On May 19, 2026, Google DeepMind hired more than twenty researchers from Contextual AI and paid roughly $100 million for a license to the company's technology. The deal, reported by Bloomberg, was not an acquisition. Contextual AI, the startup co-founded by Douwe Kiela in 2023, still exists. It still has servers. It still has a board. What it does not have, as of May 19, is Kiela or the research staff that made it a target.
This was the fourth lab-to-startup deal in seven days. Anthropic, Mistral, Google DeepMind, and Meta each closed a similar arrangement in the week of May 18 to May 24, 2026. Each one followed the same template. Hire the people. License the code. Leave the company.
The corporate form has come apart at the seams. The acquisition, which for a century was an all-or-nothing transaction between two indivisible legal entities, is now a procurement process across three separate categories of asset. Talent is one category. Technology is another. The corporate shell, with its customers and brand and bank account, is the third. Each category prices independently. Each clears at a different number. Each ends up with a different owner.
Every founder and operator with a venture-backed AI company is about to discover that the company they thought they were building is, in the new accounting, a portfolio of three loosely connected line items. Two of them might be worth selling. The third might not.
The Inflection Template
The shape was set in March 2024. Microsoft hired Mustafa Suleyman and most of the staff of Inflection AI, and paid roughly $650 million for a non-exclusive license to Inflection's models, including Inflection-2.5. Inflection remained, on paper, an independent company. Reid Hoffman stayed on the board. A new CEO was installed. The Pi chatbot was wound down.
The deal was unusual enough that Fortune called it "the most important non-acquisition in AI," and the U.K.'s CMA opened a Phase 1 review of the arrangement on July 16, 2024. The CMA closed the review without further action, ruling that Microsoft had not acquired a "relevant merger situation," a verdict that surprised no one and changed the playing field for everyone.
Inflection had given the M&A bar a procurement primitive: Hire And License Out, or HALO. Hire the founders. License the IP. Leave the shell. The U.S. Hart-Scott-Rodino Act, which requires regulatory review of mergers above a size threshold, has nothing to say about a non-exclusive license and a payroll move. The Federal Trade Commission opened an inquiry into the Inflection deal in early 2025, but the structural innovation had already been absorbed into the M&A playbook.
What followed was a year of variations. Amazon hired most of Adept's team in June 2024, leaving the company technically independent. Google hired Character.AI's founders Noam Shazeer and Daniel De Freitas back into DeepMind in August 2024 and licensed the model weights. Microsoft and Inflection became the proof of concept. Amazon and Adept became the proof of replicability. Google and Character became the proof that even the labs that had spun talent out were willing to buy it back at non-merger prices.
By the end of 2025, the HALO structure had become the default.
The Receipts Got Bigger
On December 24, 2025, Nvidia signed a non-exclusive licensing agreement with Groq worth roughly $20 billion, the largest deal in Nvidia's history. Jonathan Ross, the Groq CEO and co-founder, joined Nvidia along with most of the senior leadership team. Groq itself stayed standing as an independent company under Simon Edwards, the former CFO. Senator Elizabeth Warren and Senator Ron Wyden wrote to FTC Chair Andrew Ferguson and DOJ Antitrust head Gail Slater in early February 2026 asking for an investigation. CNBC quoted one analyst calling the structure designed to "keep the fiction of competition alive."
The fiction is the point.
Meta's $14.3 billion investment in Scale AI in June 2025 followed the same logic. Meta got Alexandr Wang as Chief AI Officer and a 49% non-voting stake. Scale got Jason Droege as interim CEO. The data labeling business kept running. The legal entity stayed sovereign. StartupHub.ai, recapping the same week of May 2026 lab consolidation, also reported that Anthropic had closed a $30 billion growth round led by Sequoia, Dragoneer, Altimeter, and Greenoaks, bringing the lab's valuation above $900 billion. The labs are buying everything that moves, the deals are getting larger, and almost none of them are merger filings.
Google's split with Windsurf in July 2025 is the most cynical case so far. Google paid $2.4 billion for a non-exclusive license to Windsurf's IDE technology and the services of CEO Varun Mohan, co-founder Douglas Chen, and roughly 40 senior engineers. The remaining 250 Windsurf employees were not part of the deal. Their equity vesting was not waived. Their stock options were not converted. They learned about the deal the same way everyone else did, from a press release. Cognition Labs picked up the remaining IP, product, brand, and headcount two weeks later, and at least had the decency to waive the vesting cliffs.
The pattern is now legible. The most valuable parts of a startup are removable. They get removed. The rest is left to compost or gets scooped up by whoever wants to inherit the customer list.
Why the Shell Stays Behind
The shell stays behind because, in the HSR architecture, the shell is the company. The Hart-Scott-Rodino Antitrust Improvements Act of 1976 was written to police the acquisition of corporate entities. It assumed that corporate entities and corporate value were the same thing. If you bought the company, you got the value. If you didn't buy the company, you didn't get the value.
That assumption is dead.
A modern AI startup's value is concentrated in three places: the founders, the model weights and training code, and the customer book. The first two are portable. The third is sticky, but only because it is socially expensive to break a customer contract. The HSR test looks at corporate entities. The actual value lives in payroll and IP. The structural mismatch is the loophole.
FTC Chair Andrew Ferguson told Bloomberg on January 16, 2026, that the agency is examining reverse acqui-hires under the existing statute, looking for transactions that should have been filed and were not. Mintz's February 20, 2026 client alert notes that the FTC may treat substantially identical arrangements as "merger evasion" if the parties have structured them to avoid HSR notification. The remedy, if one comes, will involve unwinding payroll moves a year after the fact, an idea that legal scholars find amusing for the same reason that surgeons find Krazy Glue amusing.
The current administration has so far preferred case-by-case enforcement to a blanket rule. The American Action Forum's reading is that this approach will lead to two or three test cases over the next eighteen months and effectively no general guidance for the rest of the market. By the time those cases work through the courts, the HALO structure will be more entrenched than the email signature block.
What the Decomposition Means for Your Company
The corporate form is now decomposable. This is a description of last week, captured in present tense.
If you run a private company with a serious AI capability, a small senior team, and a customer base, your company is no longer the asset. It is the holding pattern for three separate assets that will be priced and traded independently. The implications, in plain English, are uncomfortable.
Your founding team has a wholesale price
The market price for senior AI engineers and researchers cleared, in the Windsurf deal, at roughly $60 million per person for the top 40 names. The Contextual deal cleared at roughly $5 million per researcher for the top 20. The Inflection deal cleared at roughly $9 million per head. Treat these numbers as talent acquisition costs paid to a third party, namely the company the people used to work for, in exchange for the right to extinguish the employment relationship and start a new one.
This price is now legible to your team. If they are in the top 20 names at any AI lab between roughly $100 million and $5 billion in valuation, they have read the Bloomberg article. They know what they are worth. They know that a labor offer wrapped in a license is the standard exit. They know that the new path to an outsized personal payout runs through a frontier lab deciding their skills are needed.
Call it a moat if you want. What you are actually building is a holding pen until the buyer comes for the seven names that matter.
Your IP has a non-exclusive license value
The license values in these deals are interesting because they are non-exclusive. Microsoft paid roughly $620 million to use Inflection-2.5 alongside any other licensee Inflection cared to sign. Google paid $2.4 billion for non-exclusive use of Windsurf's technology, knowing Cognition would buy the rest. Nvidia paid roughly $20 billion for the right to use Groq's LPU patents while Groq, in theory, could license them to anyone else.
The non-exclusive structure tells you what the acquirer actually values. What the acquirer values is the speed of getting the team on payroll. The IP license is a cover story for the FTC. The price is the cost of legitimacy.
This is good news and bad news for your company. The good news is that there is a real market for non-exclusive licenses to your stack. The bad news is that the price tracks how much regulatory cover the acquirer needs. It tracks your stack's value only incidentally.
Your residual shell is the bag holder
In every one of the HALO deals, the residual company has continued to operate. Inflection has a CEO. Scale has a CEO. Groq has a CEO. Contextual has a CEO. None of these CEOs are the founders. All of these companies are smaller, slower, and weaker than they were the day before the deal. They are run by the second-tier operations leadership, who got promoted, and they serve a customer base whose contractual relationships predate the strip mine.
If you are an investor with a check in one of these residual shells, you own the third asset, the leftovers. The HALO deal does not technically reduce your equity. It just removes the people and the IP that gave the equity its meaning. You still have your stock certificate. You just have less to point to when you ask what it represents.
If you are a customer of one of these residual shells, you are now buying a service from a company that has different leadership, less talent, and a contractual obligation to honor your renewal. You signed up for the founder's vision. You are getting the operations team's interpretation of it.
The Architecture Question
A decomposable company is an architectural choice. A non-decomposable company is also an architectural choice. The default, given a normal startup employment agreement and a normal cap table, is the decomposable one. The acquirers know this. They have built a transactional model around it. The founders, in most cases, have not yet realized that the decomposition is the deal.
There are two questions a CEO has to answer if they want their company to be more than a holding pattern for a future strip mine.
The first is whether the value of your company is concentrated in a small enough number of people that a HALO deal makes sense. If you can name the seven people who, if hired by Anthropic or Google tomorrow, would leave your current investors a footnote, then your company is HALO-shaped. The right defense is structural. Build the load so that no HALO offer maps onto a manageable headcount. Distribute the work. Document the systems. Multiply the bus factor until the strip mine becomes uneconomic. Most companies have not done this. The ones that have done it are not interesting acquisition targets, which is exactly the right answer for a company that wants to stay a company.
The second question is whether your IP has value independent of the people who made it. A model checkpoint has value if other people can run it, fine-tune it, and improve it. A codebase has value if other people can read it and ship features against it. A patent portfolio has value if other people can prosecute and defend it. Most AI startups have stacks that nobody outside the founding team can operate. This makes the IP look like ballast in a HALO deal. The license price is high because the acquirer needs the regulatory cover; the operational use of the IP, after the talent moves, is roughly zero. If you want your IP to clear at a number that means something, make sure it can run without your founders.
The decomposition is the structural reality of how AI value is distributed. Acquirers have built their playbook around it. Founders have not yet caught up. The question is whether your company is built to be sold as one piece or as three.
Why This Is The Critical Strategic Question Right Now
There is a moment, somewhere between Series A and Series B, where the founder has to decide what kind of company they are building. The default path, for AI companies, leads to a HALO exit at a number that looks great on a press release and a fractional return to the cap table at large. The alternative path leads to a company that has institutional value beyond its founders and that can be sold whole, taken public, or held for compounding returns.
The default path is fine for some founders. It is, in the median case, a faster exit than the IPO path and a cleaner one than a traditional acquisition. It also, by design, leaves the customers, the late-stage investors, and most of the employees behind. Ask the 250 Windsurf engineers what the median outcome feels like from the wrong side of the press release.
The alternative path requires architecting the company to be non-decomposable. It requires investing, before the demand for the HALO deal arrives, in the systems and the people and the IP structures that make the company more than the sum of its founders. It requires deciding that the seven names on the org chart will not, in fact, be the company. This is an architecture question. Hiring won't fix it, and neither will retention bonuses. It is the kind of thing that has to be designed at Series A or it never gets done.
If you are a CEO or founder of a serious AI company, the question of whether you want to be HALO-shaped or built to last is in front of you now. The acquirers have made their preferences known. The FTC is not coming to save you. The structure of every deal in the next two years will assume that your company is decomposable until you prove otherwise.
Architecting that proof is what Agor AI Advisory does. We work with founders and senior operators to design the technical, organizational, and IP architecture of a company that cannot be strip-mined. We help you decide whether you want to be HALO-shaped or built to last, and we help you build the systems that make the choice real. We have seen too many founders sleepwalk into a transaction that they would not have signed if they had understood the decomposition. Buying a tool will not fix this. Hiring an investment banker will not fix this. The decomposition is now the default, and designing past it is the work.
Sources
- [Google Hires Staff From Bezos-Backed Contextual AI In Licensing Deal, Bloomberg, May 19, 2026](https://www.bloomberg.com/news/articles/2026-05-19/google-hires-staff-from-bezos-backed-contextual-ai-in-licensing-deal)
- [Four labs, four acquisitions in five days: the consolidation signal hiding in plain sight, StartupHub.ai, May 2026](https://www.startuphub.ai/ai-news/ai-news/2026/four-labs-four-acquisitions-ai-consolidation-may-2026)
- [Cognition AI to purchase fellow coding startup Windsurf after Google's $2.4B acqui-hire deal for founders, PitchBook, July 2025](https://pitchbook.com/news/articles/cognition-purchase-windsurf-after-google-2-4b-acqui-hire)
- [Nvidia buying AI chip startup Groq's assets for about $20 billion in its largest deal on record, CNBC, December 24, 2025](https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html)
- [Nvidia-Groq deal is structured to keep 'fiction of competition alive,' analyst says, CNBC, December 26, 2025](https://www.cnbc.com/2025/12/26/nvidia-groq-deal-is-structured-to-keep-fiction-of-competition-alive.html)
- [Sen. Warren, others urge FTC, DOJ to scrutinize tech AI 'acqui-hiring' deals for antitrust violations, CNBC, February 4, 2026](https://www.cnbc.com/2026/02/04/sen-warren-others-urge-ftc-doj-to-scrutinize-tech-acquihire-deals.html)
- [AI Acquihires Under Fire: FTC Signals HSR Scrutiny, Mintz, February 20, 2026](https://www.mintz.com/insights-center/viewpoints/54731/2026-02-20-ai-acquihires-under-fire-ftc-signals-hsr-scrutiny-ai)
- [Why Microsoft's surprise deal with $4 billion startup Inflection is the most important non-acquisition in AI, Fortune, March 19, 2024](https://fortune.com/2024/03/19/microsoft-surprise-deal-inflection-ai-mustafa-suleyman-reid-hoffman-questions/)