MIT researchers use cooling platform for AI nuclear work

MIT researchers are using Ferveret's cooling-based computing platform to study nuclear materials with AI. The work is testing whether more computing output can be achieved within the same power limits.

The project is based at MIT's Materials in Extreme Environments Lab, where scientists are applying AI to materials for next-generation nuclear energy systems. Ferveret says its platform lets researchers train larger models, run experiments continuously and increase usable compute within a fixed power envelope.

Its Adaptive Phase Cooling system can deliver up to 15% more server-level compute than direct-to-chip liquid cooling methods currently on the market, according to the company. It says that, combined with facility-level efficiency gains, operators can achieve up to 35% more usable compute from the same power supply.

The research focuses on workloads that place sustained demands on graphics processing units, mirroring the way many commercial AI systems are run. Ferveret is using the MIT deployment to show how data centre cooling design can affect the practical output of AI infrastructure, particularly when compute is limited by power and thermal constraints rather than chip supply alone.

Reza Azizian, Co-Founder and Chief Executive Officer of Ferveret, outlined the company's approach to measuring performance. "Ferveret's philosophy for evaluating high-performance AI infrastructure is simple: rather than relying solely on synthetic benchmarks or artificial stress tests, the company believes real scientific workloads provide a more meaningful measure of system performance while creating tangible value for researchers and others who need the compute," he said.

At MIT, one line of work supported by the platform involves analysing coherent X-ray images collected from nuclear materials. Researchers said existing algorithms often struggle to interpret the complexity of advanced reactor materials and that AI-based methods are being developed to improve the accuracy and reliability of image analysis.

Ericmoore Jossou, Assistant Professor of Nuclear Science and Engineering and Electrical Engineering and Computer Science, and Head of the Materials in Extreme Environments Lab, said the additional computing resources have widened the scope of the team's research.

"Access to Ferveret's platform enables us to pursue research that simply would not be feasible with the computational resources typically available in academia. The ability to train larger models, iterate faster and run experiments continuously is accelerating our pace of discovery. Today, compute is becoming as fundamental to scientific discovery as the laboratory itself. Advances in AI will increasingly depend not only on better algorithms, but on access to the infrastructure that makes them possible," he said.

A second area of research involves machine learning interatomic potentials, which are used to predict how materials behave at the atomic scale under harsh operating conditions. These models require extensive GPU resources during training, but can enable simulations that approach the accuracy of density functional theory, widely regarded as one of the most compute-intensive tools in materials science.

Scientific workloads

Ferveret argues that this kind of academic research is a useful test bed because the computational patterns closely resemble those seen in production AI environments. That gives the company a way to demonstrate the effect of thermal management on sustained GPU performance under real workloads rather than short benchmark runs.

"Our AI workloads place sustained demands on GPU infrastructure that closely resemble the challenges found in production AI environments. That makes our research an ideal proving ground for Ferveret's technology. At the same time, the additional compute available through Ferveret's platform allows us to explore more ambitious models, complete research cycles faster and accelerate scientific discovery," said DongHun Ryu, Graduate Researcher, Materials in Extreme Environments Lab, MIT.

Jossou also described the role of generative AI in the project's materials work. "We are leveraging generative AI to solve long-standing challenges in understanding materials for next-generation nuclear energy. Our AI models are significantly more accurate than conventional approaches, but they also push the limits of today's computing infrastructure," he said.

Infrastructure limits

The announcement reflects a broader shift in the AI sector, as operators focus more on limits imposed by electricity, cooling and facility design. As models grow and training runs lengthen, available power in a data centre can become a more immediate bottleneck than access to processors.

Ferveret says its cooling system is based on subcooled boiling, a technique used in nuclear reactor systems. According to the company, the approach improves heat transfer efficiency while operating near ambient pressure, allowing GPUs to sustain higher performance with lower thermal overhead and reduced water use.

Founded by MIT alumni, the company is based in San Jose, California. It says research publications produced from work carried out on its platform are expected to acknowledge its contribution to the computing infrastructure behind the projects.

Azizian linked the MIT work to the wider economics of AI infrastructure. "The next era of AI will be constrained less by chips than by infrastructure. Every watt we save becomes another watt available for AI. Seeing MIT researchers use that additional compute to accelerate discoveries in advanced nuclear materials demonstrates exactly why better cooling matters. The future of AI depends not only on more powerful processors, but on infrastructure that allows them to operate at their full potential," he said.