DeepMind alumni raise $3.7M for Fusionality, a startup building fusion reactor control systems

DeepMind alumni launch Fusionality with a $3.7M pre-seed

Two researchers with backgrounds at Google DeepMind have founded Fusionality, a Lausanne-based startup that wants to supply the control systems fusion companies currently build themselves. According to TechCrunch, the company has raised a $3.7 million (CHF 3 million) pre-seed round from Founderful and Playfair.

Federico Felici, who serves as CEO, and Jonas Buchli, the CTO, founded the company this year. Both had spent years developing methods for controlling experimental fusion devices, and Felici told TechCrunch that fusion companies repeatedly made the same complaint: they would prefer to buy components for their control systems rather than build everything, but no vendor existed that understood fusion engineering.

The problem: every fusion startup rebuilds the same controls

A fusion power plant would generate electricity from the energy released when atomic nuclei fuse, a reaction that requires sustaining an extremely hot plasma. Holding that plasma at the correct temperature, shape and fuel level demands decisions in fractions of a second, which makes the associated control software and hardware difficult and expensive to develop.

According to TechCrunch, many fusion companies construct their control systems entirely from scratch. Felici estimates that around 80 percent of each company's control system is effectively identical, because the underlying physics is consistent across reactor designs. Fusionality plans to build on that commonality with a suite of control systems and simulation environments that startups can adapt and fine-tune for their own machines.

The company sits inside a broader supply chain forming around the fusion industry. TechCrunch points to firms such as Kyoto Fusioneering, which develops components that help fusion startups turn engineering breakthroughs into grid electricity, alongside established manufacturers producing high-precision parts. Very few companies, however, specialize in the hardware and software needed to control the reactors themselves.

From tokamak research to DeepMind

The two founders met while teaching AI to control an experimental tokamak, a donut-shaped device used to study fusion. Felici was then at EPFL in Switzerland, which houses one such machine, while Buchli was at Google DeepMind. Felici later joined DeepMind too, where he worked on simulations and machine learning interfaces for fusion devices.

Fusionality will initially concentrate on magnetic confinement, an approach that uses powerful electromagnets to keep fusion fuel dense and hot enough to feed a power plant. TechCrunch names Commonwealth Fusion Systems, Realta Fusion, Proxima Fusion and Type One Energy as examples of startups pursuing magnetic confinement devices, and Felici said systems for other approaches will follow.

AI as a component, not the whole answer

Despite the founders' machine learning pedigree, Felici was careful to bound AI's role. He told TechCrunch he would not advocate handing control of an entire fusion reactor to AI, saying current systems are more likely to complement, enhance or optimize parts of the control stack, though he expects AI to play an important role in future control systems.

The seven-person team will use its pre-seed funding to develop a tightly selected set of technologies. Felici compared them to Lego blocks — starting with a few and adding more over time — and said the ambition is to eventually cover the wide range of technology needed to operate a fusion reactor. He declined to tell TechCrunch which specific technologies come first.

Why it matters

Fusion startups have attracted heavy investment on the promise of grid-scale power, yet many are independently solving near-identical engineering problems. If Fusionality's bet on shared control infrastructure is right, it could shorten development cycles across the industry, much as standardized tooling did for software. The launch is also a grounded example of applied AI: machine learning expertise from DeepMind deployed as an optimization layer inside reactor control rather than as a wholesale replacement for conventional systems. For the grid, the open question is whether a common control layer can establish itself before the first wave of commercial fusion plants locks in fully proprietary stacks.