Forty Founders and a Breakfast: DeepMind's Alumni Wave Bets Big Money on Post-Transformer AI

The most consequential competition facing Google DeepMind may not come from OpenAI or Anthropic. It may come from its own departed scientists — and they are raising money at a pace the industry has never seen for ideas that explicitly reject the architecture everyone else is scaling.

Bloomberg reported on September 26 that Nando de Freitas, who spent roughly a decade at DeepMind before a stint at Microsoft, is in talks to raise at least $100 million to launch Revolution Labs, a startup built around diffusion models as an alternative to the transformer architecture that underpins every major large language model today. Accel is in talks to lead the round, according to the report.

The story behind the story

De Freitas’s round does not stand alone. It is the newest entry in what has become a systematic talent migration:

• David Silver, the reinforcement-learning legend behind AlphaGo, has raised roughly $1 billion for his London startup Ineffable Intelligence — a record-shattering $1.1 billion seed at a $5.1 billion valuation announced in April, with Sequoia and Nvidia backing.
• Jack Parker-Holder’s Emulate, a DeepMind offshoot building world models for robots, is in discussions to raise a $700 million seed round, as Bloomberg reported earlier this month.
• Thore Graepel, a veteran DeepMind scientist who left this summer, is seeking substantial funding for Metis Reasoning, a reasoning-model startup.
• Jeff Dean and Sanjay Ghemawat, Google’s most senior systems engineers, co-founded Discovery Loop, a research venture valued in the tens of billions.
• Andrew Dai’s Elorian, which debuted in April, is pursuing visual reasoning that its founders argue transformers cannot efficiently deliver.

The detail that crystallizes the trend: fifteen current and former DeepMind employees recently met for a breakfast in central London — effectively a support group for founders-to-be comparing notes on raising capital. Bloomberg’s counting now puts Google and DeepMind as the origin of more than 40 “neolab” founders, more than OpenAI, Anthropic, and Meta combined.

Why “post-transformer” is the pitch

The technical premise deserves attention beyond the funding headline. Diffusion models — the family behind image generators like Midjourney and Stable Diffusion — generate output by iteratively denoising a random seed rather than predicting one token at a time. Applied to language, this offers a coarse-to-fine approach where the model refines an entire answer in parallel instead of writing left to right.

The commercial proof of concept already exists. Inception Labs’ Mercury line of diffusion LLMs runs at over 1,000 tokens per second — several times faster than comparable autoregressive models — with claims of less than half the inference cost. Mercury 2.5, shipped this month, pushed intelligence benchmarks up 40% over its predecessor while keeping that speed edge. Parallel token generation also changes the economics of agentic workloads, where chains of model calls multiply latency and cost.

Revolution Labs’s wager, as framed by Bloomberg, is that diffusion can scale to frontier-level intelligence — and that a transformer incumbent’s entire inference stack, with its sequential bottleneck, becomes a strategic liability at agent scale. Google itself has hedged: its own Gemini Diffusion experiments showed the family working at the tech giant’s scale. Accel putting growth-stage money behind the thesis would be the first hard signal that post-transformer architectures now attract more than seed checks.

The exit motive, in their own words

Janusz Marecki, a former DeepMind lead scientist now an AI partner at Ahren Innovation Capital, put the departure rationale bluntly: “There’s a top-down pressure to drop everything you’re working on and focus on large language models. Then you just leave, there’s no point. These are smart people, smart scientists.”

He also flagged the market distortion: “Valuations used to be 10 times revenue. Now it’s 200 times, 300 times.”

Demis Hassabis — who in August handed day-to-day control of the lab to technology chief Koray Kavukcuoglu and moved to Alphabet’s chief scientist role — pushed back on the exodus framing: “There’s a lot of talent movement between all the leading labs and we win our fair share of the top talent. We have by far the biggest and broadest research bench of any of the labs out there.”

The numbers partially support him: DeepMind’s arrivals-to-departures ratio has fallen from roughly 12:1 in Q2 2023 to about 2:1 in Q3 2026, per Bloomberg — attrition elevated, but hardly a collapse, even as Anthropic and OpenAI poach aggressively.

Why this matters

Three stakes stand out.

First, architectural diversity. Every frontier lab except a handful of challengers now runs on transformers. If scaling laws plateau — or if inference cost becomes the binding constraint on agent deployment — the industry’s monoculture becomes a systemic risk. DeepMind alumni are uniquely positioned to test alternatives, having spent years on world models, RL, and non-transformer research that got deprioritized as Google consolidated around Gemini.

Second, Google’s retention math. Google is a founding investor in several departing-founder startups, which functions either as a durable hedge on disruption or as a subsidy to its own future competitors. The breakfast-club dynamic suggests the pipeline of departures is self-reinforcing: each record raise lowers the social cost of the next exit.

Third, capital discipline. A $100 million round for a pre-product lab testing an unproven-at-scale architecture is either prescient contrarianism or the top of the 200x-revenue froth Marecki describes. The honest answer is that nobody knows yet — and the Revolution Labs close, when it happens, will be watched as a referendum on both diffusion specifically and the neolab model generally.

What is certain is that the most expensive question in AI — what comes after the transformer? — now has roughly forty well-funded teams of ex-DeepMind scientists racing to answer it, and at least a billion dollars of Silver’s Ineffable already deployed. Google built the deepest research bench in the industry. Its alumni are betting they can build the next one without the LLM constraint.