Former Google DeepMind top researcher launches London “compute-efficient” AI company
Tech.eu John Reynolds
A former DeepMind lead is launching Metis Reasoning in London. It wants a cheaper, more verifiable path to AI reasoning than giant frontier models.
Based on reporting by Tech.eu, John Reynolds — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
A new London AI startup is betting that bigger isn’t better. Metis Reasoning, co-founded by former Google DeepMind research lead Thore Graepel, is launching with an argument that AI should reason more like a system checking its work than a machine that simply spits out fluent guesses.
Graepel left Google DeepMind this summer, and he’s bringing some heavyweight company with him. Metis’s founding team also includes Piers Ingram, a technology and biotech entrepreneur who has raised more than $350m, plus Michael Warren, who worked on robotics and autonomous systems at Boeing, and David Berman, who heads AI at Cambridge Consultants and advises the UK government and NATO on AI.
The company says it is building a “compute-efficient” alternative to the frontier model race led by Anthropic and OpenAI. Ingram says Metis has been “funded ... adequately to date,” while declining to discuss reports that it is initially raising tens of millions of dollars. He also says the startup’s architecture is “much much more” compute efficient than frontier models and that it is not chasing ever larger models or ever larger training budgets.
That is the core of the pitch: large language models can already do a lot, but Metis says they still struggle with reliable reasoning and leave too little audit trail for high-stakes work. The startup says that makes adoption harder in areas like robotics and engineering, where decisions need to be checked, not just admired for sounding smart.
Metis says its approach should be more verifiable, robust and safer, and that it has already produced encouraging results on frontier reasoning problems at a fraction of the cost of frontier models. The first public model is not on a timetable yet. And it is entering a London scene that already includes other startups with similar ambitions, from Ineffable Intelligence to Recursive Superintelligence.
My take — AI-written commentary, not fact-checked reporting
This is the right fight to pick. The industry has spent years acting as if every problem needs a larger model and a bigger cheque, which is a nice theory right up until someone needs an answer they can actually trust. London seems to be collecting the adults in the room, at last.
Read more about this at: Tech.eu