Ten thousand times per second, a neural network decides the voltage on 19 magnetic coils. The coils hold a ring of plasma hotter than the core of the sun.
By early 2026 the same method had been demonstrated on DIII-D, one of the world's largest tokamaks, and peer-reviewed in Nuclear Fusion — without the equilibrium-reconstruction step every classical controller depends on.
A commercial supply chain is now forming around it. A spinout from the Swiss Plasma Center began selling fusion control and data systems in April 2026, founded by two ex-DeepMind robotics engineers.
The control layer of the fusion industry is quietly becoming an AI business.
A tokamak is a magnetic bottle with no lid. The plasma inside it — ionized hydrogen at over a hundred million degrees — wants to drift, stretch, and crash into the vessel wall within milliseconds. Holding it in place is a real-time control problem that has been solved with the same tools since the 1970s: equilibrium reconstruction feeding a cascade of hand-tuned PID loops.
That stack is about to be replaced. Not because it failed, but because it does not scale to the machines now being built. ITER, the next-generation tokamak under construction in France, needs controllers for plasma shapes that have never been run on a real device. Private fusion companies need to iterate on control systems faster than they can hire the handful of engineers in the world who know how to do it.
One policy, nineteen coils, no reconstruction
The classical approach works like this. Sensors read the magnetic field around the plasma. A separate algorithm, called an equilibrium reconstruction, converts those raw readings into a picture of the plasma's shape and position. A cascade of PID controllers compares that picture to the target shape and adjusts each coil's voltage accordingly.
The reconstruction step is the bottleneck. It is slow, it needs constant re-tuning for every new plasma shape, and it was never built for the exotic configurations researchers now want to test.
DeepMind's team, working with the Swiss Plasma Center at EPFL, trained a single reinforcement-learning policy that reads raw magnetic diagnostic signals and commands all 19 coils directly. No reconstruction in the loop. The policy learned to hold standard elongated shapes and ITER-like shapes. It also produced a snowflake configuration with multiple exhaust legs. And it held two separate plasma droplets in the same vessel at once — a feat no conventional controller on TCV had managed before.
Control-loop speed of a learned fusion controller
A single neural network adjusts coil voltages at kilohertz rates, replacing a hand-tuned stack of reconstruction and PID logic. · DeepMind / Nature, 2022
The agent did not learn in the reactor. It trained offline in a simulator, playing millions of trial-and-error episodes where a bad action costs nothing. The finished policy was then transferred to the machine — what the field calls zero-shot sim-to-real transfer. The team made the underlying simulator code, TORAX, open source in May 2024, which lowered the barrier for every other lab working on learned control.
What DIII-D changed
The 2022 TCV result was a proof of concept on a small research machine. The question was whether the approach would survive contact with a bigger, faster, less forgiving tokamak.
In January 2026, a paper in Nuclear Fusion demonstrated it did. A team led by G. F. Subbotin of Next Step Fusion ran a reinforcement-learning controller on one of the largest operating tokamaks in the world. General Atomics and the University of California San Diego contributed engineers. The controller mapped raw magnetic diagnostic signals straight to actuator commands, eliminating equilibrium reconstruction from the control loop entirely. No per-shot tuning was needed to reach target parameters on early discharges.
| Parameter | Classical control (isoflux + PID) | RL controller |
|---|---|---|
| Equilibrium reconstruction | ✔ Required every cycle | ✗ Skipped entirely |
| New plasma shape | ✗ Days of manual re-tuning | ✔ Retrained in hours |
| Advanced configurations | ✗ Struggles (snowflake, droplets) | ✔ Demonstrated on TCV and DIII-D |
| Safety authority | ✔ Deterministic, fully auditable | ◐ Learned, wrapped in safety layers |
The result on the largest device matters for two reasons. First, the learned policy transfers across machines. The training simulator, NSFsim, reproduces kinetic parameter evolution alongside magnetic equilibria — the team identified that as the key factor in control quality. Second, it removed the reconstruction dependency that every commercial reactor design inherited from the research era.
The paper also recorded where the method still falls short. The vertical X-point position — the point where the plasma's magnetic field structure splits — showed errors of up to 4 cm, attributed to current-density mismatches between simulation and experiment. That gap between the simulator and the real machine is the central risk of the whole approach.
The supply chain problem nobody was solving
Fusion has a skills bottleneck that has nothing to do with physics. Dozens of well-funded private companies are building prototype reactors, and almost all of them need the same thing: plasma control, instrumentation, and data systems. There are too few experts globally to build those in-house for every company.
That gap is where Fusionality sits. The company was founded in April 2026 in Lausanne, near the Swiss Plasma Center, and spun out decades of control technology developed on TCV and other tokamaks. Its two co-founders are Federico Felici and Jonas Buchli — both ex-Google DeepMind, both former EPFL researchers, and both named authors on the 2022 Nature paper that started this thread.
The company sells control, instrumentation, and data capabilities as a service to fusion builders, positioning itself as a supply-chain company for today's prototype reactors and tomorrow's power plants. In July 2026 it received CHF 40,000 from the Swiss Venture Kick program to commercialize its control systems. The money is small; the signal is not. Venture Kick backing means the control layer of fusion is now treated as a fundable business, not a research byproduct.
First public funding for a fusion-control vendor
Fusionality received the award to advance control and instrumentation systems for the emerging fusion industry. · CB Insights / Venture Kick, 2026
The commercial logic mirrors what happened in other capital-intensive industries. When every builder needs the same specialized subsystem, a dedicated supplier emerges to serve them all. The control and data layer is precisely such a subsystem — high technical barrier, tiny talent pool, identical requirements across competitors.
Washington entered the picture
Governments have noticed that AI is now a fusion dependency. In November 2025 the US administration launched the Genesis Mission by executive order, a national effort to apply AI to scientific discovery across Department of Energy laboratories. STELLAR-AI, launched in January 2026, is the fusion arm of that program. It is led by the Princeton Plasma Physics Laboratory (PPPL), with partners including the UK Atomic Energy Authority, MIT, and the University of Wisconsin-Madison.
STELLAR-AI pairs AI with high-performance computing to speed the simulations that fusion research depends on. A single high-fidelity fusion simulation can take months on existing infrastructure. STELLAR-AI's stated goal is to compress that timeline and connect computing resources directly to experimental devices, such as PPPL's NSTX-U, so researchers can analyze data as experiments run.
Its sister project, StellFoundry, applies AI to the design of stellarators — the twisted, pretzel-shaped fusion devices that require sifting through enormous configuration spaces. Stellarator design has historically taken months or years per candidate; the AI-assisted workflow is designed to compress that into days.
The institutional money matters for the same reason the venture money does. When a national laboratory builds its AI-fusion platform around these tools, it de-risks the technical path for every private company in the sector.
Where the learned controller still answers to physics
The honest limits are worth spelling out, because fusion is an industry built on overpromising.
What is still unresolved with learned plasma control
One policy, one machine. A controller encodes one device's geometry and dynamics. Each new tokamak needs its own simulator and training run — what transfers is the method, not the weights.
Safety authority. A learned inner loop must stay wrapped inside deterministic interlocks that retain final authority. That layering is standard practice now, but it means no learned component ever runs without oversight.
Verification at reactor scale. On a power plant, an off-normal event is far more costly than on a research machine. The bar for certifying learned control will be higher than any demonstration so far.
None of these limits invalidate the direction. They define the timeline. The credible near-term path is hybrid: learned inner loops handling shape and instability avoidance inside mature, classical safety systems, with verification maturing before learned components take on more authority. That is exactly the architecture every serious actor in this space is converging on.
The learning curve runs through real machines
The DIII-D result did not arrive in a vacuum. It is the latest step in a lineage that started with control-adjacent AI work on other devices. In 2021, researchers at KSTAR, South Korea's superconducting tokamak, demonstrated deep-reinforcement-learning feedforward control of plasma beta — the ratio of plasma pressure to magnetic pressure. In early 2024, a team led by Princeton Plasma Physics Laboratory and General Atomics published a Nature paper on that machine. Their controller predicted tearing-mode instabilities and steered the plasma away from them before they grew. Each step moved learned control from the periphery toward the core of the magnetic-confinement loop.
The pattern across all of these results is the same. A simulator absorbs the expensive part of the learning. A policy transfers to the real machine with minimal or no re-tuning. And a deterministic safety layer keeps final authority. The reconstruction-free controller is significant because it removed the last classical component from the inner loop — but it sits inside that same now-standard architecture.
That architecture is worth understanding as an investment concept in its own right. The control software is not the reactor. It is the layer that makes every reactor design testable. A company that owns proven, transferable control software can attach it to whichever reactor design wins — without betting on the plasma physics itself. That is a structurally different risk profile from betting on a single tokamak geometry or a single fuel cycle.
It also explains why the talent pool matters so much. The people who can build these controllers are counted in the dozens worldwide. Fusionality's founding team is a direct transfer of that expertise into a commercial vehicle — the same human-capital arbitrage that seeded the semiconductor supply chain. The scarce resource is not the physics or the compute. It is the small group of engineers who have run learned controllers on real tokamaks and know where they break.
Why this looks like an industry forming, not a research trick
Three independent signals point the same way. The peer-reviewed demonstration on a major tokamak. The dedicated supplier that emerged to sell control systems commercially. And the national platform that built its research agenda around the technique.
When all three appear inside eighteen months, it is usually safe to stop calling something a research novelty.
Fusion is a complex system of systems. We need AI and high performance computing to really optimize the design for economic construction and operation.— Jonathan Menard, deputy director for research, Princeton Plasma Physics Laboratory
The investment angle is less about any single company than about where value accrues in a supply chain that is still taking shape. The reactor builders are the visible names. The control layer — the software and instrumentation that every one of them needs and none of them wants to build from scratch — is the quieter position. The technique was proven on TCV. It proved transferable on the larger device. The spinout is betting that the world's private fusion companies will pay rather than rebuild that expertise internally.
That bet is early but legible. The same pattern ran through aerospace, then semiconductors, then battery manufacturing. A critical subsystem gets spun out of a research institution. A small vendor captures the demand from every integrator. The vendor compounds while the integrators compete on the parts they actually differentiate.
Signals to track
A second peer-reviewed RL control result on a different tokamak class — proves the method generalizes beyond DIII-D and TCV.
The first named fusion company that contracts a dedicated control vendor rather than building in-house — confirms the supply-chain thesis.
NSTX-U coming online under STELLAR-AI — the first large device where AI control and data infrastructure are designed in from the start.
Any Fusion Energy Design or Nuclear Fusion paper reporting sub-centimeter shape error from a learned controller — the main credibility metric for closing the sim-to-real gap.
