First, some org chart history

A company founded in 1995 had a room full of clerks. A company founded in 2013 did not, and the reason was not that anyone set out to eliminate clerks. It was that finance lived in a SaaS product, and support lived in a SaaS product, and payroll, recruiting, pipeline, and infrastructure all lived in SaaS products, and the jobs simply never appeared on the org chart. We called that era software eating the world and pointed at the products being sold. What few people realized was that the more consequential thing was happening inside the buyers.

We are in the early part of the same transition, one layer up, and almost everything currently being written about startups and AI is looking at the wrong side of the transaction.

The substrate, not the product

Shipping a product with a model inside it is not a differentiator. It will not be a fundable thesis for much longer, and it will not distinguish a startup from an incumbent, because the incumbent will ship one too. That is the visible half of the story and it is the boring half.

The question worth asking about a company founded this year is not what it sells. It is what the company runs on.

The functions that survived the software transition are the coordination and judgment work that sat on top of the SaaS tools: the person who reconciles what the tools disagree about, the person who reads the contract, the person who routes the exception, the person whose job is to know which of nine dashboards is lying. That layer is now being absorbed. Finance, HR, payroll, compliance, and data operations run something like 30–40% of operating cost at a typical mid-size organization, and where automation is genuinely deployed the reduction runs 25–50%. The point is not the cost saving. The point is that operating capacity gets uncoupled from headcount, and once that happens the shape of the company changes permanently.

Why this is an early-stage phenomenon

The advantage a five-person company has here is structural, and it has nothing to do with talent, ambition, or how modern anyone feels.

A startup adopting AI internally is just building. An incumbent adopting AI internally is running a change-management program against installed systems, negotiated vendor contracts, a data governance committee, an audit trail somebody put their name on, and a meaningful number of employees whose job description is the thing being automated.

The evidence that this is hard is not subtle. Roughly four in five organizations report challenges adopting AI, up double digits year over year. More than half of C-suite executives say adoption is tearing their company apart. Only about a third report meaningful ROI. Change management and workflow redesign now outrank the technology itself as the binding constraint. And a striking share of companies are planning layoffs for non-adopters, which tells you the internal resistance is real enough that leadership has resorted to threats.

A company with no installed base, no committee, and nobody to threaten has none of this. It runs lean from the first week and never notices what roles it didn’t hire.

The headcount data is already moving. Median team size at seed is four people. Average Series B headcount has fallen from 53 to 45. Average Series D headcount is down 29% from its 2023 peak. AI-native companies run roughly 40% smaller teams than comparable software businesses at similar revenue.

I want to be honest about what that data does not prove. Some of the thinning is post-2021 capital discipline rather than technology, and the two are genuinely hard to separate right now. Some of it will reverse, because a fraction of the headcount an incumbent carries is not waste but control, and startups that learn this the hard way will hire back into it. Notice that the thinning is dramatic at seed and modest at Series D, which is roughly what you’d expect if both stories were true at once. Discipline reverses when capital gets cheap, but substrates do not. The 2013 company never rehired the clerks and the current generation of startups will never hire into the roles that AI already automated.

The objection, and why it doesn’t save us

Here is the strongest argument against everything I’ve just said, and it comes from my own analogy.

The software substrate shrank per-company headcount, but it did not shrink venture. It did the opposite. AWS took the cost of starting a company from $2M to $50k, and the industry responded by starting twenty times as many companies, inventing the seed fund, and doing more financings than ever. If AI does the same thing, the answer is more startups, not fewer doors, and this whole thesis is inverted.

I think this objection is right about the aggregate but wrong about the thing that matters.

Venture returns do not come from the aggregate. They come from a handful of companies per vintage, and the only question that has ever mattered is whether you can get meaningful ownership in one of them. Cheap formation produces enormous numbers of new companies at the top of the funnel, which is where entry has never been the constraint. What it does not produce is more entry points into the winners.

And the winners are where the compression bites hardest, because leverage compounds fastest exactly where the business is working. A company with real traction, a four-person team, and a large round in the bank does not come back to market in fourteen months. Median seed-to-A is already out to twenty months. Extend that: a company that once did seed, A, B, C, D and two extensions instead does seed, A, and a growth round, and then never needs outside money again.

Every financing is a door. Fewer financings means fewer doors, and the doors that remain are available first to the people already inside the house. So the proliferation objection is correct ... and irrelevant. You get more companies to invest in and fewer chances at the ones you’d want.

What this does to liquidity

The consequences run in both directions, and the second one is where the market has been slower to notice.

The exit story right now is optimistic. IPO counts are recovering, several long-awaited names have listed or filed, and the trade press has moved on to writing about the great unlocking. All true, but none of it changes the arithmetic. The 2021 vintage has returned about eight cents on the dollar. Median DPI at year eight is down to roughly 0.7x. Distributions have run at 14–15% of NAV since late 2022, about half the historical norm, while hold periods stretched past seven years. A few trillion-dollar listings do a great deal for a few cap tables and nothing at all for the several hundred billion dollars sitting in funds that will never have a company anyone wants to take public.

Liquidity is recovering at the very top and remains structurally broken everywhere else. That is not a rate-cycle condition. It is what concentration looks like from underneath.

Meanwhile the people inside the winners are stuck in a longer version of the same problem. A founder whose company will now stay private for twelve years instead of eight, with three financings instead of seven, has fewer of the incidental liquidity moments that used to come bundled with a Series C. They still have mortgages, children, and a reasonable objection to holding one illiquid position for a decade...

Which is the whole argument

Investors can’t get in through the front door. Insiders can’t get out through the front door. Both pressures resolve in the same place.

The US venture secondary market did roughly $106 billion in 2025 and about $122 billion in the trailing year through Q2 2026. In the first quarter of 2026, direct secondary volume surpassed IPO volume for the first time in the history of the asset class.

That is not a workaround anymore, and it is not a distressed trade. It is the primary market for exposure to companies that no longer need a primary market, and the primary source of liquidity for the people who built them.

Venture spent forty years on the assumption that ownership is acquired at financings and returned at exits. Both halves are coming apart at once. Secondaries are what is underneath.