NVIDIA and the U.S. Department of Energy are building a 100,000-GPU AI supercomputer called Solstice at Argonne National Laboratory, the centerpiece of the DOE's Genesis Mission to put frontier AI behind American science. Solstice will run on NVIDIA Vera Rubin GPUs and deliver 5,000 exaflops dedicated to science — five times larger than the entire TOP500 supercomputer list combined. A second system, Equinox, is already being stood up with 10,000 NVIDIA Grace Blackwell GPUs. The work spans 17 national labs and 20 years of NVIDIA partnership with them.
U.S. Energy Secretary Chris Wright and NVIDIA Vice President of Hyperscale and High-Performance Computing Ian Buck laid out the case Thursday morning at the SCSP AI+ Expo in a 30-minute fireside chat moderated by SCSP president Ylli Bajraktari. Their pitch: American leadership in AI runs through American leadership in energy, and the same GPUs training frontier models will now be pointed at fusion, materials and the electricity grid itself.
"NVIDIA is 100% committed and invested in Genesis," Buck said. "I've never seen more excitement across the lab and industry." Buck stressed that the labs are getting the same hardware and software stack the major AI labs use commercially — not a science-only side branch.
Key facts
- 01NVIDIA and the DOE are building Solstice at Argonne with 100,000 Vera Rubin GPUs, delivering 5,000 exaflops for science.
- 02A first system, Equinox, is being stood up now with 10,000 Grace Blackwell GPUs.
- 03Blackwell delivers 30x the performance of Hopper and 25x the performance per watt, per NVIDIA's Ian Buck.
- 04The DOE will bring three small modular reactors critical by July 4, 2026 to feed grid demand.
- 05An open-source NVIDIA model was trained on 1.5 million physics papers and fine-tuned on 100,000 fusion papers for DOE researchers.
The scale claim is the headline. "To put that 100,000 in perspective on the next-generation GPU, which is dedicated to science, it's 5,000 exaflops," Buck said. "That's a big number that actually is five times larger than the entire TOP500 supercomputer list combined." Equinox, the smaller of the two, is using the same Grace Blackwell silicon shipping into commercial AI build-outs today.
“Solstice's 100,000 next-generation GPUs will deliver 5,000 exaflops dedicated to science — five times larger than the entire TOP500 supercomputer list combined.”— Jaeden Schafer
Buck offered a concrete example of what the systems will run. NVIDIA has trained an open-source model on 1.5 million physics papers and fine-tuned it on 100,000 papers specifically about fusion, producing a specialized agent DOE researchers can interrogate to move faster on reactor design. That kind of domain-narrow model is the template Genesis is meant to industrialize across the labs.
Wright framed the energy side bluntly. Over the last 20 years, the U.S. has tripled oil production and doubled natural gas production but barely grown electricity production — and electricity is what AI runs on. His department is leaning into natural gas, nuclear and coal simultaneously, with three small modular reactors set to go critical by July 4, 2026 and both new large reactors and additional SMRs behind them.
"We have to fix this bureaucratic and complex electricity grid so that it can grow fast, so that it can grow like our primary energy production and it can keep up with AI," Wright said. "If we don't do that, we're going to slow down AI." Grid interconnection studies that currently take years are one of the first AI targets — Wright wants them down to weeks or hours.
On the chip side, Buck pointed to per-watt efficiency as the offset. "We went from the Hopper generation to Blackwell," he said. "We increased performance by 30x. We actually increased performance per watt by 25 times." Those gains are the basis for the argument that more data centers, paired with more generation, lower electricity costs rather than raise them.
Wright set a 12-month bar for Genesis: concrete deliverables in fusion, materials and grid interconnection. "We will have deliverables that we're going to point to — we couldn't do that before, and now we can," he said. "That's the goal of Genesis: drive discovery and bring the benefits to humans." The DOE has also stood up a strategic fusion office to coordinate the work.
The skeptical read is that 5,000 exaflops on a slide is not 5,000 exaflops in a building. Solstice depends on Vera Rubin shipping on schedule, on Argonne's site power and cooling keeping pace, and on the grid build-out Wright described actually clearing permitting fast enough to feed it. Public concern that AI data centers will push up household electricity bills is also real, and the political answer — build more generation faster — is harder than the engineering one.
Genesis is the clearest signal yet that the U.S. government wants frontier compute treated as national infrastructure rather than a private-lab perk. If Solstice lands on time, it gives the national labs access to the same class of system that OpenAI, Anthropic and xAI are racing to build for themselves, but pointed at problems with no obvious commercial buyer. That's the part of the AI build-out that doesn't get funded by ad revenue or API pricing, and it's the part that decides whether the next decade's scientific breakthroughs are American.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
