Post

AI Factories 2025H2

🌍 AI Factories Top 20 — 2025 H2 Edition

Inspired by the TOP500 Supercomputer List, this list ranks the largest known or announced AI training infrastructures worldwide — often called “AI Factories”.
Figures are based on public announcements, credible leaks, and infrastructure analysis as of November 2025.


🏅 RankName / Project🌐 Location🏢 Owner / Operator⚙️ GPU Count / Compute Scale🧩 Platform / Architecture📆 Status📝 Notes
1Stargate (OpenAI + Microsoft)Quincy, WA, USAMicrosoft / OpenAI~150,000+ NVIDIA Blackwell GPUs (planned)Azure AI FabricAnnounced (2024)Part of $100B multi-phase AI infrastructure expansion
2Frontier AI Factory (NVIDIA + CoreWeave)USACoreWeave / NVIDIA~100,000 GPUs (H100/B100 mix)DGX CloudUnder ConstructionNVIDIA’s flagship partner cloud cluster
3xAI “Colossus” Cluster (Elon Musk / xAI)USAxAI~100,000 GPUs (H100 + B200)Custom Tesla Dojo hybridBuilding (2025)Colossus 1 & 2; dedicated to Grok model suite
4Amazon Bedrock / Titan ClusterUSAAWS~80,000 GPUs equivalentTrainium + H100 mixOperationalInternal Amazon FMs training platform
5Google Gemini Training InfrastructureUSAGoogle DeepMind~70,000 GPUs (A3 Mega + TPUv5e)TPUv5e / TPUv6OperationalPowers Gemini and multimodal DeepMind systems
6Meta “AI Research SuperCluster” (RSC 2)USAMeta~60,000 GPUs (H100/B200)PyTorch / NVIDIA DGXUnder ExpansionSuccessor to RSC Phase 1 (16k A100s)
7Azure “AI Supercomputer” ClusterUSA + EUMicrosoft Azure~60,000+ GPUs (H100s)Azure NDv5OperationalUsed for OpenAI + Copilot workloads
8G42 Condor Galaxy (UAE)Abu Dhabi, UAEG42 + Cerebras~54,000 GPU-equivalentsCerebras WSE + NVIDIAExpansion (CG-2, CG-3)Key MENA AI hub, multi-region rollout
9China “Baidu AI Cloud” ClusterBeijing, ChinaBaidu~50,000 GPUs (H800 / Ascend 910B)Kunlun + NVIDIA H800OperationalBaidu’s core generative AI infrastructure
10Alibaba Tongyi ClusterHangzhou, ChinaAlibaba Cloud~45,000 GPUs (H800 / Ascend 910B)PAI / Lingjun AI InfraOperationalPowers Tongyi Qianwen LLM family
11Tencent AI SuperCenterShenzhen, ChinaTencent~40,000 GPUsH800 + Blackwell (planned)ExpansionSupports Hunyuan model suite
12DOE “Frontier AI Discovery” ProjectOak Ridge, USAU.S. Dept. of Energy (DOE)~35,000 GPUs equivalentAMD Instinct / HPE CrayBuilding (2025)Research-oriented, exascale AI-HPC hybrid
13DOE “Lux” AI FacilityUSAU.S. Dept. of Energy~30,000 GPUs equivalentHybrid DOE HPC / NVIDIAAnnouncedNext-gen AI scientific computing center
14EU “JUPITER AI Extension” (EuroHPC)Jülich, GermanyEuroHPC / Forschungszentrum Jülich~25,000 GPUs (MI300A + H100)Modular SupercomputingUnder ConstructionEurope’s flagship AI-HPC hybrid
15Saudi NEOM AI FactoryNEOM, Saudi ArabiaNEOM Tech / NVIDIA~24,000 GPUsNVIDIA + Oracle CloudAnnouncedAI hub for NEOM smart city projects
16ETH Zurich “Alps” AI-HPCZurich, SwitzerlandCSCS / ETH~22,000 GPUsNVIDIA Grace HopperLaunching 2025European sovereign AI compute center
17France “Jean Zay 2” AI UpgradeParis, FranceGENCI~20,000 GPUsNVIDIA / Atos BullSequanaConstructionFrench national AI infrastructure
18Japan “ABCI 3.0”Tokyo, JapanAIST~18,000 GPUsNVIDIA H200Planned (2025)Asia’s open AI research cluster
19Amazon / Anthropic “Claude Factory”USAAWS + Anthropic~15,000 GPUsTrainium2 / H100OperationalDedicated to Claude 3+ and next-gen models
20OpenAI “Legacy Azure Cluster” (SuperPod)Iowa, USAMicrosoft / OpenAI~14,000 GPUs (A100s)Azure AIOperationalFormer GPT-4 / GPT-3 training supercluster

Notes (2025):

  • 🏭 “AI Factories” are large-scale GPU or accelerator clusters dedicated to training or serving AI foundation models.
  • 🧠 GPU counts are based on credible estimates; real totals may vary due to hybrid architectures (TPUs, WSEs, Trainium, etc.).
  • 🌎 The U.S. leads in installed capacity, but UAE, China, and EU projects are rapidly expanding.
  • 🚀 Projects like Stargate and Frontier AI Factory mark the beginning of the 100k+ GPU era of AI infrastructure.
  • ⚡ Expect several >200k-GPU builds to break ground before 2026.

Last updated: November 2025
Source: company filings, press releases, HPC community trackers, and infrastructure research.

This post is licensed under CC BY 4.0 by the author.