Artificial intelligence has advanced rapidly by learning from abundant digital information—text, images, code, and video gathered at internet scale. Fusion energy presents a very different challenge. In fusion, data can be scarce, expensive to generate, spread across legacy systems, and tied to physical processes evolving in real time. Computational models must deal with nonlinear plasma behaviour and complex engineering systems to predict the behaviour of machines functioning in regimes (operating conditions) never before achieved. Those themes emerged at the ITER Public-Private Fusion Workshop, held on 28 and 29 April 2026, where four speakers—Simon McIntosh of ITER, Antonio Policicchio of NTT Data Italy, Tom Gibbs of NVIDIA, and Thomas Kopinski of Gaia Lab—outlined how artificial intelligence is beginning to support construction, maintenance, simulation, and future operations at fusion projects.Their broader message is that fusion is not simply another use case for AI. In some respects, fusion may be one of the forces pushing AI to evolve.When data is difficultMany modern AI systems benefit from the fact that digital data is plentiful. “For large language models and agents, the data is cheap, but the training is very expensive,” Gibbs said in discussion during the workshop. ”For physical AI (applications that learn from real-world systems rather than digital content) the data is expensive, but the training is not.”That distinction is significant. A plasma pulse on a tokamak is not equivalent to collecting another billion web pages. With fusion, experimental campaigns are costly, machine access is limited, and each shot may explore only a narrow operating window. Useful data must often be extracted carefully from relatively small (at least compared to internet scale) but highly valuable datasets.According to McIntosh, fusion organizations are also facing another challenge: decades of historic data stored in formats never designed for modern AI workflows. “Essentially, we have to do data archaeology to use what was produced before,” he says.He cites archives from JET, which operated for more than 40 years and changed configuration many times over its lifetime. Valuable experimental records, software tools, and engineering context still exist—but not always in forms that are easy to search, compare, or reuse. Using AI agents, McIntosh says, teams have begun tracing files, identifying patterns, and reconstructing historical operational knowledge from past and currently operating tokamaks that might otherwise remain buried.Recovering existing data is only one of the tasks. Others are the standardization of the data format and verification of the data itself. According to McIntosh, the fusion community has spent close to 20 years developing common data formats so software built for one machine can be tested and reused on another. “It’s called the IMAS Data Dictionary,” he says, describing a shared framework, developed at ITER, that allows information such as coil currents or diagnostic signals to be stored in consistent ways across facilities. Data verification is a parallel task that needs to be addressed with some urgency as the scientists responsible for these extremely valuable datasets reach the end of their careers. That may sound administrative, but it is strategically important. Standardized data enables machine learning models, control tools, and analysis software to move more easily across tokamaks, accelerating development across the sector.But organized data still has to be accessible. In many industrial settings, the first AI challenge is not intelligence but access. McIntosh points to Lucy, an AI-enabled assistant connected to ITER’s internal document systems. It is already helping staff retrieve engineering records, technical notes, and project documentation in seconds rather than through lengthy manual searches.Perhaps the most ambitious concept discussed by McIntosh is ITER’s effort to prepare for operations by training what he calls a fusion “world model.” “We have a GPU cluster that we’re planning to use to train what’s known as a fusion world model,” he says.In practical terms, such a model would learn from data gathered across existing tokamaks and create a simplified but useful representation of machine behaviour. It could allow teams to simulate pulses before they are run, test procedures, anticipate faults, and improve readiness before ITER begins full experimental campaigns. McIntosh compares the concept to the intuitive model humans carry in their own minds—knowing what the next step on the stairs will feel like before the foot lands. For ITER, the goal would be similar: predict the next state of the machine before it happens.Such world models will transfer decades of knowledge gained on past and existing tokamaks to ITER in such a way that data collected on the first day of ITER operations will be immediately available in the form of improved second-day predictions. “This pattern of continual recalibration and retraining will ensure that the capabilities of our data-grounded machine-learning models will improve in lockstep with the knowledge frontier that ITER will push into yet unseen reactor-relevant conditions,” says McIntosh. Simon Macintosh stands in front of the eight newly installed GPUs that arrived in March to help process ITER's AI needs. Adapting AI to plasma physicsGibbs described another route to acceleration during the workshop at ITER: AI surrogate models trained on conventional plasma simulations. Traditional high-fidelity codes can consume hundreds or thousands of GPU hours. Once trained, surrogate models can approximate those results in milliseconds, creating possibilities for faster analysis and eventually real-time decision support.Asked how much faster such systems can be, Gibbs estimates gains ranging from 10³ to 10⁶ times. That speed could support digital twins—virtual representations of fusion devices continuously updated with sensor data and capable of helping operators understand evolving plasma conditions.According to Gibbs, digital twins already exist for several experimental machines, and there is no fundamental reason an ITER twin could not begin development before operations start. But if AI models become central to real-time fusion control, some believe faster models alone may not be enough. The underlying hardware—and software to connect the live data with the models—may also need to evolve.While GPUs dominate today’s AI landscape, Kopinski used his workshop presentation to argue that fusion will require alternative computing architectures. Plasma systems are highly dynamic and nonlinear, he says, and some control problems demand responses faster than conventional approaches can comfortably deliver.“They were talking about milliseconds; we are talking about microseconds,” Kopinski says, referring to comparisons with GPU-based systems. According to Kopinski, Gaia Lab is developing a fusion world model based on Q.ANT's photonic hardware, using their light-based processors to predict and control plasma behaviour “faster than real time.” He says that prototypes from Q.ANT are deployed today, and a commercially viable world model may be available before the end of this decade.Such timelines remain early-stage, but the underlying point is broader: fusion may become a proving ground not only for new energy systems, but for new forms of computing.Practical gains nowEven with longer-term ambitions, speakers repeatedly emphasized that some of AI’s greatest value may come much sooner. Because ITER is executing assembly and commissioning, McIntosh says construction support is a near-term priority.“Help with construction is something we can use today,” he says. “If we can use AI to inform better decision making that is grounded in our data, we can possibly accelerate the construction schedule, de-risk items, and address problems more quickly.”Policicchio, who presented practical AI applications at the workshop, describes projects involving LiDAR-based progress tracking and predictive maintenance for equipment such as pumps and cooling systems. “I believe it will be one of the key enablers for reducing schedule slippage in this large project,” he says.For all the momentum around AI, ITER’s purpose remains experimental science. It is intended to test operating regimes, pulse durations, and integrated performance levels never before demonstrated at this scale. Those results must be measured in reality, not inferred from software alone. While AI cannot replace an experiment of a scale that has never been attempted before, it can shorten the path to the resulting discoveries.ITER is not merely adopting off-the-shelf artificial intelligence tools. It is confronting some of the hardest questions in industrial AI: how to learn from scarce data, unify fragmented archives, predict complex physical systems, and act in real time when mistakes are costly.That may make fusion one of AI’s toughest applications—but also one of the places where its future capabilities are being forged.

With the delicate touch of a butterfly settling on a branch, engineers have successfully transferred the 440-tonne load of vacuum vessel sector #6 from its overhead radial beam to its gravity support below. The first-of-its-kind operation—an essential benchmark in vacuum vessel assembly—means the vacuum vessel sector has been decoupled from the central column and bioshield wall that had previously shouldered its weight. It is now fully supported from underneath by the vacuum vessel gravity support that sits on the cryostat base and is buttressed from below by the tokamak’s crown support structure. Stability clamps are also holding the sector steady until all nine vacuum vessel sector modules are installed in the tokamak pit and torus welding has progressed sufficiently.The load transfer process required manoeuvring the enormous vacuum vessel sector into position over the gravity support with only three millimetres of margin. Then, the hinge mechanism of the support was used to raise it into contact with the lower port stub extension of the sector, and shims were inserted to ensure a perfectly secure interface between the components. Finally, the sector was bolted to the gravity support and its weight was released from the overhead radial beam on Tuesday 5 May.  The transfer of the sector loads to the vacuum vessel gravity supports (VVGS) has to be accelerated due to the new welding strategy. “The preparations and alignment had been executed perfectly so there was no movement during the load transfer,” says Vincent Hanser, the contract manager for the project. “The real achievement is reaching this stage, as the ITER teams and contractors had to invest a huge amount of effort over two years to obtain this result.”Originally, the vacuum vessel sectors were to be lowered onto their dedicated supports only after being welded into “triplets.” However, when the assembly strategy evolved toward simultaneously welding all nine sectors after they had landed on their gravity supports, the landing process had to be moved forward. The transfer of the load was monitored by the SIMIC team from a temporary control room set up in a tokamak port cell. “The change in welding strategy was important for achieving homogenous shrinkage across the nine sector joints and mitigating deformation, but it did require a significant adaptation to the assembly schedule,” says Sébastien Koczorowski, ITER’s Deputy Program Manager for the Machine Assembly Program. “The successful landing highlights the agility, resilience, and collective problem-solving capacity of the teams.”The installation of the vacuum vessel gravity support and the stability clamps, as well as the landing itself, were executed by ITER contractor SIMIC. With the successful touchdown of the first sector, teams can now turn their sights to landing vacuum vessel sector #7 in July.

The European Domestic Agency Fusion for Energy delivered sector #3 of the ITER vacuum vessel on Thursday 7 May. Only one more sector is needed to build the ITER plasma chamber. The arrival of vacuum vessel sector #3 marks another major milestone for the ITER assembly program. Manufactured in Europe by the AMW consortium—Ansaldo Nucleare, Westinghouse, and Walter Tosto—the 440-tonne component reached the ITER site last week after a multi-stage journey from Italy by sea, barge, and road. Europe is responsible for five of the nine sectors that will form the torus-shaped vacuum vessel at the heart of the ITER tokamak.Sector #3 is the fourth supplied by Europe and the eighth vacuum vessel sector delivered to ITER overall. Once inspection and preparation activities are completed, teams will integrate the sector with its thermal shield and a pair of toroidal field coils to create a fully equipped sector module for installation in the tokamak pit. The Procurement Arrangement for the sectors was signed between the ITER Organization and Fusion for Energy in 2009. With only one sector now outstanding, the conclusion of a nearly 20-year industrial adventure is coming into view.The AMW consortium worked with more than 15 European companies and directly mobilized 150 professionals to manufacture the sectors. From the machining of raw forgings to the assembly, welding and final machining of completed sectors, the process required hundreds of intermediate tasks and interaction between teams across Europe.The final sector, vacuum vessel sector #2, is scheduled for delivery this autumn. 

As part of an ongoing initiative to encourage young talent to consider a future in fusion, the ITER Organization participated in the 17th edition of the Olympiades des Sciences de l’Ingénieur, a student engineering competition organized by the public education service in the Aix-Marseille region and the National Academy of Technologies on 6 May. The annual national competition gives French high school students the opportunity to develop innovative projects in engineering sciences. This year’s regional edition saw record participation, with 66 teams from 15 schools in the Aix-Marseille school district and 300 students presenting projects on the themes of "Engineering in the Service of Art" and "The City of Tomorrow."The jury included ITER’s Fabien Lassueur, Section Leader in the Engineering Services Department, and DeLeah Lockridge, Deputy Director-General and Head of the Engineering Services Department. ITER also sponsored the prize for the best project presented in English, awarded to an innovative proposal for cleaning monuments using drones.“It was great to see the level of participation and engagement from the students,” said Lockridge. “Their enthusiasm and passion were refreshing, because this is our future. I see the event as a real success, and I hope ITER participates in future competitions.”After the presentations, students took part in round-table sessions where they could learn more about ITER and ask questions. The top three teams from the regional competition have been invited to the national finals, which will take place in Paris on 28 May. The high school students were also invited to learn more about ITER during informal meetings with senior project leaders such as DeLeah Lockridge.