I have been reading Y Combinator batch pages the way some people do crosswords, and recently I started writing down team sizes. The AI-native service companies I tracked across the recent batches — an accounting firm run by two people, insurance operations run by five, an AI-native law firm of ten, agents doing policy servicing for major carriers with twenty-one — line up like this: 2, 2, 5, 6, 10, 15, 21, 25, 30, 50.
A Fibonacci-minded reader spots the 2, the 5, the 21, and starts squinting at the rest. Let me save you the squint: no, Y Combinator startups do not literally grow along the Fibonacci sequence, and anyone who tells you otherwise is selling a newsletter. What caught my attention is something better. These companies exhibit the two properties that give the sequence its name: each stage is built by recombining the two stages before it, and the growth ratio between stages converges on something small and stable — roughly the golden ratio’s 1.6, not the 2.0-with-a-hiring-plan of the SaaS era. The numbers are decoration. The recurrence is real, and it is the most useful mental model I have found for what is happening in these batches.
Here is the recurrence I keep seeing, told in Fibonacci’s own numbers.
1 + 1 = 2: the wedge. A founder and an agent, pointed at one painful, repetitive, compliance-heavy workflow. Cranston AI (Fall 2025) is two people — one of whom previously owned a tax practice that filed ten thousand returns — shipping agents that work tax returns inside the software CPA firms already use. Rote is two founders building an AI-native insurance department for auto body shops, disputing denied claims. The wedge stage is deliberately narrow: one workflow, owned end to end, in a domain where the founder has scar tissue.
2 + 1 = 3: the trust layer. The wedge plus the thing that makes a regulated buyer sign: auditability. Every drafted clause, reconciled balance, or filed claim traceable to its source, reviewable in one click. This is the stage most generic AI tools never reach, and it is where vertical companies earn the right to charge service prices. Casey (Fall 2025, five people, founders out of Swiss insurtech) runs this split explicitly: licensed producers own every client relationship while the machinery handles carrier paperwork behind them.
3 + 2 = 5: the swarm. The trusted service plus the wedge playbook, repeated — more workflows, more agents, the same tiny human core. Moritz’s ten humans run an entire law firm this way: AI executes roughly 80% of first-pass legal work, elite lawyers own the final 20%, and the firm reports $3 billion in deals closed at a four-hour average turnaround. Toothy AI (fifteen people) swarms dental insurance verification with voice agents that call payers — reclaiming the 160-plus monthly hours clinics spent on hold.
5 + 3 = 8: the vertical operating system. The swarm plus the service, extended across the whole vertical until the company is a layer the industry runs on rather than a tool the industry uses. Strada (twenty-one people, Summer 2023 — an early mover) runs policy servicing and claims operations for major US carriers with native integrations into their core systems. Harper (twenty-five) is building the autonomous commercial brokerage, its agents learning continuously from the human brokers beside them. Kinter (thirty) sells AI accountants running the financial close inside enterprises, and Artisan (fifty, Winter 2024) reports thousands of customers running its AI business-development employee — over 90% of them fully autonomously.
Every company above I checked against its live YC page this week, and the directory turned out to have its own archaeology. Kinter wears a “W20” badge that looks impossible for an AI accounting company until you learn it inherited the label from its previous life as Alloy Automation. Moritz now claims $3 billion in closed deals at a four-hour average turnaround, half a billion more than the figure circulating in older summaries. And Rote already carries a Winter 2027 badge in August 2026, because YC lists early admits before their batch begins. The future cohort is, in a small way, already public.
Notice what the recurrence forbids: skipping. A company cannot jump from wedge to vertical OS, because the OS is made of the intermediate structures — the trust layer is a component of the swarm, the swarm a component of the OS. Every failed “we’re building the AI operating system for industry X” pitch I have seen failed exactly here: it tried to build the 8-square without laying the 2 and the 3.
The second Fibonacci property — the stable, small ratio between terms — is where the economics live. In the SaaS era, doubling revenue meant something close to doubling the org chart. The companies above grow revenue on a compounding curve while headcount moves in reluctant golden- ratio steps, and sometimes not at all.
The cleanest data comes from Stripe’s analysis of the top AI companies on its platform: a median of 11.5 months to $1 million in annualized revenue — months faster than the best SaaS cohort ever managed — and the number of solo operators clearing $1 million more than doubled between 2023 and 2025. At the extreme end: Cursor’s maker Anysphere reached roughly $100 million in annualized revenue with about twenty people (per Sacra’s estimates — label them as such), and Lovable crossed $100 million eight months after launch with about forty-five employees.
This is the context for the famous bet. Sam Altman told Alexis Ohanian back in February 2024 that his tech-CEO group chat keeps a betting pool on the year the first one-person billion-dollar company arrives. As of this writing, no verified one-person unicorn exists, which needs saying plainly because the claim gets recycled as if it had happened. But the trend line the bet sits on is real, and the YC companies above are the early terms of that sequence.
Start with the venue, because YC itself changed shape while this was happening. Since 2025 it runs four batches a year — Winter, Spring, Summer, Fall — and the Spring code is, delightfully, “X” (X25, X26), because S was taken. The AI-agent share of these cohorts has been climbing: PitchBook counted ~46% of the Spring 2025 batch as agent companies, up from ~36% the batch before. Garry Tan told CNBC the Winter 2025 batch was growing about 10% per week in aggregate — the fastest in the fund’s history — and that for a quarter of that batch, 95% of the code was AI-written. (For balance: at least one independent analysis disputes how exceptional the growth really is. Aggregate weekly growth claims deserve the same skepticism as any other unaudited number.)
The products tell the same story. Each era of AI products has been built from — not instead of — the one before it.
The strategists have names for the destination. Foundation Capital calls it service-as-software and sizes the opportunity at $4.6 trillion — the spend on services and the labor behind them, against the few hundred billion the world spends on software. a16z has documented the pricing shift from seats to outcomes. PitchBook’s Q1 2026 analyst note is titled, without apparent irony, “SaaS Is Dead, Long Live SaS”. And YC’s own requests for startups stopped being subtle: the “full-stack AI companies” framing from partner Jared Friedman put it as bluntly as venture prose allows — instead of selling to the dinosaurs, you could make them extinct.
And one layer further down, the substrate itself is being industrialized. A whole cohort of companies now sells the picks and shovels of agentic operation — sandboxed computers for agents (E2B, Daytona, Browserbase have collectively raised nine figures for exactly this), payment rails for agents (Stripe’s agentic commerce suite; the x402 protocol, whose foundation now counts Visa, Mastercard, Google, and AWS as members), and authentication that lets agents act across third-party apps. Every wedge founder who starts today inherits this infrastructure the way a 2010 founder inherited AWS. The early terms of a Fibonacci sequence are slow; the compounding comes later, and the substrate is what the compounding runs on.
This is where the analogy has to admit its limits: Fibonacci’s original rabbits were immortal, and startups are not. The sequence above says nothing about churn, about the 40% of agentic projects Gartner expects to be canceled, or about what happens to a ten-person law firm’s margins when its underlying model provider changes pricing. The golden ratio describes the survivors. And the recurrence cuts both ways — a company whose trust layer fails loses not just a stage but every later stage built on top of it, which is the polite mathematical way of saying that in regulated verticals, one audit failure can unwind the whole spiral.
So no — the YC batches do not follow the Fibonacci sequence, and I will not pretend the 2, the 5, and the 21 in my team-size list are anything but a coincidence that made me smile. What they follow is the sequence’s logic: build each stage out of the stages you already have, keep the human core growing slower than the value it governs, and let the compounding — which always looks unimpressive in the early terms — do what compounding does. The companies above are at 2, 3, and 5. For the batches now forming, the question is who gets to 8.
References
All company details verified against live Y Combinator directory pages in the week of 24 August 2026. Where a figure is a company’s own claim or an analyst estimate, the text says so.