[Ranking 04/05] · 2026 Edition

Top 100 GPU Clusters

The largest GPU superclusters powering frontier AI — ranked by GPU count, interconnect fabric, and training throughput.

Top 100 Datacenters →Top 100 Operators →Top 100 AI Factories →Top 100 Infra Cos →Compare →
[AI Summary]

The Top 100 GPU Clusters ranks the world's largest production GPU superclusters by accelerator count, interconnect fabric (InfiniBand NDR, NVLink, Spectrum-X), and training throughput. The 2026 edition is led by xAI Colossus and Meta's Hyperion build, with H100, H200, B200, GB200 NVL72, and MI300X clusters from hyperscalers, neoclouds, and sovereign operators — each profiled at cluster-level, not facility-level.

[Cross-Reference]
Top 100 AI Factories →
See the physical training factories these GPU clusters sit inside, at hundreds of MW to gigawatt scale.
Top 100 AI Datacenters →
Map clusters to their parent facilities — power envelope, cooling, and network fabric.
Top 100 Operators →
See which hyperscalers and neoclouds operate the largest H100, H200, and B200 fleets.
Top 100 Infra Cos →
Trace the silicon, NVLink, InfiniBand, and Spectrum-X vendors behind these clusters.
[T1]

Tier 1 — 100k+ GPU Superclusters

Operational clusters above 100,000 accelerators.

10 entries
  1. 001xAI Colossus (200k H100)
  2. 002Meta Hyperion (180k GB200)
  3. 003OpenAI Stargate Abilene (150k GB200)
  4. 004Anthropic Project Rainier (140k Trainium2)
  5. 005Google TPU v6 Council Bluffs (130k)
  6. 006Microsoft Mt Pleasant (120k H200)
  7. 007Meta Prometheus (110k B200)
  8. 008AWS Project Ceiba (100k GH200)
  9. 009Oracle Stargate I (100k GB200)
  10. 010CoreWeave Plano GB200 (100k)
[T2]

Tier 2 — 50k–100k GPU Clusters

Major training clusters in the 50–100k range.

10 entries
  1. 011Crusoe Abilene GB200 (80k)
  2. 012Lambda Allen H200 (70k)
  3. 013Nebius Kansas City H200 (65k)
  4. 014Microsoft Atlanta H200 (60k)
  5. 015Google St Ghislain TPU v6 (60k)
  6. 016Meta Eagle Mountain B200 (55k)
  7. 017Azure Sydney H200 (55k)
  8. 018AWS Indiana Trainium2 (55k)
  9. 019CoreWeave Las Vegas H200 (55k)
  10. 020Oracle Phoenix GB200 (50k)
[T3]

Tier 3 — 25k–50k GPU Clusters

Large clusters in the 25–50k range.

10 entries
  1. 021Together AI Reno H200 (40k)
  2. 022Voltage Park Houston H100 (40k)
  3. 023Nebius Mäntsälä H200 (40k)
  4. 024Applied Digital Ellendale H200 (35k)
  5. 025Lambda Las Vegas H100 (35k)
  6. 026CoreWeave Lancaster GB200 (35k)
  7. 027Crusoe Wyoming GB200 (35k)
  8. 028Microsoft Quincy H100 (30k)
  9. 029Google The Dalles TPU v5p (30k)
  10. 030Meta Luleå H100 (30k)
[T4]

Tier 4 — Sovereign Clusters

Sovereign AI clusters operated by nation-scale platforms.

10 entries
  1. 031G42 Condor Galaxy 1 (CA)
  2. 032G42 Condor Galaxy 2 (CA)
  3. 033Humain Riyadh GB200
  4. 034Humain Dammam H200
  5. 035DataVolt NEOM GB200
  6. 036Saudi Aramco Dhahran H200
  7. 037Khazna Abu Dhabi 1 H200
  8. 038Moro Hub Dubai H100
  9. 039Bharti Airtel India H200
  10. 040Reliance Jamnagar GB200
[T5]

Tier 5 — China Clusters

Largest disclosed China-region AI clusters.

10 entries
  1. 041ByteDance Tianjin H20 (60k)
  2. 042Alibaba Zhangbei H20 (50k)
  3. 043Tencent Guian H20 (45k)
  4. 044Baidu Yangquan Ascend 910C (40k)
  5. 045Huawei Ulanqab Ascend 910C (40k)
  6. 046Moonshot Beijing H20 (30k)
  7. 047Zhipu Beijing H20 (25k)
  8. 048iFlytek Hefei Ascend (25k)
  9. 049SenseTime Shanghai (20k)
  10. 050Kuaishou Hohhot (20k)
[T6]

Tier 6 — EU AI Factory Clusters

EuroHPC AI Factory clusters.

10 entries
  1. 051JUPITER (Jülich, DE)
  2. 052LUMI-AI (Kajaani, FI)
  3. 053Leonardo-AI (Bologna, IT)
  4. 054MareNostrum 5 AI (Barcelona)
  5. 055MeluXina AI (Luxembourg)
  6. 056Karolina AI (Ostrava)
  7. 057Discoverer+ (Sofia)
  8. 058Deucalion (Minho, PT)
  9. 059Vega AI (Maribor)
  10. 060Daedalus (Athens)
[T7]

Tier 7 — APAC Research Clusters

Large APAC research and lab clusters.

10 entries
  1. 061AIST ABCI 3.0 (Tokyo)
  2. 062RIKEN MN-Core (Kobe)
  3. 063NAVER Sejong H200
  4. 064Kakao Ansan H200
  5. 065SK Telecom Gasan H200
  6. 066Samsung SDS Pyeongtaek H200
  7. 067NTT Inzai H200
  8. 068SoftBank Hokkaido H200
  9. 069NCI Australia Gadi-AI
  10. 070NSCC Singapore ASPIRE 2A+
[T8]

Tier 8 — US National Lab Clusters

US DOE and national lab AI clusters.

10 entries
  1. 071Frontier (ORNL)
  2. 072Aurora (ANL)
  3. 073El Capitan (LLNL)
  4. 074Perlmutter (NERSC)
  5. 075Polaris (ANL)
  6. 076Vista (TACC)
  7. 077Delta (NCSA)
  8. 078Bridges-2 (PSC)
  9. 079Expanse (SDSC)
  10. 080Anvil (Purdue)
[T9]

Tier 9 — Custom-Silicon Clusters

Largest non-NVIDIA clusters in production.

10 entries
  1. 081AWS Trainium2 (IN)
  2. 082Google TPU v6p (IA)
  3. 083Google TPU v5p (OR)
  4. 084Meta MTIA v2 (OH)
  5. 085Microsoft Maia 100 (WI)
  6. 086Cerebras CS-3 (G42)
  7. 087Groq LPU DFW
  8. 088SambaNova SN40L
  9. 089Tenstorrent Wormhole (Austin)
  10. 090Graphcore Bow (Bristol)
[T10]

Tier 10 — Emerging & Specialty Clusters

Smaller but notable clusters rounding out the top 100.

10 entries
  1. 091Hut 8 Salt Creek H200
  2. 092TeraWulf Lake Mariner H200
  3. 093Iris Energy Childress H200
  4. 094Core Scientific Denton H200
  5. 095Cipher Black Pearl H200
  6. 096Marathon Granbury H200
  7. 097Galaxy Helios H200
  8. 098Riot Corsicana H200
  9. 099Cleanspark Wyoming H200
  10. 100Nscale Glomfjord H200
[FAQ]

Frequently asked questions

How are GPU clusters ranked?
By installed GPU count, peak FP8/FP16 throughput, interconnect topology (NVLink, InfiniBand NDR/XDR, Spectrum-X, ICI), and verified training jobs run on the cluster as of Q1 2026.
Which GPU generations are eligible?
NVIDIA H100, H200, B200, GB200, GB300, AMD MI300X/MI325X/MI355X, Google TPU v5p/v6, AWS Trainium 2, and Intel Gaudi 3 superclusters of 1,000+ accelerators.
Are inference-only clusters included?
Yes, but ranked separately within tier blurbs. Training-first superclusters dominate the upper tiers because of interconnect requirements.
What's the difference between a cluster and a factory?
A GPU cluster is a tightly-coupled compute fabric (one or more pods on a unified InfiniBand/NVLink domain). A factory is the building. One factory can house several Top 100 clusters.
How can I compare specific GPU generations?
See our H100 vs H200 vs B200 deep-dive in the blog for per-GPU TCO, memory bandwidth, and FP8 throughput comparisons.
[Knowledge Graph]

Company → Datacenter → GPU → Customer → Industry

Every entity on this site is cross-linked. Follow the graph from operators down to specific facilities, GPU clusters, customers, and sectoral context.

[Companies]
[Rankings]
[Comparisons]
[Best of 2026]
[Pricing & Calculators]
[Infrastructure profiles]
[News & analysis]