Zenithon, a London startup that announced a $10 million seed round on Sept. 30, launched NOVA on Oct. 20-21 and describes it as "the first world model for extreme physics," capable of predicting system behavior in seconds on a single GPU. The company says that could replace fusion simulations that now take hours, but no independent results or public access have emerged to back the claim.
The pitch, in short, is speed.
What NOVA is supposed to do
According to Zenithon's LinkedIn announcement, NOVA "predicts how complex physical systems evolve and respond to control inputs, in seconds on a single GPU." The company says it is built on a single transformer architecture, the same family of neural network behind modern language models, and was pretrained across multiple physical systems and simulation codes.
Training data covers plasma physics, fluid dynamics and other systems at varying spatial and temporal resolutions, per the announcement. Predictions are action-conditioned, which means the model forecasts how a system will evolve once a given control input is applied.
Zenithon also says NOVA can adapt to new operating regimes from only a few examples. In one tokamak case, a doughnut-shaped fusion device, it reportedly needed none, a capability known as zero-shot prediction.
Fusion comes first
Fusion is the opening showcase. The company says it has demonstrated NOVA across magnetic, magneto-inertial and inertial fusion, which it presents as spanning the main approaches to confining and igniting plasma.
Its website names two further target industries: semiconductors, for reactor and chemistry simulation, and aerospace, for de-risking mission-critical hardware. The tagline is "World models for extreme physics."
CEO Alex Higginbottom laid out the ambition in a SemiWiki interview published Oct. 9 and updated Oct. 20. "We're building models to explore a million design points in the time a conventional solver takes to run one," he said. The interview predates the launch and did not concern the NOVA announcement itself.
What is not public
Access is the sticking point. IsItSOTA, a model tracker that has followed NOVA since Oct. 20, lists "API: not yet" and "Open weights: unknown," and it reports no independent results. Availability has not been stated, and no pricing has been published.
No customers are named, either. A LinkedIn repost quoting the CEO alludes to "leading fusion companies," but as of Oct. 29 the research turned up no named users.
The route to market, for now, is a "Book a demo" button on the website. Higginbottom said in the SemiWki interview that Zenithon pursues pilots, custom deals and fine-tuning to customers' own systems. A job listing for a forward-deployed engineer, based in "London or San Francisco," hints at a push into the US. LinkedIn puts the company at 11 to 50 employees.
The supporting research
On Oct. 16, Zenithon authors posted an arXiv preprint titled "ModeCast: One-Shot Prediction of Linear MHD Eigenmodes and Growth Rates in Tokamaks." The paper claims inference 10,000 times faster than JOREK linear runs on a single GPU, covering n=1-4 modes. It was accepted to the NeurIPS 2026 Sim2Science workshop.
Still, ModeCast is a separate paper and not a NOVA benchmark. Earlier signals are thinner. A 1cFE blog post cited a Z-pinch surrogate that ran in under 8 milliseconds versus roughly five hours for a conventional approach, drawn from a December 2025 workshop result, but noted that hardware and validation details were missing.
Founders and investors
Higginbottom is founder and CEO. Co-founder Abetharan "Abby" Antony serves as chief technology and product officer and holds a PhD in plasma and high-temperature physics from Imperial College, according to BaeHQ and LinkedIn profiles.
The $10 million seed, reported by Tech.eu, Sifted, SOSV and Tech Funding News, drew BACKED VC, Lunar Ventures, Seraphim Space, MMC Ventures and SOSV. Caplight lists the round at $10.13 million. Seraphim published the NOVA announcement post on Oct. 21.
A crowded field
Physics AI is hardly empty. NVIDIA expanded its PhysicsNeMo libraries for surrogate modeling across computational fluid dynamics, weather and nuclear engineering, according to an August 2026 newsroom post. UK-based PhysicsX sells what it calls "Large Physics Models" and announced a CoreWeave partnership in March 2026. Neural Concept, an EPFL spin-out, raised a Series C in December 2025, per DClaw Research.
Zenithon is wagering on one pretrained model for extreme regimes rather than a separate surrogate for each product. Whether engineers can put that claim to the test depends on when an API or an independent benchmark appears. As of Oct. 29, neither was public.
