AI Is Buying AI: What 30 Deals in 2026 Tell Founders

For two years the story in AI M&A was Big Tech reaching into frontier labs and pulling out the people. Microsoft and Inflection. Google and Character.AI. Amazon, Adept and Covariant. The structure was a licence plus an acquihire, the motive was regulatory, and we covered it at the time in AI Roll-Ups Are the New PE Playbook.

In 2026 the buyer side changed. The companies doing the acquiring are, overwhelmingly, AI companies themselves. Not Microsoft. Not Google. OpenAI, Anthropic, Cohere, Harvey, Legora, Sierra, Nscale, Cyera. Businesses that were raising seed rounds three or four years ago are now running corporate development functions and closing acquisitions at a rate that would look aggressive for a listed strategic.

We have just finished verifying and publishing that wave into the Exit Mode M&A archive: thirty AI-on-AI acquisitions announced between January and September 2026, each one checked back to a primary or source-grade reference before it went in. That gives us something more useful than a narrative. It gives us a sample.

Here is what the sample says, and what it should change about how you think about your own exit.

Who is actually buying

Ten buyers accounted for all thirty deals. The distribution is lopsided:

OpenAI, 11 deals. Legora, 5. Harvey, 4. Sierra, 3. Cohere, 2. Then one each from Anthropic, Nscale, Cyera, Sword Health and SpaceX.

Twenty-one of the thirty had completed by the end of September. The other nine were announced with firm intention and were still working through closing or clearance. This is not a pipeline of speculative talks. It is a market that is already transacting.

The detail that matters most is the one nobody puts in a headline: with the single exception of SpaceX, which listed in June, every buyer on that list was a private, venture-funded company at the point it signed. The dominant acquirer in AI right now is not a public company with a market capitalisation you can look up and a currency you can sell. It is a private company with a valuation set by its last round.

Hold that thought. It changes what a term sheet from one of them is actually worth.

OpenAI is running a product roadmap through the M&A market

Eleven acquisitions in nine months is roughly one every three and a half weeks. What is striking is not the pace but how legible the logic is. Line the targets up against the products and the roadmap reads itself:

Into Codex: Astral, the developer-tools company behind uv, Ruff and ty. Ona, for AI-agent cloud execution and orchestration. Into ChatGPT: NextSlide, an AI presentation software team. Into Prism: Crixet, a cloud LaTeX and scientific writing platform. Into the safety and evaluation stack: Promptfoo, for evaluation, red-teaming and compliance. Into new verticals: Torch Health for unified medical records at a reported $60 million, Hiro Finance in personal finance, Glass Imaging in AI imaging software and hardware at a reported $300 million. Into delivery capacity and geography: Tomoro in the UK for enterprise AI deployment, Convogo for applied business software, and the open-source agent project OpenClaw out of Austria, classified as an acquihire.

Two of those eleven had a disclosed price. Nine did not.

If you are a founder, that list is the clearest statement of buyer intent you will see all year. OpenAI is not buying revenue. It is buying specific missing pieces of named products and the small teams that built them, and it is doing it faster than it could hire or build them. The thing being valued is the gap you close, not the P&L you bring.

Legal AI is consolidating faster than anything else

Nine of the thirty deals, almost a third of the entire sample, came from two legal AI companies buying other companies: Legora with five and Harvey with four.

Legora bought Qura in Sweden for AI-native legal research, Walter AI in Canada for legal-agent capability and a Canadian footprint, Graceview in Australia for regulatory intelligence, Cadastral in the United States to enter commercial real estate legal intelligence, and Wexler AI in the UK for litigation fact intelligence and a London engineering hub. Harvey bought Hexus, Lume AI for customer data and AI-system integration, Benchmark to move into asset-management decision infrastructure, and Guardrails AI for agent safety, testing and monitoring.

Not one of those nine deals disclosed a price.

Read them together and the strategy is unmistakable. Two venture-funded companies are racing to own a professional-services category before the other one does, and they have decided the fastest route is to buy the adjacent workflow and the local team rather than build either. Four of Legora's five targets were outside the United States. This is land-grab M&A, and it is being run by companies that are themselves still private.

If you run a small, sharp software business adjacent to a professional services workflow, in law, accountancy, insurance, surveying or compliance, you are closer to this market than you think.

Where the money actually gets named

Seven of the thirty deals disclosed a value. Those seven are worth listing in order, because the pattern in them is sharper than anything in the undisclosed majority:

SpaceX and Cursor, $60 billion, all stock. Nscale and Anyscale, $1.65 billion, to own more of the AI compute and workload stack. Cyera and Oasis Security, $1 billion, combining data security with non-human identity security. Anthropic and Coefficient Bio, a reported $400 million in stock. OpenAI and Glass Imaging, a reported $300 million. Sword Health and Kaia Health, $285 million. OpenAI and Torch Health, $60 million.

Strip out the SpaceX outlier and the other six disclosed deals total roughly $3.7 billion between them.

Now look at what those six have in common. Compute infrastructure. Security. Biotech and healthcare data. Medical records. Imaging. Every one of them sits in a category where the asset is infrastructure, regulated data, or both, and where the buyer could point at something tangible to justify the number. Prices get named when there is a defensible asset behind them. Prices stay quiet when what changed hands was a team and a feature.

Twenty-three undisclosed prices is the real headline

Seventy-seven per cent of this market transacted without a public number. That is not an accident of reporting and it is not modesty. It is a structural feature of acquihire-adjacent M&A, and it has a direct consequence for you.

It means there are no comparables. If one of these buyers approaches your business, you cannot go and look up what the last four deals in your category went for, because nobody published them. The buyer, who has done this nine or eleven times, knows exactly what the range is. You do not. That asymmetry is the single biggest disadvantage a founder carries into one of these conversations.

There is only one workable response, and it is the same answer as always, just more urgent here. You negotiate from your own numbers, not from the market's. What would this business be worth to you if you kept it and ran it for five more years? What is the cost to the buyer of building what you have built, in time rather than money? What does your revenue, retention and team cost in a straight hiring market? Those three figures are available to you, they are defensible, and in a market with no comps they are the only anchors you have.

If you cannot answer all three on the day a buyer calls, you are negotiating with your eyes shut.

You are being offered private paper

We wrote recently about how Musk weaponised SpaceX stock as an acquisition currency, and the central point of that piece applies here in reverse.

SpaceX could buy Cursor for $60 billion in all-stock consideration because, four days before, its shares had become a listed, liquid, exchange-traded security. The seller received something with a transparent price and a route to cash. That is the exception in this sample, not the rule.

The other twenty-nine deals were done by private companies. Where consideration was reported, it leaned on stock: Anthropic and Coefficient Bio was reported as a stock deal. Private stock in a fast-growing AI company may turn out to be the most valuable thing you ever hold. It may also be paper in a business whose last round was priced in a very different market, with no liquidity event scheduled, no public price to mark against, and a transfer restriction that stops you doing anything about it.

If you are offered equity by one of these buyers, the questions are not optional. What was the last round, when, and at what price? Who set that price, and did any new outside money lead it? What are my registration and transfer rights? Is there a tender offer or secondary programme, how often does it run, and have employees actually been allowed into it? What happens to my shares if the company raises a down round, or sells, before I can sell? And the one founders skip: what proportion of the headline number is cash on completion, and would I accept the deal if the stock went to zero?

If the honest answer to that last question is no, you are not being paid what you think you are being paid.

Europe is inside the buyer universe now

Thirteen of the thirty targets were outside the United States. Germany supplied four of them: Aleph Alpha to Cohere, Reliant AI to Cohere, Ona to OpenAI and Kaia Health to Sword Health. The UK supplied two, Tomoro to OpenAI and Wexler AI to Legora. The rest came from Sweden, France, Austria, Australia, Canada, Japan and Israel.

The motive is rarely revenue. Sierra bought Opera Tech to expand enterprise AI agents in Japan and Fragment to add operational automation and a French presence. Legora took a London engineering hub with Wexler. Cohere framed Aleph Alpha explicitly as European expansion and went on to buy biopharma AI company Reliant AI for sovereign enterprise work. These buyers are crossing borders for teams, for a regional foothold, and for the credibility of being local in markets that increasingly care about where their AI is built and hosted.

For UK and European founders the practical consequence is blunt. Your most likely AI acquirer is no longer a Fortune 500 corporate development team that will take nine months and a banker to reach you. It is a three-year-old company on another continent that will find you through your product, move in weeks, and pay in a currency you cannot immediately value.

What the incumbents are doing about it

While the AI natives buy each other, the large corporates are making a different move, and it is worth watching because it is where most mid-market businesses will actually get bought.

On 31 August 2026, ITOCHU announced a tender offer for DENTSU SOKEN, the Tokyo-listed IT services and consulting subsidiary of Dentsu Group, at 2,880 yen per share for a planned total of up to 215.2 billion yen, roughly $1.3 billion. The structure takes DENTSU SOKEN private into a joint venture between Dentsu Group and ITOCHU's vehicle, which will hold about 38 per cent, with a delisting from the Tokyo Stock Exchange to follow. The European Commission is reviewing it as merger case M.12635, and the offer is targeted to open in early November once clearances are in.

ITOCHU's own stated rationale is the interesting part. It describes a market in which digital technologies such as AI are evolving so quickly that the winning model is end-to-end: understanding a client's business problem, then designing, building and running the system that solves it. So it is not buying a model lab. It is buying delivery capacity, consulting depth and customer relationships in the manufacturing and financial sectors, and pairing them with its own IT and retail media assets.

That is the quieter half of AI M&A, and for most founders it is the more relevant half. The AI natives are buying capability. The incumbents are buying the businesses that implement it for customers. If you run a services, integration or vertical software business, you are more likely to meet the second buyer than the first.

Five things to take from thirty deals

1. Diligence the buyer as hard as they diligence you. Nine of these ten acquirers were private companies spending paper of their own creation. You are being asked to take a position in their business. Read their last round like an investor would, because for the length of any lock-up, that is exactly what you are.

2. Accept that there are no comps, then stop negotiating as if there are. Twenty-three of thirty prices were never published. Build your own floor from hold value, build cost and replacement cost before anyone calls, and hold it.

3. Be a named gap on somebody's roadmap. Every deal in the OpenAI list maps to a specific product it was folded into. The businesses getting bought are not the ones with the best general story. They are the ones that are visibly the missing piece of a particular buyer's plan. Work out whose plan you complete, and make that legible.

4. Infrastructure and regulated data get a published price. Teams and features do not. If you want a number you can point at, the asset has to be something a buyer can defend in a board paper: proprietary data, a regulatory position, security, compute. If what you have is a very good team and a neat product, expect an undisclosed deal and price it yourself.

5. Be ready before you are approached. Twenty-one of thirty deals were already complete inside the window. These processes move in weeks, not quarters. Clean accounts, a tidy cap table, assigned IP, documented customer contracts and resolved contractor agreements are not things you can assemble after the first call without losing leverage.

The shift underneath all of it

The thing that has genuinely changed in 2026 is not the volume of AI M&A. It is who holds the chequebook. For a decade, an acquisition offer in technology came from an incumbent with a public share price, audited accounts and a corporate development process you could research. Increasingly, it comes from a company that is younger than your own, is not required to publish anything, has never completed a liquidity event, and is paying you in its own valuation.

That is not a reason to refuse the call. Some of these will turn out to be the best deals anyone signed this decade. It is a reason to do the work that makes you the better-prepared party in the room: know what you are worth on your own numbers, know exactly what the consideration is and when you can touch it, and know whose roadmap you complete before they tell you.

Thirty deals, twenty-three of them silent on price, nine buyers who did not exist in their current form five years ago. That is the market you are selling into. The founders who do well in it will be the ones who treated the preparation as the deal, long before a buyer ever showed up.

Adam J. Graham