Nuclear fusion: AI overturns 59-year-old stellarator conjecture

A fusion headline that came out of a proof, not a reactor

One of this autumn's most talked-about fusion results came from a chat window, not a tokamak. On September 22, 2026, University of Maryland plasma physicist Matt Landreman posted exact solutions to the equations that underpin stellarator design. Harold Grad conjectured in 1967 that solutions of this type did not exist.26 According to the paper's acknowledgments, OpenAI's GPT-6 Astra Pro found the solutions, and Landreman checked every equation by hand.26

Grad's idea was a real problem for stellarators. These machines are deliberately twisted, so they lack the symmetry of a doughnut-shaped tokamak. Grad argued that smooth plasma equilibria with varying pressure would need an extra symmetry of that kind.26 If he was right, the theory behind stellarators stood on shaky ground. Coverage says the model returned two families of counterexamples in about two days. The first, which appeared roughly 20 minutes after Landreman's prompt, had an integer rotational transform. The second had irrational transform values, a property that matters for good confinement.9 Both families were checked with numerical codes and symbolic algebra software.9

The AI was not the only route to the answer. One day before Landreman's first version, mathematicians Javier Gómez-Serrano, Lukas Liehr and Mitchell Taylor posted a proof that such counterexamples exist, certified in the Lean proof assistant.26 Other reporting describes that work as a 147-page proof that also used AI models, including GPT-5.6 and Claude.9 A generalization followed on October 5.26 Google's Gemini 4 Argon was then reported to have extended the solutions further.27 So the conjecture fell to a group of human-AI teams within roughly two weeks, not to a single system.

Where the coverage agrees, and where it inflates

The basic facts are consistent across outlets: Grad's 1967 conjecture, Landreman's role, the September dates and the GPT-6 Astra attribution all match.92627 The framing does not. One headline declared "It's over, AI won." The same article then said plainly that the work concerns how to model magnetic fields that hold superheated fuel. It does not produce electricity or make fusion commercially viable.27 A more careful account noted that neither paper says anything about when a reactor will run.26

That is the right reading. A counterexample to Grad removes a theoretical doubt about one class of fusion device. It does not show that a working stellarator is any easier to build. The bigger lesson is about which fields AI is reaching first. Results that can be checked mechanically are moving fastest: Lean proofs, closed-form equilibria that can be verified symbolically, and calculations anyone can rerun.26 The same coverage cites an AI agent that solved 49 of 50 unpublished physics problems for an average of under $3 each.26

The mathematics wave around it

The fusion result is part of a much larger run of AI-assisted mathematics. On October 6, OpenAI released 377 results. One of them claimed to settle the "quasi-Riemann hypothesis", a weaker form of the Riemann hypothesis. Stanford's Jared Duker Lichtman called it the most significant AI-for-math achievement so far, by a wide margin.11 Number theorists say the proof brings no new concepts. Instead, it finds unexpected links between existing techniques, including ones developed years earlier by Maksym Radziwiłł and Alexander Dunn.11

On October 26, OpenAI announced three more claims. The first is a disproof of the Exponential Time Hypothesis, a computer science assumption dating to 1999. The second is a proof of the local smoothing conjecture in every dimension, which would also settle the Kakeya and Stein restriction conjectures. The third is an existence proof for four-dimensional quantum field theories, including quantum chromodynamics.1719 The company published natural-language write-ups and Lean proofs for the mathematics and computer science results. It also brought in Russell Impagliazzo, Ramamohan Paturi and Ryan Williams to lead an expert review of the complexity result.1720 The physics materials have not been released in full.17 Even taken at face value, the Yang-Mills work covers existence only, not the mass gap, so the Clay Millennium Prize is still unclaimed.14

The reports differ most on verification. One summary said the findings had been verified by outside experts.18 Other coverage stresses that a Lean-checked result still needs people to examine it and understand what it means.20 The more careful framing is the one to trust. Lean confirms that each step follows from the last. It does not confirm that the formal statement matches the intended theorem, and it does not tell anyone why the proof works.

The human cost is also becoming visible. Reporting on OpenAI's Navier-Stokes claim describes an agent swarm that reportedly used $15 million in computing. It also describes a company that, by its own researcher's admission, had no research-level expertise in fluid dynamics when it announced the result.13 In a separate case, OpenAI's Yang-Mills work beat an independent proof that Anthropic's Levent Alpöge and Northwestern's Ethan Sussman had been developing in public.12 Machines are producing proofs faster than mathematicians can absorb them, and fusion theory is now caught up in that.

Materials science: the reality check

Materials discovery is where AI's record looks most mixed, even though many fusion researchers see it as the most promising use. Columbia's Carlos Paz-Soldan said he is personally more excited about AI for materials discovery than about AI chips for watching plasma.21 The record so far is less impressive. In January 2026, Nature corrected Berkeley Lab's robot-chemist paper. The corrected count is 36 compounds made from 57 targets, and "new" turned out to mean new to the prediction platform, not necessarily new to science.22 The same analysis notes that only 30% of the lab's 353 recipes worked. It also says Google DeepMind's 2023 claim of 2.2 million predicted stable crystals is still disputed on novelty.22

A new fusion result shows why materials matter so much. A Princeton Plasma Physics Laboratory model suggests that heating plasma before compressing it could sharply cut the energy needed for ignition.8 The same work warns that tungsten at just one part in 10,000 could double the pressure ignition requires. That is a serious concern given that more than a dozen next-generation machines use tungsten walls.8 The researchers suggest a liquid lithium coating could keep tungsten out of the plasma. The findings are still theoretical.8 The wall material, it turns out, can make ignition much harder.

Meanwhile, the machines keep moving

Hardware progress is steady but small. Canada's General Fusion says its LM26 prototype reached electron temperatures of 1.1 keV, above 12 million degrees Celsius. The company says it got there with mechanical compression rather than superconducting magnets or lasers.10 It has published the data and submitted it for peer review.10 In June 2026, Lawrence Livermore's National Ignition Facility achieved ignition for the 11th time, with a yield of 7.9 MJ and a target gain of about 3.8.5 The facility's record is still 8.6 MJ, set in April 2025.5 China's EAST tokamak reported operating in a "density-free regime", staying stable well past long-standing empirical density limits. It did this by managing how the plasma interacts with the reactor wall from the start of each shot.2 Helion reported a 150-million-degree plasma and measurable deuterium-tritium fusion in February, but no net energy gain.1

AI is also showing up in reactor control. Brookhaven National Laboratory and the start-up Great Sky unveiled a superconducting neuromorphic chip. In tests, it kept pace with recorded DIII-D tokamak data at a million samples per second. The chip has not yet been connected to a live machine.21 Peer-reviewed AI plasma control does exist. Princeton's controller warns of a tearing instability about 200 milliseconds ahead. However, it was tested on a single tokamak, and human operators still set its parameters.22 The broader timeline is still long. ITER's plasma operations have slipped from 2025 to 2034.22

The reading

Seen together, this coverage points to one conclusion. AI is now good at solving well-posed, checkable problems in fusion theory and mathematics, and the Grad counterexamples are an example. It is not yet good at the hard physical problems that decide whether fusion reaches the grid: wall materials, impurities and engineering. The stellarator result is a real intellectual milestone. Treating it as an energy breakthrough confuses removing a theoretical doubt with delivering power. The next tests are whether AI-designed materials hold up at the wall and whether AI controllers can manage live plasma. The coverage so far suggests both will be judged more strictly than the proofs were.