By Proxima Co-Founder and Chief Engineer Martin Kubie

Most people assume a fusion company is a physics simulation company or an engineering design company. We're both, but we’re also an advanced manufacturing company, a specialist procurement company, and an industrial facility development company. At its core, Proxima Fusion is fundamentally something deeper: a technology company.

We don't just design stellarators at Proxima. We attack every problem we face with advanced technology. If problems have a technical solution, we'll adopt tooling, write software and build platforms to support any function across the company. In 2026, that’s why AI shows up everywhere – it’s an advanced technology capable of solving broad business problems.

That points to a bigger bet, which I'll come back to: the next era of hardware won't just be accelerated by AI, it'll be designed by it… but only if we have an engineering platform making that possible. We're building that at Proxima.

From simulation-driven to data-driven

Two years ago, we called ourselves a simulation-driven company. I wrote that our simulation-driven approach was a stepping stone towards the data-driven approach we were really aiming for. And now we’re here.

The path was a progression: people wrote software to support their analysis, those tools matured into simulation environments we used for production design, we began generating very large datasets and using them to drive design decisions directly.

That splits our use of AI into two pillars I find useful to think about separately – machine learning (data-driven methods, training models on large datasets, integrating them into how we make design decisions), and the application of LLMs (coding agents, autonomous agents, the bigger generative shift of the past two years).

Pillar one: machine learning

Pillar one is that data-driven approach we have been developing since Proxima was founded. Stellarator design is, at heart, a search problem in a brutally high-dimensional space, and machine learning lets us run that search at a scale no human could manage. We use it three ways:

1) We train surrogate models that learn the physics of a sub-system and then stand in for expensive simulators, collapsing the time it takes to assess a design and letting us evaluate millions of candidate designs instead of hundreds.

2) We explore the parameter space: engineers and physicists can see trade-offs between plasma performance, manufacturability, and machine cost on data we generated by simulating at scale.

3) We discover better geometric representations that are a better fit for our stellarator design optimisation tasks.

Pillar two: large language models

The change I didn't see coming this fast is the way LLMs have democratised writing software across Proxima.

We have over 100 GitHub users, but only 8 professional software engineers. All types of engineers, from mechanical engineers to magnet modellers – people with world-class judgment in their domain, who maybe used to struggle to write code – are now building software tools to directly solve design problems.

Our team is genuinely AI-first – agents accelerate building software, finding that old doc you half remember, automating data analysis, and generating diagrams from whiteboard sketches. We’ve even built our own: AlphaClaw, an AI agent for engineers, in collaboration with Google. It runs in the cloud – when it is tagged in Slack, it jumps on a design question or a data analysis task and reports back.

Where this goes next

The last twenty years of the global economy were revolutionised by software. The next twenty will be revolutionised by AI-driven hardware, and we intend to be at the centre of that shift.

Software development has been transformed by AI because software has especially agent-friendly infrastructure: artefacts as versioned code, executable tests, explicit interfaces, reproducible environments, and tooling for reviewing changes. A similar collaboration layer doesn’t exist for hardware engineering yet, so we're building it – the infrastructure that enables frontier models to do engineering alongside humans.

Our simulation- and data-driven approach is the foundation for fully agentic engineering systems where AI drives the design loop, rather than simply assisting it. We’re building the team, tooling, and processes to enable a new AI-driven hardware paradigm, because you can’t build the next energy era's infrastructure on the engineering practices of the last one.

To do that, we need a different kind of team: physicists who think about platforms, software engineers who think about machines, mechanical engineers who write code, and AI practitioners who want to point their work at something that deeply matters.

If that's the kind of team you want to join, we're hiring.