 Warthog when I printed a paper. It was impressive because it basically “typed” faster than any human.

Subscribe now

My classmates using the Wang constantly complained about the fact that the two word processors were oversubscribed by all the experimental sections and to use one required making a reservation. What the experiment failed to consider was that students would actually write at the computer not simply type. They originally thought students would use paper and pencil for a draft then reserve time to type the result in. Using the word processor was just a fancy typewriter.

The process of writing with a typewriter varied depending on whether it was an essay, report, or research paper. Generally one had handwritten notecards that could be sorted and organized. Following that, a hand-written outline was created. From that outline a draft was written out longhand. Some assignments required going over that with the prof or TA. Corrections were made with red ink on that draft. Then a final draft was typed from that. As the semester progressed and deadlines approached, many of those steps got skipped. It turns out those analog tools were ripe for misuse as well.

This was a really important lesson. The people controlling things had no idea how the actual “work to be done” would change because of the tool. Writing on a typewriter while difficult was also expensive. One could only correct small mistakes, and awkwardly so, with whiteout or just a pen. Once you removed the paper from the typewriter aligning any fix was impossible. A typewriter was not for writing or thinking. It was simply better handwriting.

I had no paper drafts. I was using a word processor to outline, write, rewrite, arrange, spell check, and more. I was somewhere between annoyed and puzzled at the experiment being done. I wrote my entrance essays using a word processor. In fact, it soon became provable that using a word processor freed a person to—potentially—be far more productive and produce much better output.

As it would turn out so would everyone else. While home for winter break January 1984, a tv commercial ran during the Super Bowl. While we thought of Orwell and Van Halen as 1984, the commercial Apple Computer brought a new meaning to the year. The incredible marketing and distribution plans from Apple made getting the new Macintosh on campus vastly easier than acquiring a computer at home and the relative portability was a dream. They were cheaper than an Osborne and came with a free word processor and paint program. Dozens of my dorm mates got one. Working with Apple (and the brilliant Dan’l Lewin, who recently stepped down as CEO of the Computer History Museum and was one of the few people to successfully work at a high level for both Bill Gates and Steve Jobs in his career) Cornell put Macintoshes in several labs and public spaces. There were probably 200 Macintosh computers available for the Spring semester.

The writing program experiment no longer mattered. I don’t even know if they ever filed a report and I can find no evidence of one.

Reality rolled right over testing the theory. Even the need for “letter-quality” output was relaxed as ImageWriter dot-matrix printers with tractor fed paper were standard. Late night on Thursday or Sunday you could sit in the courtyard between our dorms and hear the ever-present grinding of those printers as students put the finishing touches on their assignments. The tractor feed edges of paper would be strewn throughout the dorm the next morning. Every artistic and handwritten sign and announcement hanging on cork boards around campus was replaced with a “ransom note” of outline and shadow printing by an ImageWriter.

Yes I said “reality rolled over” on purpose. The logic of “preemptive regulation” or “presumptive negatives” is what incumbents use to slow technology adoption. They do so using the language of fear or appeals to morality to just slow down and consider... The problem is you really can’t consider a technology that is already out there. You can’t put it back in a bottle. You can’t make people stop using something they like out of concern without making a case that the mere worries negatives provably exceed the realities of the benefits. The ability to express concern does not replace the need to demonstrate harm. Preemptive regulation is the enemy of progress. We would definitely not have had the 20th century with such a mindset.

I know for many there is a visceral reaction to this. The government should protect us. Industry should not move forward until it can prove something is both good and safe. The easiest algorithm I have for this is to ask if some new technology will be used by the government alone before it is approved or neutered for use by everyone else. If that is the case then history tells us the technology is being used for control or even oppression first and foremost. I do not intend to be political here, but practical. In the case of AI, the connection of AI to freedom of expression is unshakable and irrefutable.

Make no mistake, students were still writing shitty papers. They just lacked spelling errors if they used the right word processor and they might have had some bold and italics. An important lesson was being learned—one that matters for today—which is for all the fears that the tool would make things too easy and take away the process of writing, the reality was that there were those that used the tool effectively and those that did not. The mistakes were new. There were weird typos. There were duplicate paragraphs from errant block moves aka copy and paste. There were loads of half-finished sentences. These were all things that you didn’t see when writing a paper meant transferring a hand-written draft to a typed final draft, which is all one did with a typewriter.

In my opinion based on my own real world experience leading up to that point, this new normal diffusing across campus was completely obvious.

Over the next three years of college, the entire curriculum and teaching process were upended. Everyone became connected, not just those in the computer science department. Every professor used a word processor for preparing papers. Every numbers-oriented department was using Excel on Macintosh. Every engineering department was using Macsyma (or μMath) and early design software on PCs. Databases that were kept on mainframe cards and tape were migrated to dBase or File. The first connected computers over BitNet allowed us to “chat” in real-time with other students.

This wasn’t just a university thing. In the four years since Harvard had temporarily banned computers the working world had begun to acquire IBM PCs and use word processing, spreadsheets, databases, and soon “presentation graphics.” Most everyone technically inclined was using a PC for their internships and all of us had stories to tell about how we impressed our bosses with cool computer tricks. In my job I turned around an idea my boss had for an inventory report on Javelin handheld anti-tank weapons and Pershing Missile repair histories using Lotus 1-2-3 and CrossTalk that had been backlogged with the COBOL team for months.

We were in a new era. The optimism for the possibilities drowned out those that thought of these tools as shortcuts, cheats, or unfair advantages. It was obvious by 1985 or so that not using a computer was a disadvantage, in academics or business.

This all has me thinking about the University of Chicago. While many immediately came to the defense of the wisdom of the school, my experience put me right back talking to the Dean about using a word processor. What seemed obvious to him seemed like a risk to teaching and learning. The short term accolades that came from pacing the introduction and experimenting with the right way to use the technology were basically steamrolled by reality of the “native” generation.

That is what is happening right now with AI. The generation entering law school and frankly those entering the profession are AI natives. They are going to use AI no matter what anyone says. Trying to stop them is like trying to stop a new banking associate from using 1-2-3 in 1983. Not only is it impossible but it makes little sense. This is not about not having rules—for example the rules for cheating, stealing, hurting people and more all still exist and are in full force. This is about realizing that freedom of expression and use of information tools are intrinsically linked and controlling information tools is akin to controlling expression. Again, this isn’t political, just practical.

Many (many!) will dismiss this assertion—and that is all it is—immediately because as we know will be said, “this time is different.” The narrative goes something like AI is different. AI is not simply faster typing. AI replaces the whole process of writing (or analyzing or learning or reading or whatever).

There is some truth to this. But you cannot say those things about AI without recognizing that is precisely what was said about the rise of the PC. I sat at Thanksgiving dinner with some relatives who were first year banking associates who had gone to a fancy MBA program that had not yet adopted 1-2-3 or VisiCalc. Their view was that banking was about knowing what questions to ask and then getting an intern to use a computer to work out the answer.

What they didn’t see was that the new tools were not about automating the answer to a question but about taking a problem and turning it into 100 variables, analyzing the interaction of those variables, and simulating thousands of potential questions. The idea of “solve for” completely changed how banking operated day to day. Modeling was the new thinking. And yes, if you had bad assumptions or bad formulae then you would be in a lot of trouble. The expression GIGO was already hanging in every computer operating room for a generation already—Garbage In, Garbage Out. I learned that from my NASA buddies at computer club.

If that isn’t convincing, this same logic was said about the rise of the internet. Many critics of the internet (and I’m not talking about Paul Krugman in 1998 saying the internet would amount to nothing more than a fax machine by 2005) were saying that the internet with its ability to find the exact right answer would replace thoughtful research and consideration. It would replace output with theft. The internet would make us all lazy at work.

No one in 1995 saying those things was predicting the true impact of the internet. Not only would it fail to have those negatives so adamantly declared, but the positives came in entirely new ways. Jobs were not eliminated but unleveled. People did not become lazy about doing new things; they just expected old things to be more efficient. We all know this now. And yes, whole new kinds of mistakes or bad behaviors arose. The same could be said of electricity or cars, or fire.

I can’t prove that AI will evolve in as favorable a manner as PCs or the internet, or even my TI-35 calculator. I can say that the forces in play already point in that direction. We have a generation of natives going to college and joining the workforce. We have a technology that so clearly does new things in new ways.

Yes we have risks. Hallucinations are supercharged AutoCorrect typos. But you know, AutoCorrect (or spell check) typos were simply a much more efficient way to have typos than simply mistyping. AI might bring more versions of the Excel date formatting problem. I grew up in an era where we called things like these “stupid computer tricks” after our hero David Letterman. AI will have its own versions of stupid computer tricks.

One of the most popular ones to talk about now is the rise of hallucinations in legal briefs in court. Suddenly there is an explosion of court-submitted documents with made up citations or quotes. You know when you get a new car and suddenly you see that car everywhere. That’s what’s going on now. I ran across a compilation of all the court cases where hallucination came to the attention of the judge. There were just over 1,000 cases. That seems huge. You know what the denominator is? It is about 65,000,000 new cases filed every year (most lasting more than a year.) What is the rate of disciplinary action for shady filings without AI? Turns out that is common too. In the 1980s courts began a concerted effort to take note of attorneys stretching the facts in briefs—human hallucination—and since then most federal judges encounter that at least every year. There are almost 1800 federal judges. The reasons are different.

I am not dismissing any concerns at all, but a reminder that all the millions of laws we live under and rules in place at every institution still apply. No one is confused about accountability. This is even true with nascent agents as they are running as someone in an organization. AI is already a governed and regulated technology. AI is still overseen in its entirety by the post-war regulatory apparatus. This includes everything from liability to copyright to civil rights to antitrust.

New tools give us new ways to fail doing old things. And more importantly new ways to succeed at doing new things.

To anyone who made it this far, be a native AI user. Hire native AI users. Let them do native AI things. Let’s together see what a new technology can do. That’s how optimists and builders approach new things.

Author’s note: LLM was used for typos and grammar. I asked for pointers on where a skeptic would be provoked and in my own words made about five additions. FWIW, the model said those changes would result in simply putting down the essay and “reacting poorly to what you wrote.” Thank you Clippy.

Below is the full article on Harvard Law banning the use of computers.

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

AI was supposed to replace human labor.

It did the opposite.

For the first time in history, humans are cheaper than software.

And AI is creating more jobs than it eliminates.

Technology has always solved one problem by creating another.

In the 1830s, the advent of the railroad drove one of the largest infrastructure buildout the world had ever seen. American track mileage 120Xed in a decade.

Then the system broke.

On October 5, 1841, two trains fatally collided on the Western Railroad in Massachusetts due to a simple coordination failure.

As the complexity of the railways grew, individual conductors were no longer enough to keep train travel safe. The railroad companies thus began a decades-long effort: hiring managers for each geography, defining roles in writing, and establishing clear hierarchies with reporting lines. Modern management was born. With it, rail became among the world’s first billion-dollar industries, at its peak representing roughly 60% of the stock market.

AI is breaking the system again.

We just gave every employee, even the worst ones, effectively unlimited headcount and budget.

Managing AI is harder than managing people, because AI scales dysfunction instantly. Fortunately, we can learn from the past:

Agent workforces and human workforces fail in the same way.

Understanding the 7 major parallels between the two will unlock the next trillion dollars of AI value creation:

The 7 Parallels of Agent and Human Workforces

1. Tokenmaxxing is throwing bodies at the problem.

The tokenmaxxing hype cycle ran its full course in under a month.

But the amount of tokens spent was never the real problem.

People are spending so much on tokens because they don’t know how to use them.

Maybe 1 in 100 employees knows how to give AI context. It’s a rare breed of person that can articulate a process clearly, that has the patience to empathize with a polluted context window, or even understands what that means.

Give an agent harness to the other 99 people and they will produce “loops”.

2. Loops are meetings about meetings.

In Claude Code/Cowork, Copilot, Karpathy’s Autoresearch, or any harness, loops are a bandaid for the fact that almost nobody can successfully prompt.

Loops are a brute force attempt to compensate for human inadequacy. Agents call themselves to fix themselves only because a human never articulated the task cleanly. Brute force becomes the system’s only path to progress. This all stems from human failure to successfully understand the task in the first place.

You are spending tokens on spending tokens.

3. Wasted tokens are the new headcount bloat.

Most companies today are mismanaged.

The vast majority of workers don’t meaningfully impact the business at all. They are cogs in the machine, stamping approvals at each layer and hiring more cogs to fuel a machine that exists to exist.

They’re looping.

Often it’s more efficient to cut the loop. Elon cut 80% of X’s staff and the company performed better. Private equity operating partners make a living arbitraging this simple fact.

Just like 80% of employees do nothing, 80% of tokens today do nothing.

People create more people. Tokens create more tokens. Looping is the new empire building.

4. 100X tokens are the new 10X engineers.

The promise of software was that we’d build it once, run it forever at low cost, and never need to supervise it. AI broke that promise. As soon as software could do anything, it couldn’t do any one thing predictably.

Tokens behave like a workforce, and as soon as you see tokens as employees, the promises of AI start to break down:

• “Tokens are more accurate than humans” but only when prompted correctly.

• “Tokens are faster than humans” yet speed means nothing across 100 retries.

• “Tokens don’t play politics” but they build empires of token spend.

• “Tokens don’t quit” but they die between new model releases and new sessions.

• “Tokens can be trusted” yet they fail confidently with perfect formatting.

The one place AI really does beat humans is scalability. Scaling humans burns enormous energy across recruiting, onboarding, and attrition. Scaling tokens is instantaneous. Which is exactly why mismanaging them is so expensive, and why you must find and scale the 100X token.

The 10X engineer built the last era of companies. The 100X token will build the next.

In the same way a handful of employees make others 10X as productive, for any given job some amount of token context can cut AI effort down by orders of magnitude. There exist tokens that will give you 100X as much leverage.

Humans are cheaper than tokens on average, but good tokens are cheaper at scale.

Management converts one into the other.

5. Context hoarding is the latest job security tactic.

There’s a massive political problem with AI inside the firm, and it will only get worse.

Employees don’t want to teach AI systems their secret sauce.

They’re starting to pick up on the fact that these systems aren’t only there to “help them” or “increase productivity.”

Look at Meta, where stock-owning employees, who are wildly incentivized to get AI right, are outraged that the company is using employee context as training data. That’s at a tech company… a conflict that’s a microcosm of what is about to happen across every industry.

Tribal knowledge has been job security for centuries. Medieval guilds kept their methods secret. AI is the first technology that asks workers to hand all of it over at once.

Nobody trains their replacement for free.

The people who hold the 100X tokens have the least incentive to surrender them. Emotionally, structurally, and politically, firms are wired to reject the most important technology for their future.

6. Evals are the new OKRs.

The best way to manage a token workforce is the same as the best way to manage humans: by defining what good looks like.

The one breakout AI use case that escaped politics is coding. It expanded the pie and made engineers better.

The mechanism is evals. 99% of AI revenue today is coding because coding has built-in evals. Code runs or it doesn’t.

Broader, cross domain AI use cases will only come online when someone builds the requisite evals. Specific evals matter more than teaching your employees to prompt or giving them a chat harness. With them, AI will eat the parts of the economy code could never touch.

The real work of managing is turning fuzzy human processes into code, expressing the qualitative as quantitative.

A firm’s eval suite will become its most valuable resource.

Just like OKRs are key to leveraging a human workforce to optimal output, evals will be key to leveraging an infinitely scalable token workforce. Evals are the path to running 100X tokens.

Furthermore, no two firms will have the same eval set. Evals will be key to competitive advantage. An organization running generic evals or generic agents has no edge.

7. The next trillion-dollar opportunity is the transformation company.

Enterprises have been buying foundation model commits, the application layer, and internal builds for years now. All of it conceals a brutal truth about the economics:

Nobody has AI working reliably yet.

Silicon Valley is so convinced of this failure that its latest obsession is betting against the business of today. “Neofirms” or “AI Native Services” startups are being funded to capture the $21 trillion of services spend across the knowledge economy on the theory that incumbents, mired in their own politics and processes, will never manage the transition themselves.

Neofirms may well provide competitive pressure that catalyzes “tradfirm” AI adoption. But the greatest AI assets still sit inside the incumbents: differentiated processes that already work, scalable through distribution channels that already exist.

In fact, the next biggest businesses won’t be eating existing services spend. They will sell a net-new type of service to existing players:

“AI transformation companies” will be 10X larger than any neofirm.

Transformation sounds like a one-off project. But there’s a Jevons paradox at work: every use case an organization adopts can surface ten more. The more AI-enabled a firm becomes, the more transformation it consumes, while the frontier of what’s possible advances daily. Ongoing AI transformation efforts may become the only way to compete.

Consider Palantir, on paper the most Claude-disruptible company in software: a half-trillion-dollar business hand-building bespoke applications for the enterprise. By the logic that has made SaaS nearly uninvestable, $PLTR should be a zero before $NOW.

It’s not, because Palantir was never selling software. It was selling transformation.

But transformation itself has evolved since Palantir’s days of old. In an AI first world, it’s more than ontologies, custom software, and the rare bespoke prompt. The real work is in the evals, in token minimization, in understanding a business so deeply that you can program it.

Encoding each firm’s nuances into agents may become the largest economic task of the decade ahead.

It’s time to manage.

Every phase of the AI boom has had its guiding cliché.

We were told to sell pickaxes during a gold rush, and we built infrastructure. We were told to sell “Service-as-a-Software,” and we built neofirms. We have enough infrastructure. We have enough services. Now the work is making the trains run on time.

It’s time to survey the enterprise: to find the 100X tokens, record the loops that work, and direct intelligence that is being massively wasted.

Humans just became cheaper than software.

Someone still has to tell them both what to do.

Thanks to Sam Wolfe, , , and for their thinking here.

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

People are happier when they make things. The table your neighbor built himself beats the one he ordered from West Elm. Kids abandon their expensive toys by lunchtime and then spend all week making a basement fort with the boxes they came in. We know this about ourselves and yet we keep half-forgetting it, partly because most of the technology of the last hundred years was built to do the reverse: to let us have more while doing less.

That technology is wonderful. I can summon a (self-driving!) car or order groceries that show up in an hour. But almost none of it is what I’d point at if you asked me what makes a life worth living. A paintbrush has never saved anyone a single minute, and we’ve kept them within reach for tens of thousands of years.

A while ago, Eugenia Kuyda, who has spent years building AI consumer products, said something to me that I keep coming back to. Most people, she said, aren’t trying to save time. They’re trying to spend it.

Almost everything anyone says about AI, for it or against it, is an argument about time. Will it take jobs or augment them? How many minutes has the model shaved off writing the memo, and then off of reading it? Were the tokens worth what they cost?

But I think the more interesting thing about AI is that it’s also the other kind of technology, the paintbrush kind, the kind that lets us express ourselves as humans. A few technologies in history have managed to be both: language, the printing press, and the steam engine are all tools that saved enormous labor and, almost as a side effect, expanded what a person could be. AI belongs to that small category. It makes making things feel more possible.

Why does making matter so much more than consuming? Oliver Sacks had an idea late in his life that I keep coming back to. He observed that most people treat being present as the goal, as if the whole point of living were to sit inside the current second. He thought that was a little sad. The people he found most alive were soaked in the past and future at once: remembering, planning and dreaming. That’s what making does to you. It stretches you across all three tenses, because you’re building toward something you can already see, out of everything you’ve ever loved. Consuming, by contrast, parks you in the now.

Take social media. At the start it was a little miraculous: people taking the time to share what was happening and what they were thinking with each other. Then it became a contest. A call for attention. And once the algorithms took over, the thing that broke through was whatever was loudest, so people optimized for loud, and lately we’ve started calling the machine-made version of it slop.

Subscribe now

But the feed was slop before any machine got involved. Slop is just what you get when everyone consumes and nobody creates, and the cure is people making things again, which AI is unusually good at enabling. That’s about the most anti-slop thing I can imagine. A companion you talk to and shape and argue with will do more for you than a celebrity you watch from across the internet, because one of them is a thing you make and the other is a thing that happens to you. It’s the difference between putting on an autogenerated Spotify playlist and making somebody a mixtape. A mixtape is you, smuggled into another person’s afternoon.

And this isn’t reserved for creatives. A master electrician in Kentucky with no computer science degree used AI to build a load-calculation tool that sells for $12.99 and replaces a $500 service call. A plumber canceled a $40,000 consulting contract after a single afternoon with OpenClaw got him further than the consultants had scoped in weeks. For most of computing history, if you couldn’t code, you were a consumer of other people’s ideas. That’s over. The cost of trying things has collapsed, and the people picking this up first are not the ones anyone predicted. Software is about to be everywhere the way YouTube made video everywhere, and most of it will be built by people who’d never have called themselves builders.

The same thing is arriving in the time you spend at home. The McKinsey-esque (and unimaginative) version of the promise of AI is that everything in your life will get ten percent more efficient. The better version is that every year or so, you’ll find one new thing to fall in love with and pour your free time into it. None of this matters, in the big serious sense of the word, which is the good news, because the things that don’t matter are usually the things people love the most. Fan fiction. A kid DJing in his bedroom. Some ridiculous game you and your daughter invent on a rainy Saturday. It’s also where nearly every great consumer company has ever come from. As my partner Chris Dixon said, the next big thing will start out looking like a toy.

Work is the same story. The good part of any job is the stretch where you’re doing the thing you’re genuinely good at, and everything around it, the politics and the status and the meetings about other meetings, is a tax you pay to get there. AI is starting to eat the tax. As it shrinks, the job feels more like yours, more like the thing you’d have done anyway. And the part of work that feels like play is where the real breakthroughs have always lived. The good ideas almost never climb out of strategy books. They come from some side quest a person went down because they couldn’t help themselves, the strange weekend project, the rabbit hole that paid off.

There’s an old story about the moment the hippies turned into the finest capitalists the country had ever produced, well-explored in David Brooks’ book Bobo’s in Paradise. All that sixties individuality, the stuff that was supposed to be the sworn enemy of money, gradually evolved into the engine of it. The war between the strivers and the rebels didn’t end with a winner; it ended with a merger. Being a distinct person stopped being a rebellion and became the whole economy (some companies that exemplify this include Ben & Jerry’s, Restoration Hardware, and Range Rover).

What I want is more of that, pushed harder. Individuality at scale. Which is why the frightened version of the AI story, the one where a few companies own everything and everyone else sinks into a permanent underclass, has always felt to me like the wrong story. The side quests and the rabbit holes and the weekend projects don’t ladder up into a neat diagram of who owns what. Whatever barbell this thing creates, you’ll feel it first in the texture of your own days, in what you get to make and who you get to become, long before it ever turns into a chart of how society is arranged.

Look at the people who already use this stuff all day. They haven’t gone limp and checked out. They’re working more, and they’re further into it. People who spent years as spectators of their own ambitions are finally actualizing them. Give someone the means to build the thing they’ve been carrying around in their head, and they don’t lean back. They lean in. They stay up too late, and they text you at 2am to show you what happened.

A few months ago I went on a strange guys’ trip where the daily routine was four hours on our laptops in the morning, then padel, then a long dinner, and the dinner was mostly four grown men showing each other what they’d built that morning. That felt like the future to me, mostly because of how badly everyone wanted to show their work.

Here’s what I think happens when execution gets cheap. For most of history, the bottleneck on making something was never the idea, it was the grind: acquiring the years of skill, raising the money, assembling the team, and getting the permission. So most people’s best ideas died inside them, unmade. Lift that bottleneck and the thing that decides what gets built is no longer whether people can justify the VC funding or the enterprise-level capital expenditure, but who has something to say. The most human thing about you stops being a private quirk and starts being the point.

Individuality was supposed to be the luxury you bought once you’d made it. I think it’s about to become the thing everyone gets to spend their life on, the work itself. So when I picture where this goes, I see a lot of people making strange, beautiful, slightly pointless things and handing them around to see what happens. The toys were fine. But we always wanted to play with the boxes.

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

This essay by David Ulevitch originally ran earlier this year, prior to our America Dynamism Summit. Have a great July 4th weekend everyone, we are off next week and we will see you on the 13th! -AD

The demand for American Dynamism has never been greater. America is and always has been a land of builders, from the framers of the constitution to industrial titans to modern technologists and engineers. This year of AMERICA 250 is a celebration of that very character, of the incredible innovations and people who spent the last 250 years turning a vast, untamed continent into the world’s superpower—and who are now working to build, defend, and power the next American Golden Age.

A confident and capable America strengthens the world. The American Dream has never been confined by geography. For generations, American innovation has set standards, opened markets, secured trade routes, and underwritten global prosperity. When America builds, allies benefit. When America leads, freedom scales.

When we launched Fund I four years ago, the relationship between Washington and Silicon Valley had long turned cold. Decades of offshoring and regulatory inertia hollowed out large parts of our industrial base while the “Magnificent Seven” concentrated almost all of the market returns. The U.S. once led in 60 of the 64 most critical technologies. Today we have almost entirely lost that lead, with China now leading in 57 and the U.S. in only 7 technologies. Early American Dynamism pioneers like SpaceX and Palantir made major technological leaps but had to fight to even be allowed to compete for their first government contracts, let alone win them. At the same time, the geopolitical threat from China was growing and the war in Ukraine had not yet begun.

Since then, the world has woken up to the need for American Dynamism. Startups in formerly off-limits industries like defense, energy, and manufacturing are winning real commercial and government programs. Defense has drawn much attention in recent years, but American Dynamism is more than defense. It covers the full “shift left” electro-industrial stack—batteries, motors, chemicals, materials, mining, and more—everything that enables the mission of domestic resilience and nearshored supply chains.

And we are only at the beginning. Today, a rare alignment of undeniable demand, technological breakthroughs, supportive policy, growing capital, and top-tier talent has created a once-in-a-generation opportunity to accelerate this sector’s growth, our national security, and our country.

What used to take decades now takes years. Legacy defence primes once needed nearly 30 years to reach their first $100M program. The first generation of startups with the U.S. government as a customer reached similar milestones in 15 years. Today, thanks to breakthroughs in AI, autonomy, and advanced engineering, along with new acquisition pathways, American Dynamism companies are hitting those milestones in under 5 years. Anduril won their first program of record in 3 years and Saronic did it in under 3 years.

These companies have now matured into talent pipelines, training the next generation of mission-driven founders, operators, and engineers galvanized to build for the national interest. Alumni from SpaceX, Palantir, Anduril, and Tesla—as well as military veterans with field experience—are founding and leading the startups on stage at Summit, and many more great companies besides that. American Dynamism founders are a new breed: deep technical operators who combine engineering mastery with mission-driven purpose to build the infrastructure that literally keeps the lights on and the country secure.

The wins of American Dynamism companies are wins for America and the American people. Historically, an American worker did not need a college degree, let alone an elite college degree, to have a well-paying manufacturing job that allowed them to live a dignified life, raise a prosperous family, and be invested in the future of the country. Recent decades have gradually eroded this status quo as middle and working class jobs have been offshored and college degrees have now become a prerequisite for most middle class jobs, regardless of actual skill.

Bringing manufacturing home is central to many portfolio companies. Last year, Anduril announced Arsenal-1, a hyperscale factory in Columbus, Ohio, which is expected to create over 4,000 direct jobs. This makes it the largest single job-creation project in the state’s history. Saronic’s Louisiana shipyard expansion is projected to add about 3,270 jobs with average annual pay around $88k (roughly 46% above the parish average). While many startups are headquartered in coastal tech hubs, the skilled labor to build at scale is in America’s heartland. Thus opportunity, training, and high-quality careers are being brought back to communities that need them.

More than ever, on the 250th anniversary of American Independence, our goal is to win: to ensure our founders win, our portfolio companies win, their incredible talent pools win, the best technologies win, and that America and the American people win as a result. We have spent the past few years putting together an all-star team to identify the best founders and act as an accelerant to connect capital, policy, and procurement. We will continue to be the bridge between Silicon Valley and Washington D.C., to champion pro-innovation policy and make the introductions that put the best technology into the hands of the American warfighters, first-responders, and all those who work in service of the American people.

Technology determines who sets standards, controls supply chains, and secures prosperity. If freedom is the world’s operating system, we must fund it like our lives depend on it.

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

Welcome to a special Vacation Week issue of chart deep dives, where we go into a specific topic with data from a portfolio company. Today, we have a fun issue, thanks to our friends at Deel who shared some data with us on… how people take vacation.

All of the data below are from companies using Deel, so they are going to skew towards startups, tech companies, and remote-first firms (relative to average population data, anyway.) So, rather than “this is how everyone takes vacation”, this is going to more accurately represent, “how people in our industry (and other white collar vocations) take vacation around the world”, which frankly is just as interesting a comparison if not more.

Unsurprisingly, there’s a wide range of how much vacation days are offered to the median employee, and an equally wide range of, “do people actually take all of the vacation they’re owed?” North America clusters more stingily than Europe does, but there’s still a lot of variation within regions, particularly among people who claim 100% of the days off that they’re offered:

The one caveat we can think of here, which would be hard to untangle with the data we have here, is, “Do norms vary between countries on how much people formally take the vacation days off, as opposed to informally (and not officially docking the days with HR)?”

Over the course of the year, the Germans hold the OOO crown. But not consistently! Swedes and Italians trade places in July / August. The British are (proportionately) at their desks in peak summer; the Dutch over Christmas.

True to form, European out of office messages are doing a lot of work over the course of the summer.

And also the rest of the year, while we’re at it.

Interestingly, there’s a wide range of how long people take their summer breaks - and, while Europe clusters differently from the Americas on the usual pattern, Brazil breaks the trend. There’s a reason for this, it turns out! Brazil has some interesting laws about how employees take vacations. Brazilians get 30 days of vacation allowances (including bank holidays), but they have to take those holidays in blocks; a maximum of 3 per year, apparently. You learn something every day!

Still, most “vacations” are a day off, not a long holiday.

Another regional variation: vacation days in India are booked last minute. Whether that reflects a genuine last minute trip decision, or simply last minute disclosure to HR, we can’t say.

Long weekends seem to have special allure in the summertime. It makes sense that vacation days skew towards Mondays and Fridays, and also that the effect gets magnified in the warmer months. Still, this is a smaller discrepancy than we would’ve guessed!

A popular alternative, the “Summer Friday”, is more officially popular abroad than it is here. Unofficially, though, is another question. Our guess is that Americans are doing it off the books, rather than formally taking the half-days off.

Looking at last year as an example, it’s no surprise that the single most common day to take off is a “buy one get one free” situation.

Birthdays are also a popular day to take off, but more so for big ones. Turning 31? Not worth celebrating, apparently.

Also, Sweden, what’s going on here?

And then there’s the dirty little secret of vacation days… sick days. Which, of course, are only used for recovering from illness; we would never imply anything otherwise.

Of course, we’ve spent all this time on summer holidays, but the real empty office weeks are at Christmastime.

The War on Christmas never won the enterprise, apparently.

So, no matter how hard our Swedish, French and German friends work to make those vacation days count, summer break is still getting mogged by Christmas break. (Everywhere but… Armenia?)

Wherever and however you’re taking any time off this summer, enjoy it!

Have a great summer, from the folks at Deel and a16z. If you want more charts like these, you can subscribe to Deel’s substack here:

Subscribe to Deel

Aside from something at the end of the week to commemorate America’s 250th, we’ll be off until July 13th. See you then.

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.