Former DeepMind Researchers Launch Fusionality to Sell AI Plasma Control Software to Fusion Builders

The builders of the world's experimental fusion reactors face a software bottleneck: they must either code their own plasma control systems from scratch or buy a standardized layer. Fusionality, a Lausanne-based startup founded by former Google DeepMind researchers Federico Felici and Jonas Buchli, is betting they will choose the latter. The company emerged with a CHF 3 million ($3.7 million) pre-seed round led by Founderful and Playfair, aiming to sell the reinforcement learning techniques that previously stabilized plasma in a Swiss tokamak.[1][2]

The fusion industry has raised $14 billion across 56 companies since 2021, with $4.48 billion of that arriving in the 12 months ending July 2026. Yet much of that capital flows into constructing custom hardware—magnets, vacuum chambers, and lasers—while the software required to operate them remains highly fragmented. Currently, more than 30 private fusion companies and government projects are building their own control systems independently.[2]

Fusionality is pitching a "picks and shovels" business model for this emerging sector. Rather than attempting to build a reactor and generate net-positive energy themselves, the founders are offering a reusable software layer designed to require minimal customer-side adaptation. The goal is to reduce the onboarding costs and development time for companies that would otherwise spend years writing bespoke control algorithms.[1][2]

To understand why fusion requires millisecond-level artificial intelligence, one must look at the physics of magnetic confinement. Inside a tokamak—a donut-shaped vacuum chamber—a cloud of hydrogen isotopes is heated to roughly 100 million degrees Celsius, transitioning into a plasma state. At that temperature, the plasma will instantly melt any physical material it touches, meaning it must be suspended in mid-air using powerful magnetic fields.[2]

However, plasma is inherently unstable, turbulent, and highly sensitive to minor fluctuations. As it swirls, its shape, density, and position shift rapidly. A reactor's control system must read raw diagnostic data, calculate the plasma's exact state, and adjust the magnetic coils thousands of times per second to prevent the cloud from striking the reactor walls.[2]

Traditional control systems rely on pre-programmed mathematical models that attempt to predict plasma behavior. These models are computationally heavy and often struggle to react to sudden, non-linear instabilities. "We have spent most of our careers making plasmas in real fusion devices behave as we wanted," said Felici, Fusionality's chief executive. "Now we founded Fusionality to bring that expertise to our customers and partners, helping them tackle some of the hardest bottlenecks in fusion operations."[1]

The founders' technical claims rest on a landmark 2022 paper published in the journal Nature. While at DeepMind, Felici—a control engineer with a PhD in plasma physics—and Buchli—who led robotics and reinforcement learning teams—applied deep reinforcement learning to the TCV tokamak at the Swiss Plasma Center in Lausanne.[1][2]

Instead of manually programming the control algorithms, the DeepMind team trained an artificial intelligence agent inside a digital simulation of the reactor. The agent learned through trial and error how to adjust the magnetic coils to achieve specific plasma shapes. When transferred to the physical TCV reactor, the AI successfully sculpted the real plasma into complex configurations, including a "droplet" shape, without requiring manual tuning.[2]

Fusionality intends to commercialize that specific capability, moving it from a research demonstration to an enterprise product. The company's software is designed to convert raw plasma measurements into control signals for a reactor's heating, fueling, and magnetic systems within milliseconds. This real-time processing must run continuously and reliably for a fusion device to function at all.[1][2]

Despite the proven academic pedigree, the commercial viability of this off-the-shelf approach remains untested. Fusionality faces direct competition from Next Step Fusion, an established supply-chain player that already provides plasma modeling, diagnostics, and control software to tokamak builders.[2]

More significantly, the startup's biggest competitive threat may be its founders' former employer. In October 2025, Google DeepMind partnered directly with Commonwealth Fusion Systems and open-sourced its TORAX plasma simulator, applying reinforcement learning to the company's SPARC reactor.[2]

If the industry's best-funded reactor builders can access DeepMind's reinforcement learning expertise and simulation environments for free, the addressable market for paid third-party control tools could narrow significantly. Fusionality is betting that as the sector matures, fusion companies will prefer a dedicated, supported software vendor over maintaining open-source code themselves, trading licensing fees for operational reliability.[2]