At Lawrence Livermore National Laboratory, there is a structure that was intended to last for fifty years. 48,000 square feet of computer space, a domed roof, and flying buttresses. It was known as the Terascale Simulation Facility, and those who have stood inside claim that it was constructed to resemble a cathedral. It was already having trouble keeping up less than ten years after it opened. I’ve always found that particular detail to be the most accurate metaphor for the current state of supercomputing. Everything designed to house them is outgrown by the machines.
The Department of Energy is concurrently commissioning nine new supercomputers at three national laboratories: Oak Ridge in Tennessee, Los Alamos in New Mexico, and Argonne in Illinois. not improvements. not small adjustments. completely new systems, some of which are specifically made to sit at the nexus of national security and artificial intelligence. When you look at the numbers, it is difficult to comprehend the scope of it. 100,000 Nvidia Blackwell GPUs will power Argonne’s flagship system, Solstice. It’s not a misprint.
| Information Category | Details |
|---|---|
| Topic | Exascale Supercomputing & National Defense Applications |
| Key Technology | Exascale computing — systems performing 10¹⁸ floating-point operations per second |
| World’s Fastest (Nov 2024) | El Capitan — Lawrence Livermore National Laboratory — 1.742 exaFLOPS |
| Second Fastest | Frontier — Oak Ridge National Laboratory — 1.102 exaFLOPS |
| Primary Governing Body | U.S. Department of Energy (DOE) |
| Key National Laboratories | Argonne, Oak Ridge, Los Alamos |
| Upcoming Systems | Solstice, Equinox, Lux, Discovery, Mission, Vision |
| Solstice GPU Count | 100,000 Nvidia Blackwell GPUs |
| Defense-Focused System | Mission — dedicated to nuclear stockpile stewardship |
| El Capitan Cost | $600 million — AMD CPUs and GPUs, ~40 MW power consumption |
| Global Competitors | China (Tianhe-3, Sunway OceanLight), Japan (Fugaku), EU (JUPITER) |
| JUPITER Distinction | First exascale system outside the U.S. on TOP500; #1 on Green500 for energy efficiency |
| Strategic Framework | U.S. AI Action Plan — AI-enabled science and national security |
Things get really interesting—and a little unsettling—from the defense angle. Two machines called Mission and Vision are being delivered to Los Alamos, and the names aren’t random poetry. Designed to simulate the dependability of America’s atomic weapons without carrying out real tests, Mission was created specifically for nuclear stockpile stewardship. The United States last detonated a nuclear weapon in 1992. Since then, simulation has been used throughout the entire process of maintaining and certifying that arsenal. Therefore, more confident weapons assessments result from improved simulations. It’s not a lighthearted application. That’s computing at the existential level.
Whether the public understands what’s being built here is still up for debate. Chatbots and image generators—consumer goods that people can touch and debate—tend to dominate discussions about AI. In the meantime, systems that can perform quintillions of calculations per second are being used specifically to simulate fusion reactions, model warheads, and conduct the kind of climate and materials research that directly influences military strategy. The real competition seems to be taking place a few floors below the news.
This urgency is made tangible by China. According to reliable reports, Tianhe-3 and Sunway OceanLight, two of the top ten fastest systems in the world as of mid-2022, are both operational exascale machines; neither is listed on the public TOP500 list, which speaks for itself. It’s difficult to ignore how America’s response has shifted from scholarly concern to accelerated procurement as that dynamic has developed over the past few years. Three labs with nine machines isn’t a research program. It’s a stance.
Anticipated around 2028, Oak Ridge’s Discovery system is intended to greatly surpass Frontier, which is presently ranked second in the world. Frontier, which has been in operation since 2022, already achieves more than one exaFLOP. With AMD’s next-generation EPYC Venice processors and Instinct MI430X GPUs, Discovery should significantly surpass that limit. The timeline is tight and the generational gap is real, indicating that whoever is organizing these deployments doesn’t think the pressure from competition will lessen.
Europe is not excluded from this. The first exascale system outside of the US to be listed on the TOP500 list was Germany’s JUPITER supercomputer, which debuted in 2025. It is completely powered by renewable energy and is ranked #1 on the Green500 efficiency index, which is a truly remarkable technical accomplishment that is often overlooked. Alice Recoque is being built in France. In August 2024, the UK discreetly canceled the £900 million national exascale program that it had announced in 2023. That cancellation, which went virtually unnoticed, probably merits more attention than it got.
All of this points to a future in which unprocessed computing power is increasingly viewed as a strategic reserve that should be gathered, safeguarded, and used purposefully. Missiles and troop numbers might not define the battlefield of the coming decades. It might be determined by who can run AI workloads at a scale that their rivals cannot match, simulate materials more effectively, and model outcomes more quickly. America is making a huge, costly wager that those machines are being constructed at the appropriate times, by the appropriate partners, and in the appropriate locations. In all honesty, it’s still unclear if that wager will pay off.
