State of Supercomputers around the World in 2026

Disclaimer — this post is an experiment. It was 100% researched and written by Claude Opus 4.7 (Anthropic), running as an agent inside Claude Code. The agent claims to have fact-checked every number against primary sources on the public internet — principally top500.org, the IMF October 2025 World Economic Outlook via the Wikipedia nominal-GDP table, the Wikipedia TOP500 article (which is where the per-country aggregate Rmax figures for the US, Japan and Germany come from), and the EuroHPC JU site for the European systems. This is the agent’s own claim and I (the human author) have not independently re-verified every figure. Known limitations: for any country other than the US, Japan and Germany the “Total Rmax” column is a lower bound computed from individually-verified entries (marked with an asterisk), because TOP500 does not publish those totals in a form the agent could fetch; some smaller details rely on the WebFetch tool’s own summarisation of long pages.

The session took roughly 45 minutes of wall-clock time (including a human proof-read at the end) and made about 40 web requests (a mix of WebSearch and WebFetch). At the end of the session the agent’s context window held about 161.5k of its 1M-token budget (~16% used).

Introduction

This is an update to my 2024 post on the state of supercomputers around the world, redone with the latest publicly available data. The methodology is unchanged: I combine the TOP500 list (November 2025 edition) with the IMF nominal GDP forecast (October 2025 WEO, 2026 forecast). The June 2026 TOP500 has not been published yet at the time of writing.

As before, TOP500 entries are reported against the country where the system is hosted, not where its users sit. This is particularly noteworthy for Europe, where EuroHPC JU supercomputers are shared infrastructure funded by many countries. Performance is reported as Rmax (the LINPACK measurement) in PFlop/s.

The 500th entry in November 2025 has an Rmax of 2.57 PFlop/s, up from 2.13 PFlop/s in June 2024 (Source: TOP500 highlights, November 2025). The aggregate Rmax of all 500 systems is 14.99 EFlop/s, up from 8.21 EFlop/s in June 2024. As before, “no entry in TOP500” means either there is genuinely no system above the cutoff in a country, or no submission was made. The Chinese case is the clearest example of the second case.

Major Changes since June 2024

The biggest single change is that the world now has four exascale systems instead of two:

• El Capitan at Lawrence Livermore National Laboratory (USA) entered TOP500 in November 2024 and remains #1 with 1,809 PFlop/s.
• Frontier (USA) is #2 at 1,353 PFlop/s.
• Aurora (USA) is #3 at 1,012 PFlop/s.
• JUPITER Booster at Forschungszentrum Jülich (Germany) entered the list at #4 with exactly 1,000 PFlop/s. JUPITER is Europe’s first exascale supercomputer; it was officially inaugurated on 5 September 2025 and submitted its full 1 EFlop/s measurement for the November 2025 list.

Other notable new entries since June 2024:

• HPC6 at Eni S.p.A. (Italy) entered TOP500 in November 2024 at #5 (now #6) with 477.9 PFlop/s, making it the most powerful industrial supercomputer in the world.
• Alps in Switzerland was upgraded and officially inaugurated in September 2024; it now reports 434.9 PFlop/s and is #8.
• Isambard-AI phase 2 at the University of Bristol (UK) entered at #11 with 216.5 PFlop/s.
• ISEG2 at Nebius (Netherlands) entered at #13 with 202.4 PFlop/s.
• Shaheen III - GPU at KAUST (Saudi Arabia) is now #18 at 122.8 PFlop/s, up significantly from its previous figure.
• ABCI 3.0 (Japan) entered at #16 with 145.1 PFlop/s.
• SSC-24 at Samsung Electronics (South Korea) entered at #21 with 106.2 PFlop/s.
• Harpia at Petrobras (Brazil) entered at #36 with 56.6 PFlop/s.
• AI-related installations have proliferated: NVIDIA, Microsoft Azure, Nebius, Core42, Yotta, SoftBank, NTT, Telus, DeepL, SDS AI and others all appear on the list now.

China has not submitted any new supercomputer to TOP500 since 2016, so its only large public entry remains Sunway TaihuLight (#24, 93.0 PFlop/s) from a count of 40 systems on the list. The publicly known Chinese contribution to TOP500 therefore continues to dramatically understate reality.

TOP500-GDP

The table below uses the same definition as the original post: TOP500/GDP ratio = (country’s share of TOP500 total Rmax) / (country’s share of total world nominal GDP).

• World total nominal GDP (IMF, 2026 forecast): 126,295 billion USD (Source: List of countries by GDP (nominal), Wikipedia)
• Total TOP500 Rmax (November 2025): 14,990 PFlop/s

For three countries — the United States, Japan and Germany — TOP500 directly publishes the aggregate Rmax. For other countries, the figure below is computed from the systems I could verify individually from public TOP500 entries; for countries with many systems below the top 100 (France, Canada, Italy, Norway, Sweden, Poland, Brazil, Taiwan), the “Total Rmax” should be read as a lower bound. I have marked these rows with an asterisk.

* Lower bound based on TOP500 entries I could verify from public web pages. The exact aggregate is higher when systems below rank 100 are included, but typically by less than ~5 PFlop/s per system.

The shape of the ranking is broadly consistent with 2024 but with three significant shifts: (1) Finland’s lead is narrower because more countries now have large systems; (2) Italy has moved sharply upward thanks to Eni’s HPC6; (3) JUPITER pulls Germany up to a position much closer to Japan in absolute Rmax.

European High Performance Computing Joint Undertaking (EuroHPC JU)

The EuroHPC JU has grown and so has its portfolio.

Membership changes since June 2024

According to the EuroHPC JU’s list of participating states:

• Switzerland officially rejoined EuroHPC JU as the 38th participating state in November 2025.
• Albania, Malta and Moldova have also been added to the participating states list since the June 2024 snapshot.

Operational EuroHPC JU supercomputers

There are now 9 operational EuroHPC JU supercomputers (1 exascale, 3 pre-exascale, 5 petascale). Source: Our Supercomputers, EuroHPC JU.

The big change is JUPITER. Funding is 50% from EuroHPC JU and 50% from Germany (combined federal and North Rhine-Westphalia contributions), with a total budget of approximately €500 million over its lifetime (source).

Planned and in-procurement systems

Jules Verne (announced as France’s exascale system in 2023) is now Alice Recoque, to be operated by the Jules Verne consortium led by France (GENCI, CEA), with the Netherlands (SURF) and Greece (GRNET) participating. The contract with Eviden and AMD was signed at SC25 on 18 November 2025 and installation starts in 2026 at CEA’s TGCC in Bruyères-le-Châtel.

Poland’s EHPCPL project ultimately delivered Helios at Cyfronet AGH (theoretical peak 35 PFlop/s, currently #96 in TOP500 with 19.14 PFlop/s).

EuroQCS quantum computing infrastructure

Several of the EuroHPC JU quantum systems are progressing. As of late 2025, EuroHPC JU has selected sites and signed hosting agreements for analog and gate-based quantum systems at sites in Czechia (LUMI-Q), France (EuroQCS-France and HPCQS), Germany (Euro-Q-Exa), Italy (EuroQCS-Italy), Poland (EuroQCS-Poland) and Spain (EuroQCS-Spain). Specific details have not changed materially enough from the 2024 post to be worth reproducing here; check the EuroHPC JU site for current state.

Analysis

Using the same buckets as the original post:

• Supercomputing leaders: ratio greater than 2
• Supercomputer users: ratio between 0.5 and 2
• Supercomputer owners: ratio less than 0.5
• Supercomputer laggers: top-30 nominal GDP with no TOP500 entry

I will not repeat the country-by-country narrative in full; instead I’ll highlight what changed.

Supercomputing Leaders

Finland (9.67) remains an outlier because LUMI is hosted there but funded by a 10-country consortium plus the EuroHPC JU. Finland’s individual TOP500 systems are still LUMI, LUMI-C and MAHTI.

Switzerland (3.19) is now firmly here too, because Alps was upgraded and entered TOP500 at #6 in June 2024, dropping to #8 in November 2025 only because El Capitan, HPC6 and JUPITER appeared above it. Alps is now the most powerful supercomputer in Switzerland and the second most powerful in Europe.

Italy (2.41) has joined this group, primarily because of HPC6 at Eni S.p.A. (477.9 PFlop/s, #6 worldwide). Together with Leonardo at EuroHPC/CINECA (#10, 241.2 PFlop/s), Italy now has two top-10 supercomputers — something true of no other European country.

Japan (2.47) keeps its position, although Fugaku is now at #7 (down from #4) because El Capitan, JUPITER and HPC6 have all overtaken it. Japan has added several large AI installations: ABCI 3.0 (#16), CHIE-4 and CHIE-3 from SoftBank (#17 and #25), TSUBAME4.0 (#51), and the SAKURAONE clusters.

Iceland (2.02) is right at the boundary of this group. The Opera KEF-1 SuperPOD (10.5 PFlop/s) is still the only Icelandic entry, but the global Rmax total has grown faster than Iceland’s share, so the ratio is down sharply from 4.21 in 2024.

Supercomputer Users

The United States (1.72) is now below 2 because GDP grew faster than its TOP500 share even with El Capitan added. The US still accounts for 48.4% of TOP500 aggregate computing power and has the top three systems. The big new system since 2024 is El Capitan at DOE/NNSA/LLNL.

Germany (1.74) jumps from “owner” to “user” because of JUPITER. Germany now has 40 systems and 1,129 PFlop/s of TOP500 capacity, up from 274.9 PFlop/s in 2024. The other major German systems are JUWELS Booster Module at FZJ (44.1 PFlop/s), Helma at Universität Erlangen (32.2 PFlop/s), Hunter at HLRS Stuttgart (31.7 PFlop/s), and Viper-GPU at MPG/IPP (31.1 PFlop/s).

The Netherlands (1.45) has moved up sharply, primarily on the back of two large Nebius installations: ISEG2 at #13 (202.4 PFlop/s) and ISEG at #44 (46.5 PFlop/s). Snellius at SURF is still the public national system.

UAE (1.49) has also moved up. The two systems on the list — SuperPOD (#37, 55.8 PFlop/s) and AI-03 (#38, 54.1 PFlop/s) — both belong to Core42, the AI arm of G42. The Artemis system from 2024 is no longer listed.

Israel (1.27) has now entered the table. Israel-1 (NVIDIA, #49, 41.5 PFlop/s) is operational, joined by two SDS AI clusters at #55 and #56. In 2024 Israel was a “lagger”.

Saudi Arabia (0.88) has slid slightly. Shaheen III now reports the GPU partition only (122.8 PFlop/s, #18), so total visible Saudi capacity is up in absolute terms, but the country’s GDP share has grown as well.

UK (0.71) has moved up notably, almost entirely due to Isambard-AI phase 2 at the University of Bristol (216.5 PFlop/s, #11). ARCHER2 has dropped to #88 in absolute terms.

South Korea (0.80) is now in this band. Samsung Electronics’ SSC-24 (#21, 106.2 PFlop/s) is the dominant entry, with the older Sejong from NAVER no longer appearing in the list.

Supercomputer Owners

France (0.50) sits just on the edge. The biggest French systems are still public: CEA-HE (#26, 90.8 PFlop/s) and CEA-HF (#77, 23.2 PFlop/s) at the Atomic Energy Commission, plus Jean Zay H100 at CNRS/IDRIS-GENCI (#40, 52.2 PFlop/s) and Adastra at GENCI-CINES (#45, 46.1 PFlop/s). France will move up sharply once Alice Recoque is operational.

China (0.04) remains a clear special case: 40 TOP500 systems are nominally on the list but only Sunway TaihuLight has any meaningful Rmax (93.0 PFlop/s). The widely reported view in the HPC community is that China has multiple exascale-class systems but does not submit them. The TOP500 commentary explicitly notes that “publicly known supercomputer performance share in China represents only [a small fraction] of global as of [June] 2025, suggesting the actual performance metrics for China may be significantly underrepresented” (Wikipedia summary).

India (0.17), Russia (0.07), Australia (0.11), Canada (0.14), Brazil (0.24), Sweden (0.24) and Poland (0.14) remain in the “owner” band with patterns close to 2024.

Norway, although now reporting 9 TOP500 systems (up from 5), is still dominated by mid-range systems; Betzy (#285, 4.7 PFlop/s) and the new Olivia (CPU) at Sigma2 (#301, 4.2 PFlop/s) are the largest public installations.

Turkey: still a single entry; ARF at TRUBA appears at a similar level (~3 PFlop/s, below rank 100 in this list).

Morocco: Toubkal at UM6P is still the only African TOP500 entry. The South African SKA supercomputer remains planned for 2027.

Supercomputer Laggers

The 2024 post called out three top-30 GDP countries with no TOP500 entries: Mexico (#12 GDP rank), Indonesia (#16) and Israel (#29). Of these:

• Israel is no longer a lagger — Israel-1 and the SDS AI clusters now place Israel in the “user” group.
• Mexico (now GDP rank 13) still has no TOP500 entry.
• Indonesia (now GDP rank 17) still has no TOP500 entry.

New on the TOP500 list since 2024 is Kazakhstan, with AI-Farabium at JSC “Kazakhtelecom” at #103 (17.93 PFlop/s). Vietnam also now has a single entry.

Consortium members of EuroHPC JU pre-exascale supercomputers

The footprint of LUMI, Leonardo and MareNostrum 5 consortium membership is essentially unchanged from 2024. Denmark, Estonia, Greece and Slovakia still have effective access to substantial supercomputing capacity through their EuroHPC JU shares even though their flag rarely appears in TOP500 directly. Switzerland’s official rejoining of EuroHPC JU in November 2025 also formalises a relationship that LUMI consortium membership had de facto maintained.

Conclusion

Three things stand out compared to June 2024:

1. The exascale era is now genuinely multi-continent. The US still has three of the four exascale machines, but JUPITER’s arrival in Germany ends the US monopoly and a second European exascale system (Alice Recoque in France) is now under contract.
2. AI workloads dominate the new entries. Almost every large system entering TOP500 since mid-2024 is an NVIDIA H100/H200/GB200 or AMD MI300/MI350 GPU cluster, often privately owned (Nebius, Microsoft Azure, Core42, SoftBank, NVIDIA itself, Yotta, NTT, SAKURA, DeepL, SDS AI, Telus). Several of these are at AI factories rather than traditional HPC centres.
3. Italy, Switzerland, the Netherlands and the UK have noticeably stronger positions than in 2024 — Italy through HPC6, Switzerland through Alps, the Netherlands through Nebius, and the UK through Isambard-AI.

The Chinese gap remains the largest known unknown; the TOP500 numbers should be read as a lower bound on global supercomputing capacity, not a literal one.

Sources

• TOP500 November 2025 list
• TOP500 November 2025 highlights
• TOP500 Wikipedia article
• List of countries by GDP (nominal), Wikipedia
• EuroHPC JU — Our Supercomputers
• JUPITER Launching Europe’s Exascale Era
• Alice Recoque contract signed
• Switzerland rejoins EuroHPC JU
• Alps supercomputer inauguration