Covering Scientific & Technical AI

CAMBRIDGE, Mass., March 12, 2026 — The most powerful supercomputer at Harvard is about to get larger and faster. Much faster.

The Kempner Institute’s AI cluster, already one of the fastest AI supercomputers in the world, is undergoing a major expansion, adding more than 500 NVIDIA graphics processing units (GPUs) — the specialized processors that make modern graphics and AI possible — to scale-up its capacity for cutting-edge research on intelligence.

Once the upgrade is complete in Spring 2026, the enhanced Kempner AI Cluster is expected to join an elite league of systems whose performance is measured in exaFLOPS. One exaFLOP equals a quintillion (that’s a billion billion) mathematical operations per second, which means that in a single minute the cluster will be able to perform a task that would take a personal computer several years to complete.

“There are very few academic institutions on the planet that offer this scale of compute to a research community of our size,” said Kempner Institute Executive Director Elise Porter. “This expansion will allow for research in AI and natural intelligence at Harvard that would not otherwise be possible.”

The expansion will bring 424 of NVIDIA’s top-of-the-line H200 GPUs and 192 RTX PRO 6000 Blackwell GPUs into the institute’s high-performance computing environment, joining an existing array of 144 A100 and 384 H100 units and bringing the total number of GPUs to 1,144.

All the GPUs will be linked in a single, purpose-built system designed to help researchers train, test, and refine large-scale AI models that support work in machine learning, neuroscience, robotics, biomedical research, and a host of other disciplines.

Redefining What Academic AI Labs Can Do

AI clusters, however powerful they may be, can only draw on a fixed amount of computational power at any given moment, and academic researchers often come up against this reality. In many academic settings, there is a tradeoff between allowing researchers to use the cluster’s power for large-scale, computationally-intensive projects on the one hand, and reserving enough bandwidth for the larger community to run important but smaller projects.

With the expanded cluster, Kempner researchers will have enough computing capacity to undertake large-scale and small-scale projects simultaneously.

“Most institutions would have to stop all other work for months to run a large-scale project,” said Porter. “We’re not going to have to stop everyone so that one project can move forward. That’s what this upgrade enables.”

In particular, the new Kempner cluster now has enough GPUs so that a large-scale project can use the almost four hundred H200s in the cluster without interrupting all the smaller-scale projects already running on the cluster’s other GPUs.

“Being able to offer our community nearly four hundred H200 GPUs [for large-scale reservations] is unparalleled,” said Porter. “Very few universities can match that scale for a single project.”

With the expanded compute resources, the Kempner community is poised for a new and exciting chapter in intelligence research, said Porter. “This upgrade gives our researchers a combination of power and flexibility that was simply not possible before,” she said. “We are about to see research projects that redefine what academic AI labs can do.”

Designed with Speed and Flexibility in Mind

The supercharged power of the Kempner’s expanded AI cluster comes down to two factors: a huge influx of hardware that can perform bigger and faster computations than ever before, and a customized design that links the GPUs together in an efficient, integrated network.

“The upgraded Kempner AI cluster now delivers 1.79 exaFLOPS of performance, enabling much larger and faster AI training runs than before,” said Max Shad, the Kempner Institute’s Senior Director of AI/ML Research Engineering. “The heterogeneous GPU network, combining A100s, H100s, H200s, and RTX Blackwell server GPUs, lets us train larger models in a single, unified cluster.”

The mix of different kinds of GPUs also means the cluster can meet the diverse needs of the Kempner community, said Shad, who led the design of the expanded cluster.

“Some of the projects need the power of a Toyota engine for a day, and some need a Maserati engine for a month,” said Shad, likening the different types of GPUs to the power of different car engines. “With the new cluster, we optimized the InfiniBand network for our researchers’ workflows. Some workflows run for days and use large swathes of the GPU network, while others are quick single-GPU experiments that run independently.”

New Hardware with Specialized Capabilities

With the new RTX PRO 6000 Blackwell GPUs, Kempner researchers will now have the computational resources for advanced optical and physics-based simulations, as well as the technology for ray tracing, which simulates the behavior of light rays to generate realistic images and virtual worlds. The RTX chips also accelerate physics-based computations that can be used in neuroscience, including high-resolution simulations of electrical activity in the brain.

The RTX GPUs also have built-in support for working with low-precision numerical formats, which allow AI models to run faster and use less GPU memory by sacrificing a degree of mathematical accuracy. At the Kempner, this low-precision capability will support the institute’s current body of work on “quantization,” which is the process of compressing a model’s size, a line of research with implications for advancing model training and performance.

Alongside the RTX GPUs, the new H200s will provide essential horsepower for researchers working on the most computationally intensive AI models, such as large language models (LLMs), which process text, and “multimodal” AI models that process multiple data types, such as text, audio and video.

Yilun Du, a Kempner Institute Investigator, will use the expanded cluster to continue the development of “world models,” which enable robots to understand and respond to their “worlds” or environments.

“By using the Kempner cluster GPUs, we are able to build powerful foundation models that are typically only possible in large industry labs,” said Du, who is also an assistant professor of computer science at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS). “It allows us to focus on building models with fundamental new capabilities but also to scale them up to industry-level performance.”

The cluster’s expanded computational capacity will also enhance researchers’ ability to work with AI agents, which are AI models that can act independently and solve problems using chains of reasoning steps.

“AI agents are quickly learning to tackle complex tasks — from hard math problems to advanced coding — and a big part of that progress comes from giving them room to reason through ideas,” said Kianté Brantley, Kempner Institute Investigator and assistant professor of computer science at SEAS. “We can finally run large-scale training on models with long reasoning chains, opening the door to studying far more capable AI agents than before.”

About the Kempner

The Kempner Institute seeks to understand the basis of intelligence in natural and artificial systems by recruiting and training future generations of researchers to study intelligence from biological, cognitive, engineering, and computational perspectives. Its bold premise is that the fields of natural and artificial intelligence are intimately interconnected; the next generation of artificial intelligence (AI) will require the same principles that our brains use for fast, flexible natural reasoning, and understanding how our brains compute and reason can be elucidated by theories developed for AI. Join the Kempner mailing list to learn more, and to receive updates and news.

Source: Yohan J. John, Kempner Institute

Related