Net-Electricity Fusion Plant
No fusion device has yet delivered net electricity; ignition has been shown, the power plant as a whole has not.
Open in the interactive tree →Fusion merges hydrogen nuclei (mostly deuterium and tritium) at over 100 million degrees and releases per kilogram of fuel many times the energy of fission, with no chain reaction and no long-lived waste in the reactor. Two routes compete: magnetic confinement (tokamak, stellarator) and inertial fusion with lasers. So far plasma has been ignited, but no device has produced more electrical energy in steady operation than it consumes.
As of October 2026
A scientific gain (more fusion energy than laser energy on target) has only been achieved at NIF (factor 4.13 in April 2025), at about 300 MJ of grid electricity per shot. Magnetic devices are still below Q = 1 (JET 1997: 0.67). SPARC by Commonwealth Fusion Systems installed its first of 18 magnets in 2026 and targets first plasma in 2027 and Q > 1 shortly after; ITER's 2024 baseline sets the start of research operation for 2034 and deuterium-tritium experiments for 2039 (+5 billion euros), and by September 2026 six of its nine sector modules sat in the tokamak pit. Helion aims to deliver power from its 50-MW Orion plant to Microsoft in 2028 (company target).
What is missing
- A plant-level gain: more electrical energy out than heating, magnets and cooling consume
- Breeding its own tritium: a lithium blanket that makes more tritium than is burned - so far only tested at lab scale
- Wall materials that survive years of 14-MeV neutrons; there is no test facility with the right neutron flux yet (IFMIF-DONES under construction)
- Steady-state operation instead of second-long pulses, with stable plasma and manageable heat loads
- Economics: a supply chain for high-temperature superconducting magnets, licensing rules, and costs that can compete with solar plus storage
Becomes possible once solved
- Weather-independent low-carbon baseload without long-lived nuclear waste
- High-temperature process heat for industry and hydrogen
- Energy for desalination and carbon removal
- Fusion propulsion for spaceflight
Open steps
- Plasma control and disruption avoidance High AI leverageHold burning plasmas stable for long pulses by predicting and avoiding instabilities that could damage a power plant's walls.
- Operating-scenario design High AI leverageChoose heating, fueling and current profiles for ITER-class and SPARC-class shots that maximize fusion gain within limits.
- Stellarator and magnet design High AI leverageFind compact stellarator shapes and coil sets that confine well, can be built and tolerate manufacturing errors.
- Self-sufficient tritium breeding Medium AI leverageDesign blankets that breed more tritium than the plant burns and let it be extracted, within heat and structural limits.
- Neutron-proof wall materials Medium AI leverageFind structural and plasma-facing materials that keep strength and heat handling after years of 14 MeV neutrons; no matching test source exists yet.
Where AI could help
Medium AI leverage. AI helps plasma control, instability prediction and design search, but net power needs magnets, tritium breeding and neutron-proof materials built and tested.
- Control plasma shape and avoid instabilities in real time with learned controllers
- Run millions of virtual shots in differentiable simulators to choose operating scenarios
- Optimize stellarator and magnet designs and divertor heat loads by search
- Build surrogate models of turbulence and neutron damage to speed up design iteration
Shown so far
- In February 2022 DeepMind and EPFL used deep reinforcement learning to control plasma shape in the TCV tokamak, the first reinforcement-learning feedback control on a tokamak. source
- In February 2024 Princeton and PPPL researchers published in Nature an AI controller that predicted tearing instabilities up to 300 ms ahead and avoided them in DIII-D experiments. source
- In October 2025 Commonwealth Fusion Systems and Google DeepMind announced a partnership using the TORAX simulator, reinforcement learning and AlphaEvolve to plan SPARC plasmas (announcement, no results yet). source
Prerequisites
- Superconductivity1911
- Laser1960
- Tokamak & Stellarator1968
- Fusion: Ignition & Records2022–2026