AI FOR SCIENCE · INTERVIEWS
Higginbottom says Zenithon builds world models for extreme physics: the first value is speed, with fusion simulations that take 10 hours to several days returned in milliseconds with the same results, and the second step is training on experimental data to close the gap between simulation and reality.
The record has the round and one line about evaluating a million designs in the time of one simulation. The interview adds the baseline (10 hours to several days), a first model due 'next week' with new versions every month or couple of months after, and semiconductor manufacturing as the second use.
No result is shown: 'the same results' comes with no error or benchmark, the model is not yet released, no customer or academic partner is named, and closing the gap to reality is, by his own account, an objective not yet met.
His account: LLMs have changed knowledge work and vision world models are starting to matter in some sectors, but physics has not had this yet, and that is what Zenithon builds. The company began from fusion but from day one meant to be broader, aiming at world models for extreme physics.
“Zenithon is a frontier AI lab building world models for physics.”Alex HigginbottomHe spent three years commercialising fusion, working with about 20 fusion companies, concluded the core build, measure, learn loop was too slow, and began a PhD in AI for plasma physics, leaving halfway to found the company. His arithmetic: a fusion reactor can be put together in trillions of ways and one simulation takes 10 hours to several days on a high-performance computer, so you can only pick up random grains of sand on a beach; their model gives the same results in milliseconds, so a million times faster.
“it's like picking up random grains of sand on the beach.”Alex HigginbottomHe says nobody knows which fusion route wins, so the model must cover tokamaks, laser fusion and every other confinement type; the first model comes 'next week' and for the first time puts several confinement types in one model. New versions follow every month or every couple of months, as at frontier labs. The same approach is meant for etching and plasma in semiconductor manufacturing: AI for chip design has a lot of work, he says, but manufacturing very little.
Speed is today's value proposition; the second value is training directly on experimental data to close the sim-to-real gap, when the model would be not only faster but closer to reality. The larger vision: within ten years or sooner, no need to build five practice starships, do a million in a world model and build the final one straight away. He says that may never be reached, but it is the direction.
“we won't need to build five practice starships. We can do a million in a world model and then build the final one straight away.”Alex HigginbottomHe says most of the money goes straight into cloud compute, partly to generate data (they own the data stack) for models on the scale of billions of parameters, with the rest for hiring in London and San Francisco. Over the summer they worked out scaling laws on both the data and model sides, say they trust the curves, blind-tested them and use them for planning, though with an order-of-magnitude discrepancy against LLMs.
The timeline entry of 2026-09-30 records the company's own line, evaluating a million designs in the time of one simulation; neither article shows a result.
Higginbottom gives the baseline: one simulation takes 10 hours to several days on a high-performance computer and the model returns the same results in milliseconds; no error, benchmark or hardware is given.
The funding record's use of proceeds says a new model every three months.
In the interview the first model arrives 'next week', with a new version every month or every couple of months, and one model carrying several confinement types.
This site's record holds no Zenithon customer or academic partner, and the funding record's partnerships are empty.
He says several leading fusion companies already use it in daily workflows and that the names and academic partnerships will be announced 'shortly'; none is named now.
If one evaluation drops from hours to milliseconds, a fusion design can be searched widely across trillions of combinations instead of a handful of simulations. But the model stands in for the simulator, and how accurate it is decides whether the best design found can be trusted; that is where the interview gives no figure.
He sets chip manufacturing beside design: design has Nvidia, Synopsys and some startups, while manufacturing (etching, plasma) has little, though it faces giants such as TSMC, Applied Materials and Lam Research. It says Zenithon's second customer group lies outside fusion; whether any of those companies use it the interview does not say.
One fusion simulation takes 10 hours to several days on a high-performance computer, and Zenithon's model gives the same results in milliseconds, a million times faster.
Company-reported
No error, benchmark or hardware is given; 'the same results' is not quantified.
Several leading fusion companies already use it in daily workflows, with names and academic partnerships to be announced shortly.
Company-reported
Nobody is named and no usage data is given.
The first world model is released next week, the first with several confinement types in one model, followed by new versions every month or couple of months.
Target
A plan; the release had not happened at the interview, and 'first' is the company's own word.
Models are on the scale of billions of parameters; scaling laws on both data and model sides were blind-tested and are used for planning, with about an order-of-magnitude discrepancy against LLMs.
Company-reported
No curve or figure is given.
The second value is training directly on experimental data to close the sim-to-real gap and bring the model closer to reality.
Not yet closed
He presents it as a large objective and does not say how far it has gone.
Within ten years or sooner, a million starships can be done in a world model and the final one built straight away.
Guest's judgment
He says himself that this north star may never be reached.
The team includes researchers from Meta, Nvidia and Anthropic, and the chief scientist invented many of the biggest breakthroughs in AI for physics in the last five years.
Company-reported
The company's own account, with no specific work named.
About 80 fusion companies started in the last 5 to 10 years; a Lam Research CTO said in a speech a few years ago that there are 100 trillion ways to make a wafer.
Industry context
Industry context; the figures are his relay, and the original speech was not read.
Whether the first world model is released in the week after the interview, stating which confinement types it covers.
Would show up as A dated release by Zenithon naming the model and what it covers.Checked on Said: 'next week' (said on 2026-09-30)Target date passed Higginbottom said on 2026-09-30 that the first world model, with several confinement types in one model, would come 'next week'; by 11 October no dated release was found in search, the round's coverage or the investors' pages, and the last line the record holds is still the round coverage's new models every quarter. Zenithon's own site returned 403 to both fetch paths and was not read.Whether, after the first model's release, error and speedup are published against a named simulator on the same cases.
Would show up as A release or paper naming the simulator, the cases, the hardware, the error and the runtime.Whether a named fusion company or academic institution says it uses the model.
Would show up as A release by the customer or institution stating which workflow it is used in.Whether, after training on experimental data, the model's error against a device's measurements is lower than that of the simulation it replaces.
Would show up as A side-by-side of model and simulator on the same device and the same measured data.Release cadence: whether a dated next version follows the first model within two months.
Would show up as The company's release record stating version, date and changes.Later evidence that tests or limits what was said, and targets whose date has passed.
Nine days later, in a written Q&A with SemiWiki, Higginbottom names semiconductor processes (plasma etching, deposition and reactor design) as the strongest application, does not mention fusion or any released model, puts the cadence at 'a new model generation roughly every three months' with half the raise for compute and half for headcount, and says training directly on experimental output is under way. That differs from the interview's fusion first, then chips, and new versions every month or couple of months; the first model's release still has no public record.SemiWiki, CEO interview ↗
Read 2026.10.04 The full transcript of the interview, read in full. The host is Seb Johnson and the founder describes his own company; no result is shown in the video, and no technical material from the company was read. Open the interview ↗