Alex Higginbottom is co-founder and CEO of Zenithon, which he started with Abetharan Antony in July 2025 to build the first world models for extreme physics. He previously worked in deeptech commercialisation and spent time in fusion research, starting a PhD in the field before leaving after a year to build the company.
Tell us about your company.
Zenithon is building the first world models for extreme physics to keep humanity on the exponential of technological innovation. We’ve built a team comprising the minds behind the biggest breakthroughs in AI for physics in the last 5 years and have just raised $10 million to accelerate our development of world models for the world’s most critical technologies.
What problems are you solving?
The most complex machines ever built are still developed by trial and error. Build, test, fix, repeat, taking weeks or months per cycle. That was not a failure of engineering, it was the only method available because the most critical systems are the ones nobody can see inside and nobody can compute fast enough to search. A plasma etch chamber cuts features a few atoms wide using a process no instrument observes directly, under recipes with on the order of 100 trillion possible combinations. So engineering proceeds by increments, because trial and error explores by increments. We’re building models to explore a million design points in the time a conventional solver takes to run one.
What application areas are your strongest?
Semiconductor processes, particularly plasma etching, deposition, and reactor design and optimization. Currently, R&D is driven by trial and error with experiments, and this is to search a parameter space with 100 trillion combinations. Our world models can search this space a million times faster and provide far more optimal designs, far quicker.
What keeps your customers up at night?
For a processes engineer, the concerns are cycle time and the fact that intuition has run out. Beyond a few dimensions you can no longer reason your way to the right recipe, and the people who spent careers building that intuition are retiring. Yield and edge exclusion are the metrics, but the underlying anxiety is that the search space grew faster than the tools could handle.
What does the competitive landscape look like and how do you differentiate?
The thing we compete with most is the existing method. Most engineers we talk to are choosing between us and running another simulation, not between us and another model. Companies are building analogous technology for very different problems, such as world models for robotics or factory optimisation, but the physical regimes and the architectures those regimes demand are not the same.
The harder competitor is the track record. Machine learning has been applied to semiconductor process before and mostly disappointed. Each problem needed its own dataset, generating it took long enough that the answer was half known by the time it existed, and changing the design made it worthless. Our answer is that our models are pretrained across these problems rather than fitted to one, and that we build the data generation layer ourselves because no public corpus exists. Whether it holds is something we intend to show rather than assert.
What new features and technology are you working on?
We’re aiming to ship a new model generation roughly every three months, with roughly half the raise going to compute and half to headcount. The direction is scale: larger models trained across more physical regimes, so a customer uses one product across many use cases instead of commissioning a bespoke model per problem. We’re also working on training directly from experimental output rather than simulation alone, which closes the sim-real gap rather than widening it.
How do customers normally engage with your company?
Customers book demos with us through our website. We start with a pilot and then, depending on the situation, they can arrange a custom deal for access to support and our full world model, fine tuned to their system.
Also Read:
CEO Interview with Paul Butler of VeriQuantix
CEO Interview with Joel Jorgensen of FlowAccel
CEO Interview with Vivek Raghuraman of Mixx Technologies
Share this post via:
The silent killer of analog reliability: Why SPICE misses floating net