Pomegra
Zenithon, a London AI lab, has raised a $10M seed round to train world models for fusion reactors, rockets and chip manufacturing, with Backed VC, Lunar Ventures, Seraphim Space, MMC Ventures and SOSV backing it.
Key Takeaways
- Zenithon raised a $10M seed round announced on September 30, 2026, with valuation undisclosed.
- Backed VC, Lunar, Seraphim, MMC and SOSV are named backers, alongside angels.
- Funds split evenly between hiring and compute, targeting a new model every three months.
Lead
Zenithon, a London-based AI start-up founded in 2025, announced a $10M seed round on September 30, 2026, to build what it calls the first world models for extreme physics. The company targets three industries where a single simulation or experiment can take days: fusion energy, rocketry and semiconductor manufacturing. Backed VC, Lunar Ventures, Seraphim Space, MMC Ventures and SOSV are named as backers, with several angel investors also taking part. Zenithon has not disclosed a valuation.
What Does Zenithon Actually Build?
Zenithon builds world models, AI systems that learn how complex physical systems behave and predict the outcomes of different designs. The company says its models can explore a million design points in the time it takes a conventional solver to simulate one.
The models train on two inputs: simulation data and results from real-world experiments. That second input matters, because pure simulation inherits the approximations of the solver that produced it.
Co-founder Alex Higginbottom, who left a PhD programme in fusion energy to start the company, put the plan plainly: "We're using $10 million to train large world models for extreme physics." His co-founder, Abetharan Antony, is a plasma physicist with tokamak research experience.
Who Led the Round, and What Does It Say?
The round has no single named lead. The investors are listed together, and the team and backers' roles within the round are not broken out. That is common for early seed rounds, but it leaves outsiders unable to tell who set the terms or how much each fund committed.
The investor mix is more informative. Seraphim Space brings a space-sector focus, which maps to the rocket use case. MMC Ventures and Backed are generalist European early-stage funds, and SOSV's backing and its deep-tech portfolio fit a hardware-adjacent thesis. Founders and hyperscaler directors also put money in.
The company has also assembled an advisory group. It includes Charlie Songhurst, a Meta board director and former head of strategy at Microsoft, and Dan Brunner, co-founder of Commonwealth Fusion Systems and a former MIT researcher. Zongyi Li, an assistant professor at NYU, completes the group.
Why Do Extreme-Physics Simulations Need a New Approach?
Existing simulation tools are slow relative to the design questions engineers ask. Zenithon states that "when a single simulation or experiment takes days, engineering teams can explore only a fraction of possible designs."
Fusion plasmas, rocket combustion and the processes inside a chip fab involve extreme temperatures, pressures and multi-scale interactions. Each one is expensive to model and costly to test physically. A model that approximates the results quickly could let an engineering team screen thousands of candidates before running a single high-fidelity simulation or hardware test.
The caveat is accuracy. A fast surrogate is only useful if its errors are bounded in regimes where little experimental data exists, and fusion has comparatively few operating machines to learn from. Zenithon's claim of a million-fold design exploration speedup is a company statement, and no independent benchmarks have been published.
How Does Zenithon Compare With Rivals?
The closest comparison is PhysicsX, also London-based, which has raised more than $400M at a reported $2.4B valuation and applies machine learning to engineering simulation across industries. Fusionality, based in Lausanne, works on plasma control software, a narrower fusion-specific niche.
Zenithon's $10M is small against that benchmark. Its pitch is concentration: a single research focus on extreme physics rather than a broad industrial simulation portfolio. Early customers are in the fusion sector, though the company has not named them or disclosed contract values.
What Will the Money Pay For?
Zenithon will split the capital evenly between headcount and compute. The aim is to release new models every three months, a cadence that depends heavily on training budgets.
The company has 11 full-time staff and plans to reach 17 within three to six months. It will expand in San Francisco and across the US alongside its London base. Prior funding has not been disclosed, so the round's step-up from any earlier capital cannot be assessed.
What Comes Next?
The first test is whether quarterly model releases show measurable gains against established solvers on public fusion and aerospace benchmarks. A second is whether named fusion developers move from early adoption to paid, repeat use. A third is whether rocket and semiconductor customers, whose data is often tightly guarded, will share enough of it to train models.
A $10M seed gives Zenithon roughly the runway to answer those questions, but compute-heavy model training can consume a round quickly. A larger raise would likely depend on results shown in the next two to three model releases.
Outlook
Zenithon enters a crowded field of physics-focused AI companies with a narrow target, a modest seed and a team of 11. Its backers' fit with the fusion and space sectors is clear, while the valuation, the lead investor and prior capital remain undisclosed. Verified performance against conventional simulation will decide whether the company's speed claims hold.
