The most complete fusion machine ever built will never generate a single watt of electricity. That is not a delay, and it is not a failure. ITER, the 23,000-tonne tokamak now being assembled in southern France, was designed as a research facility to demonstrate sustained net-energy fusion reactions — not to produce power for a commercial grid.
It is enormous by any measure. The toroidal vacuum vessel spans 19.4 metres across. The superconducting magnets weigh 10,000 tonnes and store 51 gigajoules. The machine is expected to deliver around 500 megawatts of thermal output.
Now set that beside what surfaced days ago. A US company called Thea Energy unveiled a design for a plant that would feed a steady 400 megawatts of electricity into the grid — electric, not thermal — from a device whose major radius is just 8 metres. The plant, Helios, is slated to launch as early as the 2030s.
The gap between ITER's thermal research machine and Helios's 400 MW electric design is the week's real story.
That gap is the story of this week.
Last week the fusion conversation was about components: high-temperature superconducting magnets and AI-assisted design, pitched as the two variables that could compress the timeline. This week it moved from components to a floor plan. Someone drew the power plant.
Helios is a stellarator, the twisted-magnet cousin of the tokamak. Conventional stellarators are held back by their coils — bulky, complex, three-dimensional shapes that are brutally hard to manufacture. Thea's answer is to replace them with flat, planar magnets under software control, which the company says makes the design simpler and cheaper to build.
The plant is also designed for continuous operation, not pulses, and it is divided into independently serviceable sectors so the fusion hardware stays accessible for maintenance across its operating life.
Before any of that, though, Thea has to build Eos — described as the first large-scale stellarator using a planar-coil architecture, intended to test whether the approach is viable for a commercial plant. Which is exactly where the scepticism belongs. The company has a concept, a timeline, and an unbuilt prototype standing between them.
The same week, on the other side of the world, the case for that bet got louder.
At a fusion forum in Shanghai, Zhong Wulü, chief engineer of China Fusion, argued that high-temperature superconductors and AI together could shorten fusion's development cycle — and that China should exploit both. More than 70% of newly designed fusion devices worldwide, including tokamaks and stellarators, now use high-temperature superconducting technology, according to reporting from the forum.
The magnet numbers are already real. Song Yuntao of the Hefei Institutes of Physical Science said his institute's all-superconducting magnet reached 35.1 tesla, and that the EAST tokamak has now logged over 150,000 discharges.
The stellarator argument got a specific hearing too. Lin Zhihong of Shanghai Jiao Tong University noted that stellarators need no large driven plasma current — which means steady-state operation and no risk of the plasma disruptions that plague tokamaks. Their weakness is the coil geometry, where traditional design iteration is prohibitively expensive. AI and high-temperature superconductors, he said, have sharply improved the engineering feasibility.
One number makes the whole commercial case concrete: European researchers estimate that a stellarator's average cost of electricity over its full life cycle is only 40% that of a conventional tokamak.
That is the intellectual backbone under Helios. Not a new physical insight — a manufacturing and control insight.
The AI half is arriving through a different door. On September 14, at the Global Fusion Policy Summit in London, a proposed UK-US federation was announced that would link the UK Atomic Energy Authority's SUNRISE supercomputer with Princeton Plasma Physics Laboratory's STELLAR-AI platform.
The two machines feeding it — MAST Upgrade in Oxfordshire and NSTX-U in New Jersey — are both compact spherical tokamaks with similar designs. That similarity is the point: models trained on one fusion device tend to struggle on another, so combining datasets should produce predictions that travel.
SUNRISE is backed by £45 million in UK government funding. The plan is to build digital twins of both machines, virtual models that let engineers test changes in software before touching hardware. Rob Akers, a senior fellow at UKAEA, said the two labs could combine experimental data with simulations to explore operating conditions never yet tested.
Shantenu Jha, head of computational sciences at PPPL, put the ambition plainly: "Our goal is to let models and experiments move freely between the two systems. We will turn a collection of supercomputers into a single engine for fusion discovery."
The caveat matters. The federation is not yet an operational shared supercomputer; the partners are still working through technical requirements and next steps.
And the hard problems have not moved. China's fusion programme lists three scientific obstacles — steady-state self-sustaining burning plasma, materials that survive high-energy neutron bombardment and heat loads, and tritium breeding and self-sufficiency — plus three engineering ones: high-field superconducting magnets, plasma operation and control, and heat conduction.
Tritium self-sufficiency and 14 MeV neutron damage do not become easier because the coils got flat.
Meanwhile the slow machine keeps delivering. ITER's central solenoid, a 59-foot superconducting assembly, was completed after a 15-year fabrication effort at General Atomics in Poway, California. Six modules, each over 135 tons, wound from roughly 3.7 miles of niobium-tin cable — plus a seventh spare held in reserve, a risk-management decision that tells you how much rework costs here.
ITER now reports its project roughly 85% complete overall, running about seven months ahead of its 2024 baseline, with eight of nine vacuum vessel sectors delivered to site and deuterium-tritium burning experiments targeted for 2035 to 2039. The European Commission's benchmarked programme cost sits near €22 billion.
That project will never light a city. Its value is different. ITER's head of communication, Laban Coblentz, has noted that a decade ago there were perhaps five fusion companies globally; that number has since grown to roughly 45 — growth he linked to the manufacturing milestones ITER forced into existence. The giant built the supply chain.
So the delta this week is not about physics. It is that fusion's engineering path has acquired a shape: a metre count, a megawatt figure, a maintenance schedule, a named prototype standing between the drawing and the grid.
Helios says 400 megawatts electric from an 8-metre device in the 2030s. Eos has not been built. That distance — between a design unveiled and a watt delivered — is still the entire story.