In early September, DeepMind, Google’s AI division, acquired a little-known startup called Mechanize, Inc., a San Francisco-based company that develops AI coding agents. The tech giant absorbed more than a dozen employees in the multibillion-dollar “acquihire” deal, as Business Insider reported at the time, including four key engineers.
Now, less than two months later, Gazetteer SF has learned that three of those engineers have defected from DeepMind to launch their own AI-training startup, claiming they turned down an enormous retention offer on their way out the door.
Software engineers Max Niederman, Stephen Yang, and Byron Luk are currently meeting with investors to raise a seed funding round for a new company, Data-Limited Regime, potential investors told Gazetteer. The founders were originally seeking $3.5 million from angel investors at a $350 million valuation but were recently approached by several venture capital firms who offered to invest significantly more, the potential investors said.
Niederman will lead the startup as its chief executive, with Yang as CTO and Luk as president and chief researcher. Niederman and Luk are in their early 20s. Yang graduated from high school last year.
According to the timeline shared with investors, Data-Limited Regime is still in its early research and development stages but has possibly begun a short-term pilot with an unspecified AI lab.
In their pitch, the trio emphasizes that when they left DeepMind, they were offered more than $300 million in stock options to stay. “We declined,” the pitch sheet, viewed by Gazetteer, states.
Google declined to comment on the retention offer. One potential investor who received the pitch sheet called $300 million in stock options “absurd” for a team of engineers in non-managerial roles, including two who are barely of legal drinking age. The team behind Data-Limited Regime did not respond to multiple requests to verify the large stock option package or comment on why they allegedly turned it down.
Still, their departure certainly doesn’t look great for DeepMind or Google.
Like Mechanize, Data-Limited Regime is what’s known as a “data vendor,” or a company that collects and evaluates data that AI labs like DeepMind then use to train their models. (In DeepMind’s case, Gemini.) This data is extremely valuable to AI companies locked in cutthroat competition to build and release the best chat and coding bots; Google spent more than $2 billion in the Mechanize deal — a previously-unreported figure that is included in Data-Limited Regime’s pitch sheet — as part of an effort to improve Gemini’s coding capabilities amid what The Information called “a struggle to catch up with Anthropic.”
Since late last year, coding has emerged as one of the AI industry’s most promising commercial opportunities. Claude Code, Anthropic’s agentic coding tool, is widely considered the best on the market, with competition coming from OpenAI’s Codex as well as from Cursor, another AI coding startup that was acquired by SpaceX for $60 billion in August.
During their short tenure at DeepMind, Niederman, Yang, and Luk evaluated the datasets Google was purchasing from other data vendors to train its models — a detail they note in the pitch sheet before asking for money to fund what is essentially a competitor. (The Mechanize talent deal was reportedly a non-exclusive licensing agreement, meaning Niederman, Yang, and Luk could plausibly avoid legal trouble while they launch their own data startup.)
While Google has made progress in agentic coding with the recent launch of Gemini 3.8 Flash, corporate politics inside the tech giant have been known to slow development, especially compared to its competitors.
In August, four of Google’s highest-profile AI scientists left to start their own AI company; one of them, Oriol Vinyals, the former vice president of research at DeepMind and Gemini’s technical lead, told WIRED, “In a large organization there is always a lot of inertia you have to overcome to make any radical changes.”
Do you work at Google DeepMind? Message me securely on Signal at cyd.01.
