Graphics processing units are essential to training and deploying artificial intelligence models, but they don’t come cheap. Big Tech companies like Meta, Microsoft and xAI have spent billions, amassing hundreds of thousands or even millions of them.

For those without such deep pockets, access to this kind of computing power has gotten out of reach. Recently, the state of California launched an initiative called CalCompute to look into building its own public GPU cluster for startups and non-profit researchers to use. There are similar public compute pilots in New York state and at the federal level.

While many of the foundational concepts underpinning today’s generative AI boom were first explored in academia, recent advances have come by making models bigger — analyzing more data, adding more parameters and using more compute. Now only a handful of private companies can afford to build and study the most advanced models.

Marc Weber, a director and curator at the Computer History Museum in Mountain View, California, said the history of innovation has seldom been a story only of private companies, but of cross-pollination between academia, government and industry.

“In my mind, I see them all as complimentary,” he said. “It’s kind of an ecosystem where you need all the parts to make it work.”

Russell Wald, executive director of the Stanford Institute for Human-Centered AI, has called for government support to prevent that balance from breaking down.

“There’s a lot of raw potential that sits out there right now to have incredible scientific breakthroughs if we were able to resource appropriately,” he said, noting even the best-resourced universities typically only have access to a few hundred advanced GPUs, not the hundreds of thousands held by Big Tech.

“If it’s only five companies that are just the sole vanguards of truth in this space, you’re not going to get that curiosity-driven type of research,” he said, “the big kind of breakthroughs that aren’t profit-driven.”

He points to examples like Global Positioning Systems, Magnetic Resonance Imaging and the internet, which all came out of publicly funded academic research.

There’s also research that companies might shy away from publishing because it could open industry to regulation or liability, said Stella Biderman, executive director at the non-profit Eleuther AI — like what happened with tobacco companies and lung cancer, or fossil fuels and climate change.

“There’s a really big potential for really bad things to happen and maybe they know these things, maybe they don’t,” she said. “But they definitely don’t share it with the world.”

All commercial frontier AI models are closed to some extent, so you can’t get under the hood to run experiments like, how do you prevent a model from giving users bioweapon instructions? 

To study that question, Eleuther AI built its own models from the ground up.

“I am a very strong believer in independent research that doesn’t require buy-in, assistance or funding from these powerful tech companies,” said Biderman. While her organization runs lean on donated cloud services and grants, public compute resources could make that type of research more accessible at scale.

Hanna Hajishirzi is a computer science professor at the University of Washington who works with the nonprofit Allen Institute for AI to build totally open data sets and models that researchers can experiment with.

“It would be very good if it is not just accessible by a few nonprofit companies or only a few universities,” she said. “Because a lot of these questions could be solved together as a community effort.”

The institute has partnered with the National AI Research Resource, which is piloting a system to provide access to government GPU’s. But even most federal supercomputers aren’t optimized for cutting edge AI. Hajishirzi said we need to build new public infrastructure and likely subsidize commercial providers.

It’s a model that was common at one time, according to Weber at the Computer History Museum.

During the Cold War, he said, government urgency and funding helped launch the computer age. Some researchers say we’ll need that kind of public support again to thrive in the AI age.

Related Topics