Nvidia silicon now sits inside more than 400 of the world's 500 fastest supercomputers, or 81% of the latest TOP500 list released this week at the ISC High Performance conference in Hamburg, Germany. The share of new entrants is starker: 90% of the systems added since the previous ranking run on Nvidia technology, a 17-system gain that leaves rival accelerator stacks fighting for the remaining sliver. Nvidia also swept the Green500 efficiency ranking, with the top eight machines all running on Nvidia GPUs.
The footprint goes well beyond GPUs. Nvidia networking — mostly Quantum InfiniBand, with the remainder on Ethernet — now connects 376 of the TOP500 systems, a record. GPUs accelerate 238 systems, also a record. And Grace CPU adoption climbed to 26 systems, up eight from the previous list, with roughly 2.5 million Grace CPUs shipped to date. That puts Nvidia inside the compute, the fabric, and increasingly the host processor of the machines defining state-of-the-art science.
Performance density is the lever. Nvidia says its TOP500 systems collectively deliver more than 2x the AI training throughput and nearly 3x the AI inference throughput of every other platform on the list combined. Those ratios are the reason procurement officers keep landing in the same place: a supercomputer specced for AI alongside traditional simulation now defaults to Nvidia almost regardless of geography.
“NVIDIA systems across the TOP500 now deliver more than 2x the AI training and nearly 3x the AI inference throughput of every other platform combined”— Chris Porter, Nvidia
Key facts
- 01Nvidia technology powers 400+ of the TOP500, or 81% of the list, gaining 17 systems since the previous ranking.
- 02Nearly 9 of every 10 systems new to the TOP500 are built on Nvidia technology, with 238 accelerated by Nvidia GPUs.
- 03Nvidia swept the top eight slots on the Green500; KAIROS at the University of Toulouse leads at 73.3 gigaflops per watt.
- 04Nvidia Grace CPU adoption hit 26 systems on the TOP500, up from 18, with roughly 2.5 million Grace CPUs shipped.
- 0535 Nvidia AI HPC supercomputers are in development across Europe, serving more than 3 million researchers.
Grace-based machines anchor the top of both lists. JUPITER, hosted at the Jülich Supercomputing Centre in Germany, ranks No. 5 on the TOP500 and is Europe's first exascale system. Alps sits at No. 10. On the Green500, KAIROS at the University of Toulouse takes No. 1 at 73.3 gigaflops per watt, with Grace Hopper systems sweeping the top four efficiency slots across France, Germany and the United Kingdom.
Each of those machines pairs an Nvidia GPU with a Grace CPU inside a single Grace Hopper Superchip, letting the two share memory with minimal overhead. That design choice — collapsing the CPU-GPU boundary rather than just bolting accelerators onto an x86 host — is what's producing the efficiency-per-watt numbers at the top of the Green500. The Vera CPU announced earlier this year extends the same approach to data-center workloads where agents run code, use tools and evaluate results rather than just answering single-shot prompts.
Europe is the heaviest builder. A record 35 Nvidia AI HPC supercomputers are in development across the continent, equipping more than 3 million researchers with new infrastructure for AI, accelerated science and industrial work. JUPITER alone is being used to map the human brain at cellular scale, simulate Earth's climate, and develop the AI behind next-generation 6G networks — three workloads that would have been split across separate facilities a generation ago.
The newest TOP500 entrants are running on Nvidia's Blackwell architecture. B200 and GB200 systems debuted across Asia, Europe and the United States, with the first GB200 deployments landing in Japan. A new AI factory came online in South Africa, and national AI systems entered the rankings from Saudi Arabia, Singapore and Vietnam.
“JUPITER is mapping the human brain at cellular scale, simulating Earth's climate and advancing the AI behind next-generation 6G networks”— Chris Porter, Nvidia
The TOP500 is updated twice a year and ranks systems by Linpack performance; the Green500 reranks the same list by gigaflops per watt. Both lists historically lagged commercial AI infrastructure — most of the largest hyperscaler training clusters never submit results. That makes the 81% figure a floor, not a ceiling, on Nvidia's actual share of frontier compute.
The counterweight is concentration risk. AMD's MI300 series and Intel's Gaudi line are present on the TOP500 but have not closed the gap on new deployments, and custom silicon from cloud providers — Google's TPUs, AWS Trainium, Microsoft Maia — is largely invisible on these public rankings. National HPC programs in Europe and Asia have flagged dependency on a single US vendor as a strategic concern, and several of the 35 European builds are explicitly designed to give the continent sovereign AI capacity that still happens to run on Nvidia hardware.
For Nvidia, the TOP500 result extends a pattern that's now visible at every layer of the AI stack: hyperscaler training, sovereign AI builds, exascale science, and the energy-efficiency frontier all converging on the same vendor. The pricing power that follows — across GPUs, NVLink, InfiniBand and now Grace and Vera CPUs — is what's underwriting Nvidia's data-center revenue trajectory and the next round of $6B-plus compute deals from customers like SpaceX-backed Reflection AI. Until a competing accelerator demonstrates the same efficiency-per-watt at exascale, the world's AI buildout will keep running on Nvidia by default.
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.
