Commercial fusion ventures face an expensive engineering paradox: while clean baseload power is becoming the existential bottleneck for scaling AI compute and data centers, nearly every fusion startup insists on building its reactor control stack entirely from scratch. Lausanne-based Fusionality, founded by ex-DeepMind researchers Federico Felici and Jonas Buchli, is stepping in to turn that bespoke, redundant engineering into off-the-shelf software and simulation infrastructure. Backed by a $3.7 million (CHF 3 million) pre-seed round led by Founderful and Playfair, the team is carving out an infrastructure niche at the operational layer of magnetic confinement reactors.
The Commonality in Reactor Controls
Controlling superheated plasma requires microsecond-level magnetic adjustments to preserve equilibrium, balance fuel density, and avert disruptive instabilities. Reactor geometries vary across hardware players, but the underlying plasma dynamics and control topologies share enormous overlap. Federico Felici, CEO of Fusionality, argues that the sector's habit of reinventing foundational software stacks wastes scarce capital and delays commercial power generation.
"Across the industry, many companies make their own control systems from scratch. But actually, 80% of each company's control system is really exactly the same."
Fusionality's thesis is straightforward: supply a standardized modular suite of plasma control primitives and high-fidelity physics simulators. Hardware teams can integrate these baseline models and fine-tune them for their proprietary vessel geometries rather than burning runway building foundational telemetry and feedback loops.
Pragmatic Scope for Machine Learning
Felici and CTO Jonas Buchli initially collaborated on reinforcement learning models to control experimental tokamak plasma at EPFL before Felici joined Google DeepMind to advance autonomous fusion control architectures. While hype-fueled marketing often promises end-to-end autonomous fusion plants out of the box, Fusionality’s commercial roadmap is disciplined: selling domain-specific ML middleware to accelerate the fusion hardware roadmap before compute power demands outstrip the grid.