When it comes to nuclear fusion, almost all the attention goes to plasma: extreme temperatures, giant magnets, energy records, and the promise of replicating on Earth part of the process that powers the stars. But a commercial reactor will not just have to create plasma. It will have to continuously control it, interpret thousands of sensors, react to instabilities in fractions of a second, and coordinate vastly different physical systems without damaging a machine that could cost billions.
It is on this less spectacular but essential layer that Fusionality wants to work, a British startup founded by two former Google DeepMind researchers, Jonas Buchli and Brendan Tracey. The company is building software and control systems for fusion reactor developers and has raised an initial round led by Project A with participation from several deep-tech investors.
DeepMind had already shown that AI can control plasma
The link between artificial intelligence and fusion is not new. In 2022, DeepMind and EPFL's Swiss Plasma Center published a paper in Nature in which a reinforcement learning system directly controlled the magnetic coils of the TCV tokamak, successfully maintaining and shaping different plasma configurations.
That result was significant because controlling a tokamak is an extremely complex dynamic problem. The magnetic coils must be continuously adjusted, and decisions depend on the state of the plasma, which evolves rapidly. Instead of manually designing a separate controller for each configuration, the algorithm learned a control policy through simulations.
A commercial reactor will need much more than a good controller
Moving from a research experiment to a power plant means multiplying complexity. The system will have to integrate diagnostics, magnets, heating, power supply, vacuum, thermal management, and component protection. It will also have to distinguish between normal fluctuations and signals that foreshadow an event capable of disrupting the plasma or damaging the reactor.
Fusionality aims to build a platform that combines simulation, machine learning, control, and software tools to help fusion companies design and validate these systems before the final hardware exists. It is an interesting strategy because the sector includes numerous startups experimenting with different architectures: tokamaks, stellarators, inertial systems, and other configurations. Not all of them will need the same software, but all of them will need control.
Fusion can create a new category of industrial software
Traditional power plants already use sophisticated control systems, but fusion adds a dynamic layer that makes the use of predictive models and machine learning particularly compelling. Software can analyze high-frequency signals, predict unstable conditions, and select corrective actions before the issue becomes apparent to an operator.
This does not mean handing over a nuclear power plant to an opaque model. In a safety-critical system, AI will have to coexist with verifiable bounds, deterministic controls, and independent safety procedures. The challenge will be to leverage machine learning where it offers advantages without making it impossible to prove that the system stays within acceptable conditions.
Simulation is the real laboratory
A unique hurdle in fusion is that there are not yet thousands of power plants from which to collect operational data. Training algorithms therefore requires simulators capable of accurately representing the machine's behavior. It is the same principle used in robotics and autonomous driving, but with even more extreme physical phenomena.
The more accurate the simulator, the more possible it is to test rare or dangerous conditions without risking hardware. However, the so-called sim-to-real problem remains decisive: a policy that works perfectly in simulation can fail if the model does not reflect real-world dynamics. For this reason, the software will have to be continually compared with experimental data.
A “picks and shovels” startup for a highly uncertain race
Investing directly in a fusion reactor means betting on a specific physical architecture and commercialization timelines that remain uncertain. Building control tools can be a different strategy: selling technology to multiple developers, regardless of which project is first to deliver a working power plant.
It is the logic of businesses that sell picks and shovels during a gold rush. It does not eliminate risk—because if commercial fusion is delayed significantly, the software market will also grow slowly—but it reduces dependency on a single design.
The real test will come when AI has to be reliable, not brilliant
In the consumer world, a model is often evaluated on how well it performs on average. In a power reactor, rare events matter most. A system that works 99.9% of the time can be inadequate if the remaining 0.1% coincides with critical conditions.
Fusionality will therefore have to work on a form of AI very different from chatbots: less concerned with creativity, more with predictability, latency, and the ability to be validated. It is one of the reasons why fusion could become an intriguing testing ground for industrial artificial intelligence.
Igniting a star is not enough: it must also be controlled
Fusion research has made real progress, but no commercial power plant is yet supplying electricity to the grid, and company timelines remain ambitious and uncertain. It would therefore be premature to present Fusionality as the software for a future industry that already exists.
Its insight, however, is solid: if fusion does indeed reach electricity generation, the reactor will be both a machine of extreme physics and a computer system. The plasma may be the heart, but sensors, models, and controllers will be the nervous system. Building it in advance could prove almost as important as managing to ignite the reaction.
