When your AI vendor gets bought without being bought

Your coding assistant, your automation layer and the model behind your customer replies are all supplied by companies small enough to be bought. Over the past two years the buying has stopped looking like buying. A large company pays for a licence to a startup’s core technology, offers jobs to the engineers who built it, sometimes puts equity in on top, and everyone involved says the startup remains independent. No merger is announced. The logo on your invoice does not change. The thing you were relying on changes anyway.

This guide is for the solo operator or small team on standard click-through terms, the kind you accepted without reading and certainly without negotiating. If you have a master agreement your lawyer redlined, you already have most of the protections described below, and your time is better spent rereading those. This is not a guide to predicting which deals get blocked, and it will not tell you whether any particular vendor is about to fall over. It tells you how to notice that a vendor has effectively changed hands, what your existing terms already permit them to do, and what an exit ramp costs when you build it before you need it.

The structure that avoids the word acquisition

The shape is consistent. On 21 August 2026, Nvidia agreed to pay the AI coding startup Poolside 6 billion dollars for a licence to Model Factory, the software Poolside built to train its own models, plus a separate 1 billion dollar investment at a 12 billion dollar pre-money valuation. Nvidia offered jobs to 109 Poolside employees. Poolside’s founders stayed, the licence is non-exclusive, and the company’s investor letter said the arrangement “is not an acquisition and it is not an acquihire” [8].

By the letter of the paperwork, that is accurate. A non-exclusive licence is not a merger. An individual engineer accepting an individual job offer is not a merger either, and 109 of them accepting 109 offers is still, formally, 109 hiring decisions. The result is that a very large amount of money moves and no acquisition filing follows. Antitrust lawyers describe these arrangements as frequently evading the pre-closing review established by the Hart-Scott-Rodino Act [7].

Nvidia did not invent this. In February 2026, senators Elizabeth Warren, Ron Wyden and Richard Blumenthal wrote to the FTC and the Department of Justice about three of them, naming Meta’s June 2025 deal with Scale AI, Google’s July 2025 deal with Windsurf, and Nvidia’s 20 billion dollar licensing deal with Groq in December 2025. The letter argued the deals “function as de facto mergers, allowing the companies to consolidate talent, information, and resources” and that “the FTC and DOJ should not allow these companies to avoid the typical reviews that your agencies apply to acquisitions and mergers” [3]. Whatever you think of the politics, the description of the mechanism is the useful part.

A regulator already called one of these a merger and let it through

The clearest read on what these deals actually are came from the UK, not the US. In 2024 the Competition and Markets Authority examined Microsoft’s hiring of former Inflection AI staff, which the CMA understood to amount to almost all of Inflection’s team, including two of its co-founders, alongside a series of arrangements between the companies that included a non-exclusive deal to license Inflection’s IP [1].

The CMA concluded that it had jurisdiction to review the transaction “because a relevant merger situation has been created”, found that “Microsoft has substantively acquired Inflection’s pre-Transaction FM and chatbot development capabilities”, and stated the general principle plainly: acquiring a team with relevant know-how, even without further assets, may fall within its merger control jurisdiction [1].

Then it cleared the deal. The CMA found that Inflection was not a material competitive constraint on the consumer chatbots developed by Microsoft, by OpenAI or by other competitors, and decided not to refer the transaction under section 22(1) of the Enterprise Act [1]. So the official finding is that a company can be absorbed in substance while continuing to exist on paper, and that this is often lawful.

For you that settles a question worth settling early. You are not waiting for a regulator to confirm that your vendor changed. A regulator may confirm it two years later and permit it anyway, which is a fine outcome for competition policy and no help at all when you are deciding whether to renew in November.

What leaves is the roadmap, not the login page

The practical loss in these deals is the people who know how the product works. A model or a tool is not only its weights and its repository. It is also the group who can fix the thing you filed a ticket about. When that group moves, the product keeps running, the login page keeps working, and the next three releases quietly become smaller.

Windsurf is the case with the clearest paper trail. In July 2025, Google did a deal for Windsurf that the senators later listed alongside the Meta and Nvidia arrangements [3]. On 14 July 2025, Cognition announced it had signed a definitive agreement to acquire Windsurf, and its own post describes what was included: Windsurf’s IP, product, trademark and brand, a business with 82 million dollars in annual recurring revenue, more than 350 enterprise customers and hundreds of thousands of daily active users [4]. If you were one of those enterprise customers, you had a new supplier inside a month, and nothing you did caused it.

The FTC’s January 2025 staff report on partnerships between cloud providers and AI developers found the same pressure points from the other direction. It concluded that these arrangements may affect access to inputs “such as computing resources and engineering talent”, and may “increase contractual and technical switching costs” for the smaller partner [2]. Then FTC chair Lina Khan said in the announcement that such partnerships can “create lock-in, deprive start-ups of key AI inputs, and reveal sensitive information that can undermine fair competition” [2]. Lock-in at your vendor becomes lock-in for you one layer down.

Your terms already let them hand you over

Before you assess any deal, read one clause in the terms you accepted. It is called assignment, and it is close to identical across the large vendors.

Anthropic’s commercial terms say that neither party may assign its rights or delegate its obligations without the other party’s prior written consent, “except that Anthropic may assign its rights and delegate its obligations to an affiliate or as part of a sale of all or substantially all its business” [5]. OpenAI’s business terms say OpenAI “may assign this Agreement to an Affiliate without notice or Customer consent”, and that either party may assign to a successor to substantially all of its assets or business, provided it gives at least 30 days prior written notice [6].

Read that in the direction that applies to you. You cannot transfer your agreement without asking. They can transfer yours without asking, or with a month’s warning. That is not sharp practice, it is the standard shape of a software contract, and it is the reason “we remain independent” is a statement about corporate structure rather than a promise to you.

The notice you are owed when the service itself changes is shorter still. OpenAI’s terms provide that if an update materially reduces functionality, OpenAI notifies the customer at the account email address, and within 5 business days of receiving that notice the customer may choose to terminate by giving 30 days written notice [6]. Anthropic may update its published rates “to be effective the earlier of 30 days after the updates are posted by Anthropic or Customer otherwise receives Notice”, and may update the terms themselves on the same 30-day footing [5]. Your window to react to a material change is measured in days, and it starts when an email lands in an inbox you may not read.

Three things the announcement will not tell you plainly

The first is who is staying, by name rather than by role. “The founders are staying” and “the team that built the model is staying” are different sentences with different consequences. In the Poolside deal the founders stayed and 109 employees received Nvidia offers [8]. Both facts are in the same announcement, and only one of them is in the headline.

The second is whether the licence is exclusive. Poolside’s licence to Nvidia is non-exclusive, which means Poolside can license the same software to someone else [8]. A non-exclusive licence is genuine evidence that the seller still has a business. An exclusive one means the technology now has a single customer, and the seller’s remaining product is a courtesy.

The third is where the money goes. Poolside said it intends to distribute the 6 billion dollars to its investors by the end of 2027 [8]. Cash returned to investors is cash that is not being spent on the product you use. This is not a scandal and nobody hides it, but it is stated in the investor letter rather than the blog post, and it tells you more about the next two years of releases than any statement about independence does.

Build the exit ramp while nothing is wrong

The right time to work out what leaving costs is a quiet Tuesday, not the morning after an announcement. Three steps cover most of it.

Export what you can actually export, in a format something else can read. Your prompts, your saved outputs, your Notion pages, your Zapier or n8n workflow definitions, the configuration for whatever your Cursor or Claude setup depends on. If the answer is that the valuable part exists only inside the vendor’s interface, you have found your real dependency and it is worth fixing regardless of who owns the company.

Name your two alternatives in writing. Not a category, two specific products you could move to, with the price of each. Doing this once, calmly, converts a future emergency into an afternoon of work. If your workload is model calls behind an API, the alternatives are close substitutes and the move is cheap. If it is a bespoke agent built against one vendor’s tooling, it is not, and you should know which situation you are in before you need to know.

Price the move. The number is small enough to be worth writing down and large enough to change what you decide.

What still goes wrong

Regulators have noticed and are still deciding what to do about it. FTC chair Andrew Ferguson said in January 2026 that acqui-hiring “has become a big enough deal” that the agency is “beginning to look very closely at how these things work, including determining whether we need to promulgate additional guidance here in the coming months” [7]. Commissioner Mark Meador has raised the concern that firms may acquire talent not to use it but to stop rivals from having it, and the acting head of the DOJ’s antitrust division called acqui-hires a red flag in March 2026 [7]. The advice from the antitrust bar is to monitor developments and to push for rules that separate ordinary talent mobility from consolidation [7], which is a courteous way of saying the line has not been drawn yet. Plan on the structure continuing while that gets settled.

Nothing here tells you the product will get worse. It sometimes gets better. Cognition said that in the immediate term the Windsurf team would continue to operate as it had been, and gave the editor full access to the latest Claude models [4]. A startup running out of money and getting absorbed by a company with a working business is often the better outcome for you, and a vendor that stays proudly independent can still ship less next year than it did last year. Absorption is a reason to check, not a reason to leave.

The hardest limit is that you cannot verify the thing that matters most. Headcount figures come from reporting and from the companies themselves, not from filings you can inspect, and no announcement will tell you which of the engineers who left were the ones who mattered. What you can verify is your own position: what your contract permits, what you can export, what a move would cost, and whether the release notes six months from now still look like the ones that convinced you to sign up.