Harvard's Kempner Institute Faces Its First Test
Harvard's Kempner Institute bought computing hardware before the market exploded. That decision, once questioned by its own advisors, may have been the most consequential choice the institute made.
Founded in December 2021 with a $500 million gift from Meta CEO Mark Zuckerberg and Priscilla Chan, Kempner secured computing resources that remain scarce across academia. The timing proved crucial. Nearly a year before ChatGPT's public release, the institute began building what is now one of academia's largest AI clusters - a resource that has become central to its recruiting pitch and research capability.
Karel Svoboda, a former member of Kempner's Scientific Advisory Board, initially opposed the decision to own hardware rather than rent from cloud providers. "My opinion was that it may have been a bad idea," he said. "I was wrong."
The Case for Owning Hardware
Elise Porter, Kempner's executive director, compared the computational cluster to an electron microscope. Both are expensive, specialized tools necessary for the work.
Cloud computing would have allowed flexibility. But Kempner's researchers needed constant access. "If you're using it literally 24 hours a day, 365 days a year, the math doesn't work," Porter said about renting.
Kempner started with 144 A100 graphics processing units in January 2023. By March 2024, it had added 384 H100 GPUs. The institute is now adding another 384 H200 GPUs and 192 RTX 6000 Pro GPUs, bringing its total to more than 1,100 data-center GPUs.
The cluster is now Kempner's clearest recruiting advantage. Bingbin Liu, a postdoctoral fellow working on machine learning, said the hardware was one reason she chose Harvard. "Nowhere - at least in the U.S. - nowhere I know has as much compute as Kempner does," she said.
More than 6,000 Harvard researchers can access Kempner GPUs when institute researchers are not using them, though Kempner affiliates receive priority.
A Broader Mission Under Pressure
Kempner's full name - the Kempner Institute for the Study of Natural and Artificial Intelligence - reflects its founding ambition. The institute was designed to study intelligence in biological systems and artificial models, bringing together computer scientists, neuroscientists, cognitive scientists, and psychologists.
Venketesh Murthy, a professor of molecular and cellular biology and member of Kempner's faculty steering committee, said that cross-disciplinary exposure matters even when neuroscience experiments don't directly support AI work. "We gain a lot by the ideas and in the interactions," he said.
Some researchers use AI tools to analyze brain data. Others use ideas from neuroscience or physics to understand why artificial neural networks function. Thomas Anderson Keller, a research fellow, studies recurrent neural networks using concepts from physical dynamical systems. "I could definitely make more money in a company, but to work on some of the crazy stuff that I work on like waves and taking ideas from physics, it would be hard to convince someone to pay me to do that," he said.
But large language models have shifted priorities. Murthy acknowledged the change. "The language models have just, they just seem to overwhelm everything right now," he said.
The speed mismatch is real. AI research can cycle through testing and improvement in weeks. Neuroscience experiments often take far longer. Elias Issa, a neuroscience professor at Columbia University, described the divide bluntly. "AI is kind of running away, moving too fast from neuroscience," he said.
Where Researchers Go
Kempner's early fellows and post-baccalaureate researchers are beginning to leave Harvard. Some are moving to MIT, Johns Hopkins, and the University of Michigan. Others are heading to Meta, Apple, and startups.
The institute's fellowship program has become its most important tool for building community. Kempner research fellows receive two-year postdoctoral appointments and can work across multiple Harvard labs rather than being tied to a single faculty member's project. Svoboda said the fellows have become "the glue that holds the community together."
Kempner pays research fellows roughly $100,000 annually, compared with about $70,000 for standard Harvard postdoctoral salaries. But money isn't the only draw. Liu cited Kempner's open-science requirement as a key reason she stayed in academia. In industry, many labs are closed. "Whatever you do, you have to publish it, and then it has to be shared with the broader science community," she said of Kempner.
That commitment has limits. Industry labs pay more, build faster, and provide access to far larger systems. For researchers who want to deploy models at scale, the gap can be decisive. Itamar Kahn, a neuroscience professor at Columbia's Zuckerman Institute, said some ideas quickly outpace academic resources. "If you have an idea that will scale really fast - that within a year, you need resources that are beyond what is available in academia - then you're gonna leave," he said.
Building a Reputation
Kempner's standing outside Harvard remains uneven. When researchers at peer institutions were asked about the institute, some knew it well. Others hadn't heard of it. John Mackey, a computer science professor at Carnegie Mellon University, said he was unfamiliar with Kempner and surprised by the $500 million gift. "Somebody on the PR end of things needs to get cracking," he said.
The institute is still small, with 120 to 160 people now and expected to reach 300 to 500 within two years. Porter said the institute's youth is an advantage. "We don't have any history. We don't have that because we are so new, and it allows us really to be innovative and nimble," she said.
Kempner's first real test lies ahead: proving that academic AI research, even with substantial resources, can compete with the speed and scale of industry labs while maintaining its commitment to open science and cross-disciplinary work.
Your membership also unlocks:
