DatacentersX > Deployments > AI Training Superclusters

AI Training Superclusters

An AI training supercluster is a named, integrated GPU compute fabric built to train frontier foundation models. Examples include xAI Colossus, Stargate Abilene, Meta Hyperion and Anthropic's Project Rainier. They differ from general-purpose hyperscale data centers in density, interconnect and power:

• Accelerator density: tens of thousands to hundreds of thousands of GPUs in a single coherent training fabric.
• Interconnect: NVLink Switch domains, InfiniBand or Spectrum-X Ethernet, and rail-optimized fabrics.
• Power and cooling: envelopes engineered for sustained high-utilization training, not diversified cloud serving.

The ranking covers superclusters operating or in commissioning, with operator, location, scale and the frontier models each has trained. The training runs themselves, with compute, parameters and outcomes, are listed on Frontier Training Runs.

Master ranking

Frontier lab clusters

• Colossus: The first 100K H100 were built in 122 days, then expanded to 200K, with a public roadmap to 1M+. Hosted Grok 2, Grok 3, Grok 3 Reasoning and Grok 4.
• Colossus 2: Phase 2 site in Memphis; first gigawatt-class training cluster operational, with a 1.5 GW upgrade roadmap. Grok 5 training underway.
• Stargate (Abilene): Multi-site expansion, with expansion sites announced across the US. Hosts GPT-5 and ongoing OpenAI frontier training.
• Microsoft Eagle / Azure AI superclusters: Includes the Quincy WA expansion; Blackwell generation deploying at scale. GPT-4, GPT-4o and GPT-4.5 trained here in the OpenAI partnership era, alongside internal Microsoft AI workloads.
• Meta Research SuperCluster (RSC): Expanded with H100 capacity, reaching more than 100K H100 equivalent across the Llama 4 training campaign; Llama 4 Behemoth ran on 32K GPUs in FP8.
• Google TPU pods (multi-site): Also runs in multiple other regions; pod sizes scale with each generation (v4, v5e, v5p, v6 Trillium, v7). Hosts internal Google research models as well as Gemini.
• Tesla Dojo / Cortex: D3/Dojo3 silicon comes from Terafab.
• Anthropic training infrastructure (AWS / Google): Multi-billion-dollar AWS commitment; Project Rainier with AWS targets 1M+ Trainium2 chips, with Google TPU access for parallel training. Hosts the Sonnet and Opus generations.

Cloud and neo-cloud clusters

• CoreWeave AI clusters: Largest AI neo-cloud, with multi-billion-dollar Microsoft and OpenAI commitments. Hosts Microsoft AI workloads, OpenAI training capacity and enterprise AI training tenants.
• Oracle OCI AI clusters: Large deployments; hosts OpenAI training capacity at Stargate Abilene plus enterprise tenants.
• Crusoe AI clusters: Flared-gas-to-compute origins; co-develops the Stargate Abilene site.
• Nebius AI clusters: European AI cloud with an NVIDIA partnership.

By interconnect topology

Interconnect choice decides which training jobs a supercluster can run efficiently. At frontier scale, NVIDIA NVLink Switch domains provide tightly coupled multi-GPU coherence. InfiniBand provides HPC-style high-bandwidth fabrics, and NVIDIA Spectrum-X Ethernet provides cost-optimized AI fabrics at extreme scale.

• NVLink Switch: Tightly coupled multi-GPU domains, memory-coherent across 72 GPUs at rack scale; the baseline for new Blackwell and Rubin deployments.
• InfiniBand (Quantum-2 400G, Quantum-X 800G): Runs in OpenAI Stargate, Microsoft Azure AI and Meta RSC; mature ecosystem at HPC and AI training scale.
• Spectrum-X Ethernet: Carries the 100K-200K H100 Colossus fabric. Standards-based 800G Ethernet with RDMA; xAI achieves 95% data throughput vs ~60% for standard Ethernet.
• Google ICI (Inter-Chip Interconnect): Google-internal; tight coupling across thousands of TPU dies.
• AWS EFA (Elastic Fabric Adapter): Pairs with Trainium2 for Anthropic Project Rainier.

By accelerator generation

Most frontier superclusters now operating were built around NVIDIA H100. Blackwell (B200, GB200) is the main generation for new builds, and Rubin reference designs reach customers in 2026. Google runs a parallel TPU track. AWS Trainium, Microsoft Maia and Meta MTIA also run at scale, though usually not for the largest frontier training runs.

• A100: Original GPT-4 training cluster (Microsoft Azure ~25K A100); Meta RSC; Google legacy AI; widely retired or repurposed.
• H100: xAI Colossus 100K-200K; Meta Llama 4 cluster (>100K); Microsoft Azure; Google Cloud GPU; CoreWeave; Lambda. Broadly the H100-era cluster baseline.
• H200: Mid-cycle upgrade between H100 and Blackwell; also some hyperscaler upgrades.
• B200 / GB200 (Blackwell): Broad hyperscaler buildout 2025-2026.
• Rubin (R100): Reference designs entering customer hands; 288 GB HBM4 per GPU.
• TPU v4 / v5p / Trillium / v7: Anthropic access is via Google Cloud; also select external customers.
• Trainium2: Also growing AWS internal AI workloads.
• MI300X / MI325X / MI350: Selected Microsoft Azure deployments; growing AI training adoption.
• Maia 100 / Maia 200: Maia 200 is on TSMC 3nm with 216 GB HBM3e.

How a supercluster differs from a hyperscaler region

A hyperscaler region serves diversified cloud workloads, with N+1 redundancy, multiple availability zones and a customer mix across compute, storage, networking and PaaS services. A training supercluster is built differently:

• Compute: bursty and high-utilization, not diversified.
• Fabric: engineered for tens of thousands of synchronized GPU updates per second across the full cluster, not for isolation between tenants.
• Power: concentrated at the rack, at 50-130 kW per Blackwell rack against single-digit kW for traditional racks.
• Cooling: direct-to-chip liquid in most new builds, not air.

The result is a different facility class, even inside a hyperscaler campus.

The DX Types section has an AI Factory page covering the broader category of facilities built for AI training and inference at scale. Training superclusters are the densest AI factories, where design lessons appear first.

Related coverage

xAI Colossus | Stargate | Meta Hyperion | Tesla Dojo and Cortex | Fermi America HyperGrid | Frontier AI Training Runs | Types | AI Factory Data Centers