Zenithon AI, a London and San Francisco frontier lab building world models for physics, raised a $10m round led by backers including Lunar, Seraphim and MMC.
Alex Higginbottom says Zenithon's first world model, launching next week, covers different fusion confinement types and matches simulations in milliseconds, a million times faster.
Higginbottom predicts engineers could test designs in a world model within about ten years, rather than building five practice Starships, extending the company's push into semiconductors and aerospace.
Zenithon AI raises $10m to make physics simulation a million times faster
Zenithon AI is building world models that replace slow physics simulation with near-instant results. Hardware fields like fusion, semiconductors and aerospace iterate slowly and expensively, so a model a million times faster could change how designs get tested. The $10m round funds compute, data and hiring, and the open challenge is tying simulation to real experimental data.
Takeaways
World models can match physics simulations a million times faster. A simulation that takes ten hours to several days on a high performance computer can be matched in milliseconds. That speed lets engineers test a million designs in a world model before building the final one.
Fusion's slow build, measure, learn loop motivated the company. Higginbottom spent three years commercialising fusion and worked with about 20 fusion companies. Physics research iterates slowly, so he left his PhD in AI for plasma physics after raising a pre-seed round.
First model spans fusion confinement types, then extreme physics. Zenithon releases its first world model the following week, the first one covering different fusion confinement types. Fusion companies already use it in daily workflows, and the aim extends to semiconductor manufacturing and aerospace.
Semiconductor manufacturing is an open opening for AI. AI design has seen work from NVIDIA and Synopsys, but manufacturing has seen far less. Lam Research's CTO said there are 100tn ways to make a wafer, and Higginbottom wants to keep Moore's Law going.
Investors backed a generational problem and a strong team. Lunar, Seraphim and MMC back the $10m round. Higginbottom argues today's tools cannot keep pace with fusion, semiconductors and aerospace, and the core team includes researchers from Meta, NVIDIA and Anthropic.
Funds go to compute, data and hiring. The money pays for cloud compute, data generation and hiring in London and San Francisco near customers. Scaling laws guide planning, though Higginbottom concedes an order of magnitude discrepancy with how LLMs scale.
Next hurdle is linking simulation to experimental data. The following step is closing the gap between simulation and real experimental data. Higginbottom predicts that within about ten years, maybe less, engineers could test designs in a world model rather than build five practice Starships.
Tap a name to go there · tap a line to see why they are linked