CNBC took stock on April 28 of AI researchers who have left Meta, Google and OpenAI over the past year. The list makes one thing clear: the core AI teams inside Silicon Valley's largest platforms are losing important people.
Dealroom's numbers show the money chasing that talent. So far in 2026, venture firms have invested $18.8 billion in AI startups founded after 2025. At the current pace, that total would exceed the $27.9 billion raised by all newly formed companies in 2024.
Who Has Left
The most research-heavy examples are already raising rounds that once looked extraordinary for companies only months old:
- David Silver, formerly of Google DeepMind, started Ineffable Intelligence with a $1.1 billion seed round and a $5.1 billion valuation, betting on reinforcement learning without human data.
- Tim Rocktäschel, a former DeepMind principal research scientist, is building Recursive Superintelligence with at least $500 million and potentially $1 billion, at a reported $4 billion pre-money valuation, around self-improving AI systems.
- Yann LeCun, formerly Meta's chief AI scientist, raised $1 billion in March for AMI Labs.
- Anna Goldie and Azalia Mirhoseini, previously at Anthropic and DeepMind, raised $335 million across two rounds for Ricursive Intelligence.
- Humans&, founded by former Anthropic and xAI co-founders, raised $480 million in January.
- Periodic Labs, started by former OpenAI and DeepMind employees, raised $300 million last September.
Each of those checks is unusually large for a young startup. A nine-figure seed round was a spectacle in 2024. It is now close to a going rate for a small number of elite AI spinouts.
Why VCs Are Willing To Pay
Eurazeo managing director Elise Stern captured the logic behind the funding rush: "When you're in a race, you narrow focus. That creates a vacuum."
Large labs are locked into benchmark battles and scale races. That focus pulls research toward a small number of main tracks and leaves adjacent directions for startups to claim.
The neglected areas include reinforcement learning, where the dominant labs still emphasize data scale and post-training alignment; AI interpretability, where the commercial payoff is harder to budget for; vertical systems in pharmaceuticals, materials and robotics; and new architectures outside Transformer or autoregressive approaches.
Ineffable Intelligence is attacking the first category. Periodic Labs sits in the third. AMI Labs is pitching the fourth. In each case, the idea may be hard to protect as an internal Big Tech project but easier to explain as a stand-alone company.
Why Big Tech Cannot Keep Them
The problem is not compensation. Meta and Google already give top AI researchers packages worth more than $10 million, and OpenAI's tender offers for important employees can reach the tens of millions.
What is harder to retain is research freedom. AI work inside the largest companies increasingly resembles product development: projects must support the next flagship model, papers can be constrained by competitive intelligence concerns, and senior executives have tighter control over direction.
Venture investors have also validated the alternative path. LeCun left Meta and raised $1 billion within months. Rocktäschel left DeepMind and reached a $4 billion valuation within months. Once that pattern circulates among researchers, the startup option becomes difficult to ignore.
How Long Can The Migration Last
In the near term, the trend looks durable. Big Tech earnings still depend on visible compute spending and benchmark leadership, so research portfolios are unlikely to loosen quickly. Venture funds also still have capital to deploy.
The medium-term test has two parts. First, the initial wave of companies must show real results within about 18 months. If $18.8 billion produces few commercial products or credible research breakthroughs, later investors may hesitate.
Second, large companies could respond by giving core teams more research autonomy, spinning internal startup teams into independent entities, or reviving semi-independent structures like the original DeepMind model.
For now, that would require Big Tech to admit that its retention model is failing. Its latest earnings-call posture suggests it is not ready to say that. Over the next year, Recursive Superintelligence, Ineffable Intelligence and AMI Labs will be the three companies to watch. Their first results will decide whether this is a bubble or a new division of labor in AI research.
Sources: Meta, Google, OpenAI among Big Tech firms seeing top staff leaving to launch AI startups (CNBC, 2026-04-28), CocoLoop, Big Tech's Best Are Quitting Their Jobs to Form Startups (Autogpt.net, 2026-04-28)