Reporting confirms Tamay Besiroglu and more than a dozen former Mechanize staff have joined Google DeepMind — with final deal terms undisclosed, the more consequential signal is what the transaction reveals about where frontier modelvalue now sits.
Per reporting from Jingletree — there is no Google announcement in hand — Tamay Besiroglu, co-founder and former chief executive of Mechanize, is now a research scientist at Google DeepMind. More than a dozen other former Mechanize employees have followed, with most reported to be working on midtraining: the stage between pretraining and post-training where a general model is shaped toward specific competences. In Mechanize’s case, that competence was coding.
Mechanize was founded in April 2025 by three former Epoch AI researchers — Matthew Barnett, Tamay Besiroglu, and Ege Erdil. Reporting earlier this year placed the company at roughly 35 people, with approximately $9.1 million raised at a valuation of around $500 million. Those figures are attributed to reporting and should be read as such. Google had been in talks for a deal reportedly worth more than $1.5 billion, covering both the technology and the talent — but final terms were not disclosed. The $1.5 billion figure describes the level at which talks were reported, not a confirmed price paid.
The structure — a technology licence plus a team hire, without an outright acquisition of the corporate entity — is the same instrument Google used for Character.AI in 2024 and for Windsurf in July 2025, a deal reported at $2.4 billion. Windsurf’s former chief executive Varun Mohan now leads Google’s agentic coding programme, Antigravity. Mechanize is not reported to have been shut down or wound up.
Google Licence-and-Hire Timeline
2024
Google licence-and-hire for Character.AI — first documented use of the instrument at scale.
April 2025
Mechanize founded by three former Epoch AI researchers: Matthew Barnett, Tamay Besiroglu, and Ege Erdil.
July 2025
Google licence-and-hire for Windsurf, reported at $2.4B. Former CEO Varun Mohan now leads Antigravity, Google’s agentic coding programme.
2026
Google completes licence-and-hire for Mechanize. Besiroglu joins DeepMind as research scientist. Final terms not disclosed.
The key insight: Midtraining — the intermediate stage where a general model is tuned toward specific competences — had no visible labour market and no public price before this deal. It now has both. The implication is that marginal modelquality is believed to come from a stage that is labour-bound rather than compute-bound, and that the people who know how to execute it are scarcer than the accelerators.
Both figures come from reporting rather than disclosure. Mechanize was reported to have raised about $9.1 million at a roughly $500 million valuation earlier this year; Google had been in talks for a deal worth more than $1.5 billion covering the technology and the team. Final terms were not disclosed, so the gap between these bars is indicative of how the capability was being priced in negotiation, not a settled purchase price.
The Structural Read
Three times in roughly two years — Character.AI in 2024, Windsurf in July 2025, Mechanize now — Google has used the same instrument: licence the technology, hire the people, leave the corporate shell behind. As a business-model pattern, what that structure accomplishes is specific. The capability moves between organisations without a change of control, and a change of control is the event that ordinarily triggers merger review notification. This is offered as structural analysis, not a legal conclusion.
The timing is worth noticing precisely because the instrument is now being scrutinised elsewhere. The Justice Department has opened an antitrust investigation into NVIDIA’s roughly $20 billion licensing arrangement with Groq — a transaction of the same structural shape at a larger scale. The licence-and-hire is simultaneously routine at the frontier and under examination by regulators. To be exact about what is and is not being claimed: no regulator is reported to be examining Google’s Mechanize deal, and nothing in the available reporting says the structure was chosen to avoid review. The observation is that a transaction form which moves frontier capability without the usual notification is now common enough to be repeated, and contested enough to be investigated.
Map of AI — Layer Economics
The Frontier Is Fragmenting Into Priced Layers
For two years the frontier was discussed as though it had two phases: pretraining (compute-bound, capital-hungry) and post-training alignment. Midtraining occupied no named slot in the public market — no disclosed salaries, no transaction comparables, no price. This deal supplies one. Set against the same week’s movements — OpenAI commoditising the agent harness by distributing it freely, the memory layer repricing as high-bandwidth memory supply tightens — the picture is of a stack fragmenting into named layers that each develop their own distinct economics. Each layer is now acquiring a price, a labour market, and a competitive dynamic of its own.
There is also a second-order observation embedded in the integration structure itself. Windsurf’s former chief executive arrived at Google and was given a programme to run — Antigravity, a named agentic coding initiative. Mechanize’s former chief executive has arrived as a research scientist. Same instrument, meaningfully different integration. One transaction looks like the purchase of a product organisation with a defined delivery roadmap; the other looks like the purchase of a research capability to be deployed inside an existing programme. The difference matters for how the acquired knowledge actually diffuses into the acquiring organisation.
Where Each Frontier Layer Now Stands
Pretraining
COMMODITISING
Compute-bound and capital-intensive. Scaling returns are contested; the bottleneck is moving up the stack.
Midtraining
NEWLY PRICED
Labour-bound. Shapes general models toward specific competences. Just received its first public transaction comparable.
Post-Training / Agent Harness
COMMODITISING
OpenAI distributing the harness freely this week. Defensive moat at this layer is compressing rapidly.
Three Implications
IMPLICATION 1 — THE LABOUR MARKET FOR MIDTRAINING IS NOW VISIBLE
Before this deal, midtraining expertise had no transaction comparables, no disclosed salary anchors, and no public price. It now has a reported negotiation ceiling — more than $1.5 billion covering technology and a team of roughly a dozen-plus people — even if the final terms remain undisclosed. That number, conditional as it is, will function as a reference point for every subsequent negotiation in the same space. Founders and researchers working on model shaping now have a market signal they did not have six months ago.
IMPLICATION 2 — THE LICENCE-AND-HIRE IS GOOGLE’S STANDARD INSTRUMENT, AND REGULATORS ARE WATCHING THE CATEGORY
Three executions in roughly two years is enough to call this a deliberate strategy rather than opportunism. The instrument transfers frontier capability without a change-of-control notification — the event that ordinarily initiates merger review. Whether that is a feature or an incidental consequence of the structure is not reported. What is reported is that the DOJ is now examining an NVIDIA licensing arrangement of the same structural shape at a larger scale. The form is being tested, even if Google’s specific deals are not the ones under examination.
IMPLICATION 3 — INTEGRATION INTENT DIFFERS ACROSS DEALS, AND THAT DIFFERENCE HAS STRATEGIC CONSEQUENCES
Varun Mohan arrived from Windsurf and was assigned a programme — Antigravity — with a defined product remit. Tamay Besiroglu arrives from Mechanize as a research scientist. Same legal instrument, different organisational outcome. One integration appears designed to acquire delivery capacity inside a product line; the other to deepen research capability inside a lab. For competitors watching Google’s talent absorption strategy, the integration title is as informative as the deal size.
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
This account is based on reporting rather than a Google announcement. Google had been in talks for a deal worth more than $1.5 billion covering Mechanize’s technology and talent, but final terms were not disclosed — there is no confirmed price, and any per-person or multiple-of-valuation figure here is conditional arithmetic on that unconfirmed number rather than a disclosed metric. This is a licence-and-hire arrangement, not an acquisition: Google licenses the technology and hires people. Figures for money raised, valuation and headcount are as reported and are not company-confirmed. No regulator is reported to be examining this transaction, and nothing reported indicates the structure was chosen to avoid review; the Justice Department investigation referenced here concerns NVIDIA’s Groq licensing arrangement, a separate transaction, and the comparison is this article’s analysis. Alphabet is publicly listed and Mechanize was private. This is business analysis, not investment advice, and no view is expressed on any security. Google appears elsewhere in this week’s coverage, including as the guarantor behind Fluidstack’s data-center lease obligations. No quotes were available and none are invented here.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
