Why AI makes startups cheaper to build but harder to defend

AI makes software cheaper to build and, for the same reason, harder to defend. In 2017, 17% of the new startups incorporated on Carta had a single founder. In 2024 the share was 35%. In early 2025 Garry Tan, who runs Y Combinator, said that for about a quarter of the companies in its winter batch, 95% of the code had been written by AI. "You don't need a team of 50 or 100 engineers," he told CNBC. "You don't have to raise as much."

He is right, and it is good news for anyone who wants to start a software company. It is worse news for anyone who wants to own one. Whatever makes your product cheaper to build makes your competitor's product cheaper to build too. A company used to be protected, in part, by what it cost to make. AI is removing that protection.

This has happened before

In 2006 Amazon Web Services began renting servers by the hour. Before that, a web startup had to buy hardware before it could learn whether anyone wanted its product. Afterwards it could rent capacity for the length of the experiment. Michael Ewens, Ramana Nanda and Matthew Rhodes-Kropf studied what happened to venture capital in the sectors where this mattered most. After 2006, first financing rounds fell by about 20%. The number of first investments venture funds made in those sectors each year nearly doubled. Investors took fewer board seats, and more of the companies they backed failed. The authors called the new style "spray and pray".

Cheaper experiments meant more of them, and less of each one. AI is a larger version of the same shock. Cloud computing made the servers cheap. AI makes the engineers cheap, and engineers were most of the cost.

Cheap for you is cheap for everyone

The economics of why this lowers the value of firms is old and simple. Picture a market for some kind of software in which products differ a little: one fits retailers better, another fits banks, a third has the interface some users prefer. Each firm charges a markup that depends on how close its nearest rival is. Entering the market has a fixed cost, which is mostly the cost of building the product. Firms enter as long as the markup they can expect covers that cost.

In Research Note 001 I work through this standard model (the circular city of Steven Salop) for AI software. In equilibrium, each firm's gross profit equals its entry cost. That is not an approximation. Entry continues until it is true. When AI halves what it costs to build the product, more firms come in, each sits closer to its neighbours, markups fall, and the gross profit of the typical firm halves too. If investors value the firm at a multiple of gross profit, its value halves as well.

Two features of the result run against intuition. The first is that market size does not matter. A market twice as large attracts more entrants, and each still earns its entry cost. Founders often argue that AI expands the market so much that everyone wins. It may expand the market, and the expansion still goes to customers, through lower prices and more choice, not to the typical firm.

The second is that cheaper compute does not matter either. Everyone buys inference from the same few model providers at the same prices. When those prices fall, every firm's costs fall together, competition passes the saving on, and gross profit per firm is unchanged. Reported gross margins rise, because the same profit sits on smaller revenue. That is a change in accounting ratios, not in value.

Cheaper to build, dearer to run

There is a twist that makes AI different from cloud computing. Classic software cost a lot to build and almost nothing to run: one more user was nearly free. AI software costs less to build and more to run, because every answer and every completed task consumes compute. I call this the cost-structure rotation. Fixed costs fall and marginal costs rise.

Both halves of the rotation reduce defensibility. The fall in fixed costs invites entry. The rise in marginal costs removes the classic software advantage of scale. If serving the millionth customer costs as much compute as serving the first, size alone stops lowering average cost. The large incumbent and the three-person startup buy the same tokens at the same price.

The financial statements of listed software companies show the first signs. In the paper I follow 307 US-listed software firms from 2018 to 2025. The 32 that discussed generative AI earliest in their filings saw their gross margins fall relative to other software firms from 2023 on. Their revenue also grew faster, but so did that of firms that were already talking about machine learning in 2019, so faster growth is not specific to generative AI.

The margin result is consistent with the rotation. It does not prove it: early adopters started with margins about nine points higher, part of the decline may continue a trend that began before ChatGPT, and the result rests on 32 firms. The same data show what investors pay for. Across these firms, market value tracks gross profit more closely than revenue. When both enter the regression together, gross profit carries most of the weight. That matters for what follows. If value follows gross profit, and free entry pushes gross profit down to the cost of building, then cheaper building reaches valuations directly.

What survives

A model in which every firm is worth what it cost to build cannot produce a company worth a billion dollars. Unicorns are exceptions to the symmetric equilibrium, and the model says exactly what kind of exception they must be. There are three ways out, and each is a way of having something rivals cannot buy at the new, lower price.

A cost advantage of your own. A common fall in the price of compute goes to customers. A fall in your cost of serving a customer, which rivals do not share, is a rent. In AI software this means learning to do the same task with less inference than competitors: routing easy steps to cheaper models, caching, using data from past tasks to shorten the next one. It is the one scale economy the rotation leaves standing, and it is proprietary only if it is learned from your own usage.

Differentiation that code cannot copy. In the model, markups come from how far apart products are. AI shrinks the distances that are made of features, since features are now cheap to clone. It leaves alone the ones made of other things: distribution and an audience that already trusts you, a position inside a customer's workflow that is costly to unwind, regulatory approval, a brand. These were always sources of advantage. What changes is that they are now nearly the only ones.

Demand-side scale. Network effects, where the product is worth more because others use it, are a scale economy that falling build costs do not erode. A word of caution on the most popular candidate: proprietary data. Andrei Hagiu and Julian Wright show that learning from customer data can create a competitive advantage. The network effects it creates are usually weaker, and less likely to lock customers in, than the classic kind. "We have the data" is a moat only if the data keeps improving the product faster than a rival could catch up.

One thing that gets scarcer as building gets cheaper is the capacity to put software to work inside organisations. I took this up in an earlier essay. It is another sign of the same shift: value moves away from what AI makes abundant and towards what it does not.

The objection: is building really cheaper?

The argument rests on one claim: AI lowers the cost of building software for everyone. The evidence for it is strong but uneven. In a controlled experiment, developers given GitHub Copilot finished a programming task 55.8% faster. Across three field experiments with nearly 4,900 developers at Microsoft, Accenture and a Fortune 100 firm, access to the same tool raised completed tasks by about 26%, with larger gains for less experienced developers. But in a 2025 trial by METR, experienced developers working on large open-source projects they knew well were 19% slower with AI tools. They had expected to be faster and still believed afterwards that they had been.

This does not rescue the old moats. It shows where they still hold. AI cheapens the building of new, self-contained products most, and the maintenance of large, complex systems least. Products that are quick to build are the ones most exposed to entry. Systems that are hard to build, integrate and maintain keep more of their protection, at least until the tools catch up with them.

What follows

For founders, the implication is uncomfortable but useful. The product is no longer the hard part, so it can no longer be the moat. A startup that can be built in three months can be copied in one. The question to ask on day one is not "can we build it?" but "what will we have in two years that a well-funded team with the same tools could not reproduce?" If the answer is only the product, the plan is a race with no finish line.

For investors, the cloud precedent suggests what comes next: more, smaller first cheques, faster abandonment, and returns even more concentrated in the few firms that find one of the three exits. Valuing AI software on revenue multiples inherited from classic SaaS will overpay for firms whose gross profit is being competed away.

The argument also makes predictions that can fail. Entry into AI software categories should rise as building costs fall. Gross profit per firm in those categories should fall even as margins move with compute prices. And the premium that investors pay should go to firms with distribution, workflow position or network effects, not to those with the best features. If the most valuable AI software companies turn out to be the ones that simply built the best product, the argument is wrong.