The $1.5 Billion Hire: Google Completes Its Mechanize Talent Deal

The deal that San Francisco’s AI gossip mill has been tracking since August is done. Public LinkedIn profiles, first spotted by Business Insider on September 11, confirm that Google has completed its talent deal with Mechanize Inc. — the tiny startup founded to “fully automate the economy” — at a reported value of more than $1.5 billion. Mechanize’s co-founder and former CEO Tamay Besiroglu now lists himself as a research scientist at Google DeepMind, and more than a dozen other former Mechanize employees have updated their profiles to show new roles at Google, most of them concentrated on the company’s midtraining efforts — the increasingly critical middle stage of modern model development that sits between pretraining and post-training.

Google and Besiroglu both declined to comment, and the final terms were not disclosed. But the scale is not in serious dispute: Business Insider reported in August that the two sides were negotiating a package worth over $1.5 billion combining talent acquisition with a non-exclusive technology license. What is remarkable is the ratio. Mechanize raised just $9.1 million at a $500 million valuation in April 2026 — 103 days after opening its doors to outside capital. A team of roughly 35 people has now exited at roughly three times its last marked valuation, without an acquisition ever taking place.

What Mechanize actually built

Mechanize was founded in April 2025 by three Epoch AI alumni — Tamay Besiroglu, Ege Erdil, and Matthew Barnett. Epoch, which Besiroglu co-founded in 2022, is the research institute known for its rigorous tracking of AI compute, data, and scaling trends; its charts are cited in frontier labs’ own papers. Mechanize took that measurement mindset and pointed it at a product: reinforcement learning environments for coding agents.

The company’s pitch was blunt from day one. “The startup’s goal is the full automation of all work,” TechCrunch reported at launch, quoting Besiroglu’s ambition of “the full automation of the economy.” The launch post was instantly controversial —Besiroglu was skewered online for appearing to celebrate the replacement of human workers — but the underlying business was less sci-fi than the rhetoric. Mechanize built simulated work environments where AI agents could be trained and evaluated on long-horizon software tasks: realistic codebases, real pull-request workflows, real failure modes. Frontier labs increasingly believe that RL environments — not raw internet text — are the bottleneck for the next generation of agentic coding models, because an agent learns from thousands of verifiable task repetitions rather than static examples.

The market seemed to agree with the thesis. Mechanize itself estimated that AI labs were earning roughly $5 billion a year from coding tokens against some $300 billion in annual US software-engineer wages — a two-orders-of-magnitude gap it intended to close. When word of Google’s interest leaked in August, Deedy Das called it “the biggest AI data deal” in a category that barely existed eighteen months earlier.

Why “talent deal” and not “acquisition”

The structure is the story. Like its 2025 absorption of the coding startup Windsurf’s team — whose former CEO Varun Mohan now runs Google’s agentic coding program, Antigravity — Google structured the Mechanize transaction to bring aboard people and license technology while leaving the corporate shell independent. Mechanize continues to exist: its former chief of staff, Guive Assadi, now describes himself as the company’s CEO on LinkedIn.

There are two reasons Big Tech favors this shape. The first is regulatory. Talent-and-license deals are typically structured to avoid the antitrust scrutiny that fully fledged acquisitions attract — a pressing concern for Google, which remains under the shadow of an active US Justice Department antitrust case and has watched the FTC and DOJ signal growing discomfort with serial AI acquir-hires. Buying a $500 million startup outright is a reportable merger; licensing its technology and hiring its staff is, from a competition-law standpoint, far murkier — critics say by design.

The second reason is what Google actually wanted. What Mechanize built is knowledge: environment design, reward shaping, task generation, and evaluation harnesses for coding agents. That expertise walks out the door with the researchers. The license ensures Google can use what was already built; the hires ensure Google can build more of it internally, integrated directly into Gemini’s training pipeline.

The midtraining angle

Where the hires land matters as much as who they are. Business Insider reports the former Mechanize staff are working mostly on midtraining — the phase of model development that has quietly become one of the most competitive differentiators in the industry.

Midtraining sits between pretraining (trillions of tokens of raw text) and post-training (instruction tuning and RLHF). It covers curated continued pretraining, synthetic data generation, curriculum design, and — crucially for Google — the construction of RL-ready environments and reasoning data. IBM researchers documented across more than 500 controlled experiments this year that mid-training measurably boosts reasoning capabilities. OpenAI maintains a dedicated mid-training team, and Google’s own alignment researchers have experimented with “synthetic document finetuning” during the phase. A team that spent a year building industrial-grade coding RL environments is almost perfectly complementary to that agenda.

The subtext is also competitive. Business Insider notes frankly that Google “has been struggling to release an AI model that’s widely viewed as competitive lately,” and that coding is a specific weakness — the same article points out that Mechanize’s technology “helps tech companies’ AI models improve at coding, a challenge Google has struggled with.” While OpenAI and Anthropic trade benchmark leadership and Claude’s coding share keeps climbing, Gemini has repeatedly stumbled on agentic coding evaluations. Google is now paying roughly $1.5 billion for the people who built some of the best coding-agent training infrastructure outside the frontier labs.

The pattern: acqui-hires at acquisition scale

The Mechanize deal is the loudest example yet of a 2026 pattern: AI talent transactions priced like acquisitions but structured like hiring. Microsoft’s earlier Character-style arrangements, Google’s Windsurf deal, and Meta’s infamous $100 million signing-bonus era offers all pushed in the same direction — when a handful of researchers can move a model’s benchmark trajectory, their employer’s P&L matters less than their flight risk.

For startups, the math is seductive and destabilizing at once. Mechanize’s investors turned a $9.1 million seed into an exit reportedly worth over 30 times that in under six months. But the deal also removes the company’s founders from the company, leaving a shell with a license obligation and a new CEO. Whether “Mechanize” remains a meaningful vendor or becomes a footnote — the answer will tell us whether the acquir-hire era has genuinely replaced mid-size AI M&A.

What is certain is the signal. Google just paid acquisition-scale money for a team whose explicit mission was automating software engineering — betting that RL environments, not more parameters, are where the next coding-model gap will be won. And it did so in the one structure regulators have so far struggled to touch.

Sources

• Business Insider — Google just completed its $1.5 billion-plus talent deal for AI startup Mechanize (Charles Rollet, Hugh Langley, Katie Roof, Sept 11, 2026)
• Aventure.vc / Techmeme — LinkedIn profiles confirm deal completion (Sept 11, 2026)
• TechCrunch — Famed AI researcher launches controversial startup to replace all human workers everywhere (April 19, 2025)
• Dealroom — Google eyes $1.5B for Mechanize, a startup that raised $9.1M seed 103 days [prior] (Aug 6, 2026)