The AI chip race is forcing supercomputers to look beyond Nvidia and AMD

The AI race is creating a side effect that deserves more board attention: not every critical workload needs the same chips as the frontier-model labs.

Reuters reports that Sandia National Laboratories in the United States is testing chips from Israeli startup NextSilicon in a supercomputer built with Penguin Solutions. Sandia is one of three U.S. national laboratories responsible for developing and maintaining the country’s nuclear weapons arsenal. Its machines run some of the most demanding simulations in government, including hypersonic weapons and the behavior of warheads under extreme conditions.

This is not ordinary AI compute. These systems need high precision, especially what engineers call double-precision floating point. In plain language, they must calculate very large and very small numbers without losing accuracy through rounding errors. That matters in physics simulations. It matters less in much of today’s AI training and inference.

That is why the story is bigger than one U.S. lab. As Nvidia and AMD focus more of their roadmaps and supply chains on AI workloads, customers with different requirements may be squeezed. They may not get bad chips. They may get chips optimized for someone else’s economics.

Reuters writes that Sandia is feeling pressure both from compute demand and from the supply chain. Steve Monk, who manages Sandia’s high-performance computing team, said the outlook is stressful for the lab’s ability to deliver on its mission. That mission is not optional.

Why it matters

For executives, the practical lesson is clear. AI infrastructure is not only about securing as many GPUs as possible. It is about understanding which kind of computation the organization actually needs, which suppliers will prioritize that kind of work, and how exposed the organization becomes when the whole market chases the same AI capacity.

Universities, research institutions, energy companies, defense-adjacent organizations and advanced industrial companies often have different needs from a language-model provider. Simulation, optimization, materials research, weather, energy systems and manufacturing may require precision, memory bandwidth, resilience and operational stability more than pure AI throughput. If procurement becomes a simple question of Nvidia or not Nvidia, the decision is too narrow.

The NextSilicon test also shows how smaller suppliers can gain an opening when the large vendors move their center of gravity. The chips Sandia is testing use a different architecture from classic GPUs and CPUs. According to Reuters, the aim is partly better energy efficiency by spending less power moving data back and forth to memory. The system has now passed a key technical milestone using general supercomputing tests. This fall, Sandia will decide whether to test the chips on more demanding problems that resemble real nuclear-security workloads.

This is often how critical technology changes direction. Not through a glossy transformation campaign, but because a demanding customer finds that standard suppliers no longer cover the whole requirement. Sandia previously helped push liquid cooling for chips from an exotic concept into mainstream data-center practice.

The leadership point

CIOs and boards should treat this as a supplier-risk test. Which compute needs in the organization are not identical to frontier AI lab needs? Which workloads require precision, auditability, local control or predictable access? Which suppliers will still be able to deliver two years from now if the largest AI customers absorb scarce capacity?

For Europe and Norway, this also connects directly to the power and data-center debate. If countries want serious AI and research capacity, counting megawatts and GPUs is not enough. Leaders need to know what kind of compute is being built, whose workloads it serves, what security requirements apply, and whether the infrastructure reduces or increases dependence on a few global vendors.

The Reuters story is therefore not just a niche item about U.S. supercomputers. It shows a new phase of AI infrastructure: once AI demand becomes large enough, it reshapes priorities across the whole chip market. Everyone buying critical compute now has to be more precise.

Sources and media

Primary source: Reuters, Stephen Nellis, "As chip industry chases AI, U.S. national labs look to newcomers for supercomputers", published May 18, 2026. https://www.reuters.com/technology/chip-industry-chases-ai-us-national-labs-look-newcomers-supercomputers-2026-05-18/

Reuters image: NextSilicon chip at Sandia National Laboratories, photographed by Stephen Nellis. Used as reference/link only, not rehosted by hogby.ai.

Thumbnail: OpenAI Image 2 / hogby.ai.