Every year we gather our LP’s and spend a day sharing our vantage point of the world with the people who trust us with our capital. This is largely a closed-door affair and has become even more important as AI changes technology seemingly at the speed of life. This year was no different, and we spent a good amount of time sharing where we think AI goes from here and how we plan to deploy capital in this next phase.
This all happened before the S-1 dropped. Anthropic filed publicly just a couple of days back, and the numbers deserve a minute of envy. Over $6.6 million of revenue per employee on a trailing twelve-month basis, up from $1.4 million a year earlier. About 5,300 people. Revenue up 1,368% year over year in the June quarter. That is the benchmark now, and honestly it should be. It's what a company looks like when it runs on its own product, and it's the bar our fastest-growing companies will be measured against.
But it doesn't change what we told our LPs. If anything, it underlines it. Our headline was not “wait for the next model.” It was, if you only built on the models today - there’s 100x more value to unlock.
I live this personally. In the last thirty days, I've made over five thousand requests to my own agents. Research, scheduling, diligence, follow-ups. That number would have been close to zero not long ago. The gap between how I work now and how almost everyone else still works is the story.
Contrary to what many believe, we don't think we're anywhere near the top. We think the number is closer to 100x than 10x. And we didn't make it up out of thin air. It's backed up by a lot of data from people who are as smart as anyone about investing in this category and building in it. We sliced it four ways for our LPs.
Start with the user gap. Gavin Baker from Atreides talks about this one.¹ Fewer than 10 million heavy AI users worldwide are supporting roughly $175 billion of annualized AI revenue. Now set that next to the denominator. The global knowledge workforce is 1.5 billion people. More than 1.2 billion people already use ChatGPT. We think the 100x lives in what happens when everyone starts working like the power users already do.
The second is the enterprise gap. Fewer than one in ten enterprises have AI agents in production. Some surveys put it closer to 5%. Meanwhile, we believe the models can already handle the majority of knowledge work as it exists today. According to Anthropic's labor market study, current AI models can automate 90% of office and admin roles. The capability is already here; it just hasn't been deployed yet. Companies like Blitzy, Limy, and Voice Run are bringing this into the enterprise across every part of their business.
The third is small business. Three-quarters of small businesses already touch AI,³ but almost none of them actually run their day-to-day on top of it. Today it's a few ChatGPT prompts a week. Soon, dozens of agents per employee will run the phones, the scheduling, and even the books. Our portfolio company Topline Pro already puts more than four agents to work for every roofer, plumber, and landscaper on its platform, and those pros have booked over $655 million in jobs through it. 80% of them didn't even have a website before.
And finally, our own experience. Six months ago we launched Flybridge OS (thanks to Daniel and our team), our internal agent system. We were running low single-digit millions of tokens a month through it.⁴ Today it's 23 times that, growing 70% month over month. It helps us source deals, track thousands of early-stage and growth companies, and find customers and experts for our founders, as well as co-investors when they're ready to raise their next round. We're one small firm of 10 people. Multiply our usage by 100x, and then multiply that by every firm and company doing the same, and it adds up to an incredible amount.
From there, we moved on to the four areas we're exploring. The list is by no means complete; we dive deep into emerging areas in parallel with making investments, often with contributions from our Summer Associates, experts at some of the fastest-growing AI companies in the world, and the companies we see every day. We believe each area we shared with our LPs holds 100x upside from where it sits today.
Here they are:
The agent economy. Agents are becoming each other's customers, coworkers, and counterparties, with real money already moving between them. That world needs identity, permissions, and payments built for machines. Arcade is building the identity and permissions layer. Limy helps brands get discovered by agents, not just people, and it's already working with companies like LVMH, Amazon, and P&G. When your next customer is an agent, distribution looks completely different. We got a small taste of this a few weeks ago. Dorothy hosted a dinner in New York where 22 founders and investors ate in one room while 14 of their agents held a parallel dinner party of their own in a chat room. The New York Times covered it: 22 People Got Together for a Dinner Party. So Did Their A.I. Agents.
Models beyond the frontier labs. You can't run 100x usage on frontier models alone. It's too expensive, and language isn't every problem. The future needs special-purpose models: open-weight, small, deterministic, scientific. Arcee helps companies own their own models instead of renting intelligence by the token. Deterministic models make the same decision the same way every time. Science models like Polymathic train on experimental data instead of text. Different tools for different jobs.
One thing that surprised us at the meeting: for a lot of our LPs, this was the first time anyone had pulled back the curtain on what closed-weight, open-weight, and open-source actually mean in the age of AI. So we walked them through the short version, the same way we would for anyone:
A quick primer. A model is, literally, a giant series of equations. Billions, even trillions, of parameters. Every time you ask it a question, those equations calculate the most likely next token of the answer.
In a closed-weight model — OpenAI, Anthropic, Grok — you can't see the equations. You run it on their cloud, or Amazon's, or Microsoft's, and the provider controls the model and the price of every question you ask.
In an open-weight model, you don't get the source code, but you can see the equations. You can run it yourself, in your own environment, even on your own machine. You can see which parts fire and which don't, and you can change the weights to match what you actually need. Say you want the best customer-service model for your company. You don't need the parts that know American history or Greek mythology. Strip those out and your cost drops while your speed goes up. You can't do that with a closed model. You'd be reverse-engineering blind, and in a lot of cases it isn't even legal.
Open source is the fullest version: the code and the weights are both out in the open, the tradition we have backed for years, going back to companies like MongoDB. Open weight sits one step short of that. No source code, but the model itself is yours to run, inspect, and modify. That's the difference between renting intelligence by the token and owning the model that runs your business.
The places the frontier labs won't go. A frontier lab's reference point is a billion users. A billion-dollar category is a rounding error to them and a fund-returner to us. Agents going into production in real industries need vertical data the labs will never touch, in markets you can't just prompt your way into. You can see it in our portfolio. Blitzy works inside the ten million lines of proprietary mainframe code its customers run, code that has never touched the internet and can't be bought at any price. Crop Diagnostix is building the data to spot a fungal infection in a cornfield in Iowa. We call these the unpromptable markets: no matter how smart the public model gets, you can't get there by asking it a better question.
Physical AI. Intelligence leaving the chat window and taking action in the real world: railways, airfields, farms. Our bet is on specialist machines over general-purpose humanoids, and on the data underneath them. Physical AI runs on data that mostly doesn't exist yet and has to be created, which is its own opportunity. Three of ours are already out there doing the work. Intramotev is putting autonomous rail cars on the rail network. AeroVect moves luggage and cargo across airfields. Agtonomy manages tasks across large farms and worksites. All three sit in regulated, hard-to-reach markets, so they're likely to see two or three competitors instead of twenty or thirty.
Talk to us
If you're building in any of these areas- the agent economy, models beyond the frontier, the verticals the labs won't touch, or physical AI- we want to hear from you. Our inboxes and DMs are open:
¹ Gavin Baker, CIO of Atreides Management, on the user concentration behind AI revenue and the 1.5 billion knowledge workers still to come online: The a16z Show, “Is There an AI Bubble?”
² Anthropic, Labor Market Impacts of AI: A New Measure and Early Evidence (Massenkoff & McCrory, March 2026), on what today's models can already cover versus what's actually deployed.
³ Bluevine, 2026 Small Business AI Trends Report: 74% of small business owners are using or testing AI tools.
⁴ Pulled from our own usage data across OpenRouter, Anthropic, and other model providers.
