How AI Companies Get Valued, and Acquired Without Being Acquired

The headline numbers in AI stopped tracking anything familiar about a year ago. A company with almost no product raised at a valuation that would place it among the thirty largest in the S&P 500. Frontier labs are priced at twenty to thirty five times their revenue while mature software businesses trade at three to five. The largest technology companies are paying billions to hire the founders of startups without acquiring the startups. None of this looks like the market anyone learned about in business school, and yet the logic underneath is more readable than the numbers suggest.

There are two AI markets, not one

It helps to separate the frontier labs from everyone else, because they are priced on completely different logic.

At the top sit OpenAI, Anthropic, and xAI. By the middle of 2026 Anthropic had raised at roughly 960 billion dollars and OpenAI at roughly 850 billion, with xAI a good distance behind. Anthropic reported a revenue run rate near 47 billion on a gross basis, OpenAI somewhere in the twenties depending on how you count. Together, OpenAI and Anthropic absorbed close to half of every venture dollar deployed worldwide in the first half of the year. That is not a sector. That is two companies and a very long tail.

Everyone else lives in a different market. Application-layer startups, developer tools, vertical agents, and most infrastructure companies still get valued on something close to normal software math, adjusted upward for growth. The gap between the two markets is the whole story. The frontier is priced as if one or two winners will capture a category the size of electricity. The rest is priced as software.

Why the multiples look absurd, and the case that they are not

A software company trading at four times revenue is being told the market expects steady, defensible growth. A company trading at thirty times revenue is being told something different: that today's revenue is close to irrelevant, and what matters is the slope.

The bull case has three parts. First, the growth is genuinely without precedent. Anthropic went from about a billion dollars of run rate to more than forty in under eighteen months, the fastest revenue scaling in the history of software. Second, the buyers are strategic rather than financial. Microsoft, Amazon, Nvidia, and sovereign funds are writing checks to secure compute, distribution, and a seat at the table, and they do not need the return math to work the way a venture fund does. Third, if you believe the models keep improving on the current curve, the addressable market is not the software industry. It is a large slice of all knowledge work, and at that framing even these numbers look cheap.

The bear case is simpler. The multiples assume the slope holds, and slopes rarely do. Gross margins on frontier model serving are thinner than software investors are used to, because inference costs real money on every call. A large share of the strategic capital is circular, which is worth its own section. And a category with two or three credible players is one price war away from looking less like a monopoly and more like the airline industry.

Both cases are coherent. That is what makes this market genuinely hard to read rather than obviously a bubble.

The reverse acqui-hire, explained

The more revealing story is not the valuations. It is how the largest companies are buying their way in.

Between early 2024 and early 2026, Google, Microsoft, Amazon, and Meta spent well over twenty billion dollars hiring the founding teams of AI startups without acquiring a single one of those companies. The structure is consistent. The acquirer hires the founders and a chosen slice of the research team, signs a large nonexclusive license for the startup's technology, and leaves the corporate shell standing with its remaining staff, its investors, and a pile of cash.

• Microsoft and Inflection: Mustafa Suleyman and most of the team moved to Microsoft for a licensing fee around 650 million dollars.
• Amazon and Adept: the founders and key researchers hired, the technology licensed, the company left to carry on with what remained.
• Google and Character.AI: founder Noam Shazeer returned to Google alongside a licensing deal reported in the billions.
• Google and Windsurf: roughly 2.4 billion dollars to bring the chief executive and co-founders into DeepMind, after OpenAI's planned acquisition of the same company collapsed and a third company bought the remainder.
• Meta and Scale AI: about 14.3 billion dollars for a 49 percent stake, with founder Alexandr Wang moving over to run a new superintelligence lab. The largest version, and the least disguised.

The reason these deals are structured this way is not subtle. A full acquisition of that size triggers antitrust review under Hart-Scott-Rodino. A hiring spree plus a licensing agreement, at least until recently, did not. The structure gets you the people and the intellectual property while skipping the part where a regulator can say no.

Who pays for the structure

Every one of these deals produces a similar aftermath. The founders and the early investors are made whole, often generously. The company left behind keeps operating on paper, but without its leadership and its best researchers it is rarely a going concern in any meaningful sense. And the employees who did not make the list, the ones whose equity was supposed to be the reward for the risk, frequently end up with very little, because there was no acquisition to trigger a payout.

Regulators have noticed. In February 2026 a group of senators asked the FTC and the Justice Department to investigate the Meta, Google, and Nvidia deals specifically, and the FTC has said it will look at whether these structures are built to escape merger review. Whether that leads anywhere is an open question, but the assumption that a licensing deal is invisible to antitrust is no longer safe.

The competitive concern is straightforward. If the standard outcome for a promising AI startup is to have its team absorbed by an incumbent before the company matures, the incumbent never actually has to face the competitor. The rival is removed from the board while it is still small, and the market consolidates without a single merger being filed.

What Nvidia is actually doing

Nvidia sits at the center of a different pattern worth understanding on its own. Over the past year it has taken investment stakes in OpenAI, xAI, Mistral, and a long list of others. In September 2025 it announced a framework to invest as much as 100 billion dollars in OpenAI, later finalized as roughly 30 billion inside a larger round. It has also helped organize hundreds of billions of dollars in outside financing for data center buildouts.

The criticism is that much of this is circular. Nvidia invests in a company, the company spends the money on Nvidia chips, the revenue returns to Nvidia, which books it as growth and uses that growth to justify the next investment. Nvidia's leadership rejects the framing and notes that its commitments are small relative to its cash flow. Both things can be true. It is vendor financing at enormous scale, which is a normal move for a dominant supplier during a capital-intensive boom, and it also makes the demand signal harder to trust, because some of the demand is being funded by the supplier.

How I read it as someone building

A few things I take from all of this, as someone who leads an engineering organization and has to make build versus buy calls with real budgets.

• The frontier is not a market you compete in. Unless you are one of three companies, you build on top of the labs, not against them, and your architecture should assume the model underneath you changes vendors at least once. Design for that now.
• Strategic capital distorts the signal. When you assess an AI vendor's staying power, separate revenue from real customers from revenue that comes from an investor who is also a supplier or a partner. The second kind is less durable than it looks on a growth chart.
• The acqui-hire pattern is a talent risk, not just a headline. If you depend on a small AI startup for something important, the failure mode is not bankruptcy. It is that their four best people get hired away in a deal that leaves you with a license and no roadmap.
• If you are an engineer choosing where to work, understand which side of the structure you would be on. In most of these deals, being on the list is life changing and not being on it means your equity was a lottery ticket that did not pay.

None of this means the technology is overhyped. The capability gains are real and many of the businesses being built on top of them are real. But the capital structure around AI right now is doing something specific. It is concentrating the frontier into a very small number of hands, moving talent into the incumbents faster than regulators can track it, and funding a meaningful share of its own demand. Those are the things to watch, more than any single valuation headline.

If you want a sense of how I think about building on top of this rather than betting against it, my other writing and my projects get into the specifics.