Industries / Frontier science / Fusion energy
Power plant design
This is the work of turning a fusion machine into a power station: designing the whole plant around the machine, down to the turbine and the grid connection, and planning how it will be built, maintained and paid for.
6
Milestones achieved, from 1973 to 2023.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1973
Whole-plant paper designs
In November 1973 a team of 30 authors at the University of Wisconsin, among them Robert Conn and Gerald Kulcinski, issued UWMAK-I, a design study of a tokamak fusion reactor for making power, revised in March 1974 as report UWFDM-68. A tokamak is a doughnut-shaped machine that holds hot fuel in place with magnetic fields. The study worked out a reactor on paper decades before any fusion machine could make power, the same kind of exercise as the ARC design study below.
Credit
B. Badger, M. A. Abdou, Robert Conn, Gerald Kulcinski and colleagues, University of Wisconsin
1980
A plant written into law
The US Magnetic Fusion Energy Engineering Act of 7 October 1980 set national goals of proving the engineering feasibility of magnetic fusion by the early 1990s and of running a magnetic fusion demonstration plant at the turn of the twenty-first century, and stated that this would need the program's funding to double within seven years. No such plant was built: in April 2026 Commonwealth Fusion Systems described its request to connect a fusion plant to PJM, the largest US wholesale power market, as the first from any fusion plant developer to a major grid operator.
LessonA target date written into law did not by itself produce a plant.
Credit
US Congress (Public Law 96-386)
1998
Maintenance done by robot
After its 1997 deuterium-tritium experiments, the Joint European Torus (JET) near Oxford replaced its whole divertor, the part of the vessel that takes the exhaust heat, by remote handling alone. Operators in a control room drove a two-armed manipulator called Mascot, carried on a 10 m articulated boom, to remove 144 divertor carriers and fit 192 new ones in a shutdown planned for four months; JET's report says every planned task was done on time, along with unplanned ones such as removing damaged tiles. It was the first use of JET's remote handling system.
Credit
M. A. Pick and the JET Team, JET Joint Undertaking
2015
A compact plant design
An MIT team published ARC, a design for a tokamak power plant of 200 to 250 megawatts of electricity (MWe) with a 3.3 m major radius and a 9.2 tesla magnetic field. Its main magnets use REBCO (rare-earth barium copper oxide) superconducting tape and have joints, so the magnets can be opened and the vacuum vessel swapped out quickly. The paper was peer reviewed and published in Fusion Engineering and Design; Commonwealth Fusion Systems, founded later, now plans a plant by the same name.
Credit
B. N. Sorbom, D. G. Whyte and colleagues, MIT
2022
STEP gets a site
The UK government chose the West Burton power station site in North Nottinghamshire for STEP (Spherical Tokamak for Energy Production), a prototype plant meant to show that fusion can put net electricity on the grid. In March 2026 the delivery company, UK Industrial Fusion Solutions, named the ILIOS consortium, led by a Kier and Nuvia joint venture, as construction partner on a contract worth £200 million.
Credit
UK government; UK Atomic Energy Authority; UK Industrial Fusion Solutions
2023
A buyer for fusion power
Helion, a fusion company in Everett, Washington, agreed with Microsoft in 2023 that Microsoft would take the electricity from Helion's first power plant. The American Nuclear Society's news service reported in 2026 that the plant, Orion, in Malaga, Washington, has been under construction since July 2025.
Credit
Helion; Microsoft
Watch for
Orion online by 2028
Helion says Orion will be online by 2028 and will produce 50 MW or more after a year of ramping up to full power, with Microsoft as the buyer; it is working with the Chelan County Public Utility District on a grid connection agreement.
Achieved whenOrion sends electricity to the grid under the Microsoft agreement, confirmed by Microsoft or the utility, with the plant's net output published.
Claimed by
Helion, June 2026
ARC on the grid
Commonwealth Fusion Systems (CFS) says its first ARC plant in Virginia will deliver 400 MW of net electricity continuously in the early 2030s. It applied to connect to the PJM grid in April 2026, has signed agreements to sell power to Google and Eni, and in June 2026 published five peer-reviewed papers on ARC's physics in the Journal of Plasma Physics.
Achieved whenPJM's grid studies clear the connection and ARC sends net electricity to the PJM grid.
Claimed by
CFS, April and June 2026
Open
01
Every plant, one code
Write the published design points of at least six fusion plant designs, such as ARC, EU DEMO, STEP and a stellarator plant, as input files for PROCESS, the UK Atomic Energy Authority's open-source systems code, and publish a table that sets each design's own figures for gross power, recirculating power (the electricity the plant uses to run itself) and net output beside what PROCESS computes. A systems code is a program that balances a plant's main parts, from plasma and magnets to the turbine, in one calculation. Rerun three of the designs in a second open code, D0FUS.
The briefCredit · unclaimed02
Blanket swap in simulation
An open simulation task in which a robot lifts a heavy blanket module, a wall panel that captures the machine's heat and neutrons, out through a narrow port and seats a replacement within a set tolerance, from randomized starting positions, with fixed scoring rules and a public results table. Then score a conventional motion-planned controller and at least one robot model trained from demonstrations.
The briefCredit · unclaimed
Open 01 · the brief
Every plant, one code
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
Write the published design points of at least six fusion plant designs, such as ARC, EU DEMO, STEP and a stellarator plant, as input files for PROCESS, the UK Atomic Energy Authority's open-source systems code, and publish a table that sets each design's own figures for gross power, recirculating power (the electricity the plant uses to run itself) and net output beside what PROCESS computes. A systems code is a program that balances a plant's main parts, from plasma and magnets to the turbine, in one calculation. Rerun three of the designs in a second open code, D0FUS.
Why it is still open
PROCESS has been public on GitHub under the MIT license since 2023, but the inputs it ships with describe generic machines (a large tokamak, a spherical tokamak, a Helias stellarator, a laser plant), not named commercial designs. The published benchmark found compared PROCESS with Japan's TPC code on two DEMO concepts in 2012 and traced their differences to impurity radiation and magnetic flux models. A November 2026 search found no comparison of named commercial plant designs in one code.
Built with what exists
PROCESS (Python); D0FUS, a Python systems code published on GitHub in 2026 by the IRFM group; the published design papers; an AI coding agent such as Claude Code to draft input files from the papers, with a person checking every value against its source (the Note "Three working devices" covers how far such agents carry a design task).
First step
Run one of PROCESS's shipped tokamak inputs, then write an ARC input from the 2015 paper's design table and compare the net electric output with the paper's.
Done looks like
Six designs run, each input value linked to the page it came from; for each, the net electric output from PROCESS within 15% of the published figure, or the gap traced to a named model difference; three designs reproduced in D0FUS by a second person.
Open 02 · the brief
Blanket swap in simulation
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open simulation task in which a robot lifts a heavy blanket module, a wall panel that captures the machine's heat and neutrons, out through a narrow port and seats a replacement within a set tolerance, from randomized starting positions, with fixed scoring rules and a public results table. Then score a conventional motion-planned controller and at least one robot model trained from demonstrations.
Why it is still open
The UK Atomic Energy Authority's robotics group lists handling difficult payloads such as blanket modules, and remotely joining pipes, among the key unsolved tasks of fusion maintenance, and says future work will explore machine learning. VTT's divertor test rig for ITER in Tampere moves loads of about 10,000 kg with half-millimeter precision, and in 2025 VTT and UKAEA researchers used reinforcement learning in the Unity game engine to plan the path of a blanket segment out of a vacuum-vessel sector, as a design aid that replaces step-by-step path planning in CAD. A November 2026 search found that and digital twins of single maintenance robots, but no public benchmark or demonstration dataset for fusion remote handling.
Built with what exists
MuJoCo or NVIDIA Isaac Sim; simplified public geometry (a box-shaped module and port, not a real design drawing); an open-weight robot model such as Runway's Praxis-1 (see the Note "Open robot model weights"); few-demonstration training of the kind NVIDIA reports for GR00T 2 (see the Note "Five demonstrations per task"); scoring in the style of public robot benchmarks such as RoboTwin (see the Note "First on three benchmarks").
First step
Model a module of a few tonnes and a port with 20 mm clearance in MuJoCo, script a motion-planned baseline, and record its success rate over 50 runs.
Done looks like
200 simulated attempts per controller from starting offsets of up to 20 mm and 2 degrees, each counted a success only if the module seats within 0.5 mm without exceeding a set contact force; the task, the scoring code and the tally published, and rerun by a second group.