Zenithon Raises $10 Million to Build AI World Models for Fusion Reactors and Rockets

Simulating how a fusion reactor or a rocket engine will behave under extreme physical conditions traditionally takes enormous computing time, one design at a time. Zenithon is building AI models that promise to explore a million design variations in roughly the time a conventional simulation takes to run just one.

Zenithon, a UK-based AI startup, raised $10 million in seed funding backed by BACKED, Lunar, Seraphim, MMC and SOSV, along with a group of individual founders and hyperscaler directors. The company is developing what it calls world models, AI systems trained to learn how complex physical systems behave under a wide range of conditions, with an initial focus on technologies including fusion reactors, rockets and semiconductor manufacturing.

Why Physics-Based World Models Are a Different Bet Than Language Models

Most AI world-model efforts to date have focused on video, robotics or gameplay environments, domains where large amounts of real-world or simulated visual data already exist to train on. Zenithon's focus on extreme physics, fusion, propulsion and semiconductor fabrication, targets domains where the underlying physical processes are both harder to simulate accurately and more expensive to test experimentally, meaning a model that genuinely learns the governing physics rather than merely pattern-matching on limited data could unlock design iteration speed that is currently constrained by raw compute cost.

Why Learning From Real Experiments Matters Here

The company says its models will both simulate design variations at scale and learn directly from real-world experimental data, a combination that matters specifically in physics domains where a pure simulation, no matter how fast, can still diverge from how a system behaves in the real world under extreme conditions. Grounding a world model in actual experimental results, rather than simulation alone, is a harder data problem to solve but produces a model more likely to be trusted by engineers making real design decisions.

Why a Seed-Stage Syndicate of Specialist Funds Makes Sense Here

Zenithon's backer list, spanning climate-and-space-focused Seraphim, deep-tech generalist MMC, and hardware-heavy SOSV, reads like a syndicate assembled specifically to cover the different physical domains the company is targeting rather than a single generalist fund writing the whole check. For a company pursuing fusion, propulsion and semiconductor applications simultaneously, spreading the round across several sector-specialist funds gives the company access to domain-specific diligence and introductions across all three verticals at once, rather than depending on one fund's partners to be credible across such a wide range of physical-science domains.

What This Means for Founders

For founders building AI applications in deep-tech, physical-sciences categories, Zenithon's seed round is a signal that investors are willing to back highly technical, pre-commercial world-model bets well outside the more crowded video and robotics corners of the category, provided the founding team can credibly claim relevant domain expertise in the underlying physics. Founders in this position should consider assembling a similarly specialist-heavy syndicate rather than seeking a single generalist lead, since the diligence bar for physics-grounded AI claims tends to be higher than for a typical software pitch.

Browse the deep tech investor directory for more active funds in physics-based and hard-science AI, read the baseline mechanics of a pre-seed round if the term is new, or explore active capital on AngelLinx @ angellinx.ai/register.

AngelLinx Intelligence | angellinx.ai