The AI training-data unicorn snapped up a reinforcement-learning environment startup its own founder had quietly funded three months earlier, tightening its grip on the pipeline frontier labs use to teach agents how to work.
Teaching an AI model to write a sentence is one problem. Teaching it to actually complete a task inside Salesforce, Excel or Slack without breaking anything is another, and it’s the problem eating budgets at every frontier lab right now. Mercor, the AI training-data platform that Brendan Foody started at 19, said on July 9 that it has acquired Deeptune, a startup that builds simulated versions of enterprise software where AI agents can practice, fail and try again, according to The Information. Financial terms were not disclosed.
an insider round, then a buyout
The deal has an unusual backstory. Foody, now 23, personally wrote an angel check into Deeptune’s $43 million Series A in March 2026, a round led by Andreessen Horowitz partners Marco Mascorro and Martin Casado, according to Startup Fortune. Three months later his company bought the whole thing. Asked about the sequencing, Foody was unusually direct: “It was in a lot of ways the main motivation, actually,” he told Fortune. Whether Mercor’s board or its outside investors signed off on the stake before Foody wrote it isn’t clear from what’s been reported, Startup Fortune noted, and neither company has said. It’s the kind of related-party structure that tends to draw more scrutiny the bigger a company gets — and Mercor is getting very big, very fast.
full stack ambitions
Strategically, the logic is straightforward even if the optics aren’t. Deeptune, run by founder and CEO Tim Lupo, builds what the company calls “training gyms” — reinforcement-learning environments that recreate hundreds of enterprise applications so an agent can click through a spreadsheet or a support queue and get graded on whether it did the job right, work Lupo has compared to a flight simulator, according to SiliconANGLE. Mercor’s own network of more than five million domain experts already writes the tasks and scoring rubrics — the APEX benchmarks — that tell a model whether it succeeded. Put the two together, as Foody wrote in a blog post announcing the deal, and “the constraint has shifted to the environments themselves: the places where models practice the work and get measured on whether they did it well.” It’s a bet that owning both ends of the pipeline — the graders and the sandbox they grade inside — is worth more than owning either alone. It’s also not Mercor’s first bolt-on: the company acquired Sepal AI in February 2026, a pattern Crypto Briefing described as a deliberate roll-up of specialized training tools. Rivals Scale AI and Surge AI have made smaller moves of their own to shore up infrastructure ahead of funding talks, per Startup Fortune, though Mercor’s version — buying the team building the environments its product depends on — is more direct.
growth outrunning a breach
The numbers Foody is citing are eye-popping, if hard to independently verify beyond his own statements. Mercor’s annualized revenue run rate hit $2 billion in June, up from $1 billion four months earlier, a trajectory Foody called “the fastest growth trajectory ever” from $1 million to $2 billion ARR in 24 months. Bloomberg has reported Mercor is in early talks to raise new funding at roughly a $20 billion valuation, double the $10 billion it fetched in October 2025 when it closed a $350 million Series C, according to Startup Fortune and SiliconANGLE. All this comes months after hackers linked to Lapsus$ exploited a vulnerability in the open-source LiteLLM library and exfiltrated an estimated four terabytes of Mercor data — including, per the reporting, tax and banking records, passport scans, Social Security numbers, interview recordings and facial biometrics — prompting a class action filed in California in April. Mercor said in June that only a very limited subset of sensitive information was affected, with no evidence of fraudulent use. Foody’s read on the fallout: “Every frontier lab has expanded their relationship with us since the data breach.” Take that as either a remarkable vote of confidence from customers, or evidence that at Mercor’s scale, frontier labs simply have nowhere else to go.
what to watch
The real test isn’t the acquisition itself but whether Mercor’s board — or the investors weighing a $20 billion check — push back on a CEO who admits, on the record, that his angel investment was step one of a buyout plan. For now, momentum is doing the talking: a doubling of ARR in four months tends to make related-party deals easier to wave through, not harder. Watch whether the Bloomberg-reported funding round closes at that valuation, and whether other AI-training vendors respond with roll-ups of their own before regulators or LPs start asking harder questions about who profits when the CEO writes the first check.
Mercor Buys Andreessen Horowitz-Backed Startup Deeptune
