There was no press release, no blog post, no Sundar Pichai thread. The confirmation that Google had completed a deal worth more than $1.5 billion came from the least glamorous place in technology: LinkedIn. Tamay Besiroglu, cofounder and until recently chief executive of the San Francisco startup Mechanize, now lists himself as a research scientist at Google DeepMind. More than a dozen of his former employees have updated their profiles the same way, most of them landing on Google's midtraining teams — the decisive layer that turns a raw pretrained model into one that can actually write code. Business Insider, which first reported the talks in August, confirmed on September 11 that the transaction has closed. Google and Besiroglu declined to comment. Final terms were not disclosed.
What Google did not do is buy Mechanize. The company still exists. It still has a CEO — Guive Assadi, previously chief of staff, who updated his own title to match. It still holds its technology, which Google licensed on a non-exclusive basis. On paper, nothing merged. In practice, the founder, the core engineering team and the know-how that made Mechanize worth anything now sit inside Google DeepMind.
The economics are worth pausing on. Mechanize was founded in April 2025. It raised $9.1 million in seed funding at a $500 million post-money valuation — a number that drew ridicule at the time for a company with a handful of employees and no product. Seventeen months later, Google paid three times that valuation, and more than 160 times the seed round, for the right to hire the staff and license the code. Few venture outcomes in AI have compounded faster, and almost none of them involved an actual acquisition.
What Mechanize actually built
The startup's founding pitch was maximalist to the point of parody. Besiroglu announced it on X in April 2025 with a single line: Mechanize will build virtual work environments, benchmarks, and training data to enable the full automation of all work. He sized the market by adding up global wages — roughly $18 trillion a year in the United States alone. Besiroglu had previously cofounded Epoch AI, the research group known for tracking compute trends and evaluating models, and the pivot from measuring AI progress to accelerating it drew sharp criticism from his former community.
The actual product is narrower and far more useful than the manifesto. Mechanize builds reinforcement-learning environments in which coding agents perform real software engineering tasks — shipping features, deploying applications, chasing bugs — then scores how well they did. Those scores become the reward signal frontier labs use to train their coding models: picks-and-shovels infrastructure for the one capability every lab is competing on.
That explains the price. Google has spent 2026 fielding models that rivals mocked as behind the curve, and coding is where the gap has been most visible and most commercially painful. Buying the team that builds the training environments is a shortcut straight to the bottleneck.
The structure is the story
This is the third time in roughly two years that Google has run this play. In July 2025 it paid $2.4 billion to license Windsurf's technology and hire its cofounders and research team; former Windsurf CEO Varun Mohan now runs Google's agentic coding program, Antigravity, while the remainder was sold to Cognition AI. Before that, a similar deal brought back Character.AI cofounder Noam Shazeer. Google is not alone: Meta took a large minority stake in Scale AI and hired its CEO in a roughly $14 billion deal in June 2025, and Nvidia announced a $20 billion licensing arrangement with chip startup Groq in December 2025 that also moved Groq's founder and staff.
The design is deliberate. Hire the people, license the technology non-exclusively, leave the shell standing, and you arguably never trigger a reportable merger. Regulators have noticed. FTC Chair Andrew Ferguson said in January 2026 that the agency was beginning to look very closely at how these things work, and on March 25 the FTC and Justice Department jointly sought public comment on whether premerger notification rules should capture acquihires and reverse acquihires at all.
Congress has been blunter. In a February 4 letter, Senators Elizabeth Warren, Ron Wyden and Richard Blumenthal wrote that such deals function as de facto mergers, allowing the companies to consolidate talent, information, and resources, all while apparently attempting to bypass the scrutiny typically applied to mergers and acquisitions. Their warning was economic as much as legal: If left unchecked, these types of arrangements between Big Tech companies and smaller AI developers will accelerate consolidation of the AI sector, which in turn risks driving up prices and choking off innovation.
The legal question is genuinely unsettled. Section 7 of the Clayton Act reaches acquisitions of assets whose effect may be to substantially lessen competition. If a founding team and its accumulated knowledge count as an asset — and a $1.5 billion price tag argues that they do — the corporate shell left behind may not be much of a defense. So far it has been an effective one: none of these deals has been blocked.
There is also the question of what happens to the shell. Scale AI cut 200 full-time employees and ended work with 500 contractors within weeks of the Meta deal, prompting one investor to call the company a gutted fish. Windsurf found a buyer. Mechanize, under Assadi, keeps its technology and whatever staff stayed behind — a real business, in theory, now competing against a customer that employs its founder.
Watch three things. Whether the FTC's rulemaking produces actual reporting thresholds for acquihires, which would make this structure far harder to run. Whether Google's next model shows the coding improvement $1.5 billion is supposed to buy — the justification is entirely prospective right now. And whether Mechanize under Assadi can still sell RL environments to OpenAI and Anthropic after its founder walked into Mountain View. The corporate entity survived. Whether that was ever the point is the open question.
“These deals function as de facto mergers, allowing the companies to consolidate talent, information, and resources, all while apparently attempting to bypass the scrutiny typically applied to mergers and acquisitions.”— Warren, Wyden and Blumenthal, U.S. Senators, February 2026 letter to the FTC and DOJ