The AI Labs Replaced Their Researchers With Business Operators

Maher Saba, the Meta vice president handed the company's new Applied AI Engineering Group in March, put it plainly: building great models isn't just about researchers and compute. ^[1] The group he runs was built with a flat one-manager-to-fifty-engineers ratio — an org chart for shipping, not for discovery. ^[1] The latest to make the move is Alphabet. Demis Hassabis, who built DeepMind into the company's research engine, was moved to a chief-scientist role focused on long-term AGI work, while Koray Kavukcuoglu took operational control of Gemini — the product, the roadmap, the commercialization. Goldman Sachs read the change as a shift from a research-led organization to a scaled AI platform company. ^[2] At OpenAI, the pivot ran through the exits. Three senior leaders — chief product officer Kevin Weil, Sora lead Bill Peebles, and B2B CTO Srinivas Narayanan — resigned in April, as the company killed its million-dollar-a-day Sora video tool, shut down its OpenAI for Science division, and decentralized its research teams, all under an executive mandate to unify business and product strategy ahead of a possible IPO. ^[3] By August the dedicated ethicist role was gone too.

AI ethics does not belong to a single individual or team at the company. — OpenAI

At Anthropic, even the research-adjacent work got a revenue label. The company's Claude Science workbench, a tool for researchers, was described internally as part of a strategy to diversify revenue streams ahead of its IPO — the company reached a $965 billion valuation and filed first. ^[4][5] The leadership changes would be cosmetic if the daily work hadn't moved with them. It has. Inside Google, compute allocation now explicitly prioritizes high-revenue projects and Gemini development over experimental or academic research — a competitive internal market where researchers must align their work with corporate priorities to get TPUs. ^[6] Researchers have left for startups to get compute without asking ten layers of management for permission. ^[6] Running alongside all of this is a market that stopped rewarding the smartest model. Alibaba's Qwen open-weight models hit 3 billion downloads in six months, dwarfing Alphabet's 418 million and Meta's 227 million. ^[7] Microsoft's Satya Nadella, whose company owns 27% of OpenAI, declared the market "commoditized" in June. ^[8] Then came the price war: OpenAI cut its Luna model's fees by 80%, Anthropic released a model at half the price of its top system, and the cuts were tied to labs seeking profitability ahead of public offerings. ^[9] None of this means the research mission is dead. Hassabis still talks about AGI, proposing a U.S.-led AI watchdog and arguing for using the window before AGI arrives to shape the technology. ^[10] The labs still ship new models; Anthropic's Claude now writes 80% of its own code. ^[11] But the rhetoric persists while operational control and resource allocation have already moved — Hassabis's AGI advocacy is decoupled from the decisions about who gets compute and what ships. The clearest sign of where the frontier moved is in how the labs now describe their own goal. OpenAI put its strategy in its own words.

Our strategy remains focused on advancing both capability and efficiency so each generation of intelligence can accomplish more work at a lower cost — OpenAI

Replit's AI head, watching the price cuts land, described the new model.

GPT-5.6 Luna is the closest we've come to intelligence too cheap to meter. — Michele Catasta

The competitive metric is no longer the most intelligent model; it's the cost per unit of work. The researchers who built these companies are leaving, and the people now in charge are optimizing for a different frontier.