Zenithon raises $10M for AI 'world models' that simulate extreme physics

Zenithon, a London AI lab, has raised a $10M — about £7.4M — seed round to point one of AI’s hottest ideas at a very hard problem: simulating extreme physics. It is building what it calls world models for extreme physics, machine-learning systems meant to stand in for the painstaking engineering simulation that underpins fusion reactors, rockets and semiconductor fabrication. The pitch is startling in its ambition: models that can, the company says, explore a million potential design points in the time conventional methods take to simulate one, while learning directly from real-world experiments.

It is worth being precise about what “world model” means here, because it is not the same thing as the term we used last month for Emulate, the DeepMind spinout. Emulate’s world models are about the everyday physical world — how objects move and scenes unfold. Zenithon’s are about the brutal, narrow physics of engineered systems: plasma in a fusion chamber, combustion in a rocket, the fabrication steps that etch a chip. These are domains where a single high-fidelity simulation can take days of supercomputer time, so the design loop is agonisingly slow. If a model can approximate that physics well enough to search thousands of designs quickly and then check the promising ones against real experiments, it compresses that loop dramatically — which is exactly where the value would be.

The people around it suggest this is a serious attempt rather than a buzzword play. Zenithon was founded in 2025 by chief executive Alex Higginbottom, who left a machine-learning PhD to start it, and co-founder Abetharan Antony, who holds a PhD in plasma physics. The $10M came from Backed, Lunar, Seraphim, MMC and SOSV, and the advisory bench is the tell: Dan Brunner, a co-founder of Commonwealth Fusion Systems, Charlie Songhurst, a Meta board director and former head of strategy at Microsoft, and the NYU scientific-ML researcher Zongyi Li. That is a lot of domain credibility for a seed-stage company.

It also adds to a pattern that is becoming one of the more interesting things about UK AI. Alongside Emulate’s world models and Basecamp Research’s models of biology, Zenithon is part of a small cluster of British labs aiming frontier AI not at chatbots but at science and hard engineering — the strategically important, capital-intensive problems that the country’s industrial-policy instincts keep gravitating towards. For a seed round, the ambition-to-evidence ratio is high: a million-to-one speed-up and learning from live experiments are bold claims from a pre-product company, and extreme-physics simulation has humbled plenty of people before. The proof will be whether real fusion, aerospace and chip teams start designing with it. But the problem is enormous, the team is credible, and it is a genuinely exciting corner of the UK scene to watch.