A Superconducting AI Chip Kept Up With Fusion Data, but Hasn't Controlled a Reactor
Brookhaven and its partners report that a superconducting neural-network chip handled fusion data sampled one million times per second. The prototype also processed simulated plasma signals, but the announcement does not establish live reactor control, prediction accuracy or an independently measured efficiency advantage.
A superconducting neural-network chip has processed data relevant to fusion experiments, according to an October 19, 2026 announcement from Brookhaven National Laboratory. The work with Great Sky, Princeton Plasma Physics Laboratory and DIII-D is a hardware demonstration, not proof that the chip has stabilized a working fusion reactor.
The 30-second summary
- What happened? A prototype processed recorded tokamak data and simulated plasma signals in three tasks.
- Why does it matter? Rapid analysis could eventually help controllers recognize changing plasma conditions before they disrupt an experiment.
- What is the catch? Keeping pace with an input stream is different from making sufficiently accurate predictions and acting on them in a live machine. The team is still preparing its technical paper.
Key Number: One million samples per second was the collection rate of the DIII-D data used in two tasks, not a published accuracy score.
What the chip actually processed
Two tests used DIII-D measurements for anomaly classification and disruption prediction. A third used signals generated with TORAX to predict sawtooth instabilities. The distinction matters: recorded measurements and simulation outputs can test processing, but neither alone establishes successful feedback control.
TORAX's documentation describes an open-source simulator for transport in a tokamak's core. It supports modelling, experiment planning and controller-design workflows, including models of heat and particle transport. Its software also includes sawtooth triggering and redistribution. Simulation is therefore a way to explore specified plasma behaviour, not a substitute for demonstrating the complete controller on a physical machine.
A neural network built into circuits
In Great Sky's architecture, superconducting circuits perform neural computations while optical connections carry information. The company describes storing connection weights near the circuits that use them, reducing the need to shuttle those values between separate memory and processing components.
This is a different hardware approach from implementing neural-network operations as software on conventional processors. It is not a claim that the device is a biological brain. Nor should the company's broad cost and speed projections be treated as measurements established by these fusion tests.
Why plasma control matters
As ITER explains, a tokamak uses magnetic fields to confine electrically charged plasma inside a doughnut-shaped chamber. Keeping that hot material away from the walls is central to the design. A useful computing system must therefore contribute to an entire sensing-and-control chain, rather than merely process data quickly.
Before we overstate the result
- The announcement does not provide prediction-accuracy or false-alarm figures for these tasks.
- No closed-loop stabilization result is reported.
- The demonstration does not establish net electricity production or a system-level efficiency advantage over GPUs.
The next test is integration
Brookhaven describes a larger-network demonstration planned for early 2027, followed by deployment at DIII-D later that year. Those are plans. The meaningful next evidence will show how predictions perform when integrated with diagnostics and control during actual experiments.
Image: Great Sky's superconducting neural-network chip used in the demonstration. Credit: Great Sky Inc., via Brookhaven National Laboratory.
Verified topics and entities
Sources and citations4 sources
External references used to support the reporting in this article.
Published by
NewTqnia Energy Desk
An institutional editorial team within NewTqnia
