When Tamay Besiroglu announced Mechanize in April 2025, the reaction was brutal. The AI researcher, co-founder of the well-regarded Epoch AI research institute, declared his new startup’s mission in a single, unapologetic post on X: “the full automation of all work.” Critics called it tone-deaf. One former colleague at Epoch reportedly called the launch a “comms crisis.” The company’s own pitch — that it could capture a share of the $60 trillion humans are paid in wages worldwide every year — struck many as both morally uncomfortable and wildly premature.
Sixteen months later, Google is reportedly negotiating to pay more than $1.5 billion for Mechanize’s technology and engineering team.
What actually happened in between
Mechanize spent 2025 and early 2026 quietly narrowing its scope. Rather than chasing “full automation of all work” on day one, the company focused on a much more concrete target: building simulated work environments, benchmarks, and grading systems that train AI coding agents to handle realistic, multi-step software engineering tasks — the kind of messy, long-horizon work that today’s AI models still struggle with.
That focus paid off. In April 2026, Mechanize closed a $9.1 million seed round at a $500 million valuation, backed by a notable list of individual investors including former GitHub CEO Nat Friedman, Stripe co-founder Patrick Collison, and investor Daniel Gross. By industry standards, a $9.1 million seed at that valuation was already unusual — it placed in the top 1% of comparable seed rounds by size.
Then, roughly 103 days after that round closed, Business Insider reported that Google was in talks for a deal worth over $1.5 billion — not a straightforward acquisition, but a structured arrangement to license Mechanize’s technology stack and hire a number of its engineers, reportedly to strengthen Google’s model evaluation and coding-agent infrastructure.
Why a licensing deal, not an acquisition
The structure is deliberate, and it’s becoming a pattern. This is the third time in roughly two years that Google has used this playbook:
- Character.AI (2024): Google rehired co-founder Noam Shazeer and licensed the startup’s technology rather than buying the company outright.
- Windsurf (2025): Google paid roughly $2.4 billion to license Windsurf’s technology and install its founders — including new CEO Varun Mohan — inside its Antigravity coding platform.
- Mechanize (2026): The same shape of deal, now on the table for a company that was worth $500 million just over three months earlier.
Structuring deals this way lets Google acquire talent and technology without triggering the antitrust scrutiny a full-company acquisition would attract. Regulators have already signaled they intend to examine this pattern more closely, so it isn’t a permanent workaround — but for now, it’s a repeatable one.
The timing also lines up with turmoil inside Google’s own AI leadership. Chief Scientist Jeff Dean is stepping down after 27 years at the company to co-found a new research venture, part of a broader wave of senior departures to rivals including OpenAI and Anthropic. Bringing in Mechanize’s team appears to be, at least in part, an attempt to backfill that loss of coding and evaluation expertise.
The uncomfortable question Besiroglu still hasn’t fully answered
None of this resolves the objection that made Mechanize controversial in the first place. Besiroglu has argued that even if AI automation reduces wages for some workers, people don’t rely on wages alone — they also receive income through rents, dividends, and government transfers. Critics have pointed out the circularity in that argument: an economy that automates away most paid work also has to explain where the income funding those rents and dividends will come from, and for whom.
What’s changed is not the answer to that question, but the market’s verdict on Mechanize’s technology. Eighteen months ago, the company was mostly known for a provocative tweet. Today, one of the three companies capable of writing the check is reportedly prepared to pay nine figures more than the seed round valued it at, specifically for its coding-agent training infrastructure — a much narrower, more concrete achievement than “automating all human work.”
Whether that’s a sign the broader vision is on track, or evidence that Mechanize’s real value was always in the specific engineering problem it solved rather than the sweeping mission it launched with, is still an open question.
