Google DeepMind takes the wheel of fusion energy

Taming a 100-million-degree ball of plasma on Earth has a classic problem: it is violently unstable, prone to snuffing itself out in milliseconds, and governed by far more operational knobs than any human pilot could ever hope to manage.

Now, Google DeepMind wants to take over the controls. The lab announced a partnership with Commonwealth Fusion Systems (CFS) to deploy advanced AI models and simulation tools to SPARC, the compact tokamak CFS is currently assembling outside Boston.

Millions of dry runs before the switch is flipped

SPARC is roughly two-thirds built, with completion targeted for late 2026. Its goal is historic: becoming the first magnetic confinement machine to demonstrate net fusion energy, producing more power than it takes to sustain the reaction. But real-world pulses are scarce and expensive, making live trial-and-error nearly impossible.

That is where DeepMind's plasma simulator, Torax, steps in. Built on JAX to run seamlessly on GPUs and CPUs, Torax allows researchers to simulate heat, electric currents, and material flows inside the plasma core. CFS has already integrated the open-source tool directly into its daily workflow.

By chaining Torax with reinforcement learning and evolutionary algorithms like AlphaEvolve, DeepMind's agents can run millions of virtual operational scenarios before SPARC ever fires up. The AI scours through countless combinations of magnetic coil currents, heating parameters, and fueling rates to map out the most reliable path to net power.

An AI co-pilot on the edge of melting walls

DeepMind previously proved it could shape plasma using reinforcement learning at the Swiss Plasma Center. With CFS, the physical stakes are significantly higher.

When running at full throttle, SPARC will release an enormous amount of concentrated heat onto a tiny surface inside the vessel. To prevent critical components from degrading, the system has to magnetically sweep that exhaust heat back and forth across the protective tiles like a paintbrush.

DeepMind is training reinforcement learning agents to execute that sweep in real time. Orchestrating that dynamic maneuver while simultaneously keeping the core plasma stable requires balancing hard machine limits that human-crafted heuristics struggle to handle.

Why it matters

Plasma physics has long been bottle-necked by multi-variable chaos, and modern reinforcement learning is uniquely suited to tame it. If DeepMind can help SPARC reach net energy on its 2026 timeline, it turns AI into a core catalyst for clean baseload power rather than just an energy-guzzling workload.

Google's broader play here is transparently transactional. Just this August, Google joined Nvidia in CFS's $863 million Series B2 round, and it previously agreed to purchase 200 megawatts of electricity from CFS's future commercial reactor, Arc. The search giant is also backing rival fusion outfit TAE Technologies.

We think this is the ultimate self-reinforcing tech loop: modern AI requires massive amounts of zero-carbon power to keep scaling, and hyperscalers are now deploying their most sophisticated algorithms to accelerate the exact power plants they intend to plug into.