Deep in a nondescript building on Kirtland Air Force Base in New Mexico, liquid-cooled supercomputers work through some of the most complex calculations the U.S. government faces. These machines simulate hypersonic nuclear weapons moving through Earth’s atmosphere and model what happens when nuclear warheads detonate near each other.
For over a decade, mainstream semiconductor companies like Nvidia and AMD supplied the chips powering this secretive work. But as these firms increasingly design processors for artificial intelligence and face supply shortages, Sandia National Laboratories – one of three U.S. labs maintaining the nation’s nuclear weapons arsenal – is struggling to find computing power for high-precision scientific work.
“The pressure we’re feeling right now is on the computing front and also from the supply chain,” said Steve Monk, manager of Sandia’s high-performance computing team. “Looking to the future, it’s a bit stressful in terms of our ability to deliver to the mission.”
This challenge shows how the AI chip race is opening markets once dominated by major firms to smaller players. Sandia is now testing chips from NextSilicon, an Israeli startup, as traditional suppliers shift their focus away from scientific computing toward more profitable AI applications.
The situation highlights a growing divide in the chip industry. While AI work has become incredibly lucrative, it requires different computing capabilities than scientific research. Nuclear weapons simulations need what’s called double-precision floating point computation – the ability to calculate very large and very small numbers without losing accuracy to rounding errors.
For years, Nvidia and AMD competed to speed up this type of computing, winning supercomputing contracts with universities and government labs. But AI work doesn’t benefit from double-precision computing the same way physics simulations do. While AMD is releasing chips aimed at scientific computing, the double-precision performance of Nvidia’s upcoming Rubin chips has declined by some measures, worrying scientists in the high-performance computing industry.
Daniel Ernst, senior director of supercomputing products at Nvidia, said the company remains committed to scientific computing and aims to create balanced chips that can run both scientific applications and AI work. However, the shifting market has prompted Sandia officials to explore alternatives.
On Monday, Sandia, NextSilicon, and integration partner Penguin Solutions announced that NextSilicon’s systems passed a key technical milestone using supercomputing tests that put the chips in the running for government use. This success sets up a decision this fall on whether to test the chips with more demanding problems that closely resemble actual nuclear security work.
NextSilicon’s approach differs completely from the graphics processing units and central processing units used by Nvidia and AMD. The chips can perform double-precision computing and reprogram themselves on the fly to run more efficiently. They save electricity using a data flow architecture that spends less time and energy moving data back and forth to system memory.
Sandia’s work with chip companies often helps new technologies become widespread. The lab started pushing Intel, AMD, and Nvidia to develop liquid cooling systems over a decade ago when the technology was exotic. Now liquid cooling is common in high-performance computing.
James Laros, a senior scientist at Sandia who oversees the program testing new computing architectures, said the work with smaller players like NextSilicon aims to ensure Sandia can always get the chips it needs, even if major firms change focus.
“We have to keep available options to complete our mission, because the mission is not optional,” Laros said.
This shift reflects broader changes in the semiconductor industry. As AI applications generate massive profits, chip companies are prioritizing these markets over specialized scientific computing needs. For critical government work like nuclear weapons research, this creates supply chain risks that national laboratories must address through partnerships with emerging chip companies.
The collaboration also demonstrates how government research facilities serve as testing grounds for new technologies. Sandia’s validation of NextSilicon’s chips could help the startup gain credibility with other customers and potentially challenge established players in specialized computing markets.
