DataCentersX > Sites > AI Training Superclusters

AI Training Superclusters

An AI training supercluster is a named, integrated GPU compute fabric purpose-built for training frontier foundation models. Superclusters differ from general-purpose hyperscale data centers in three ways: extreme accelerator density (tens of thousands to hundreds of thousands of GPUs in a single coherent training fabric), specialized interconnect topologies (NVLink Switch domains, InfiniBand or Spectrum-X Ethernet, rail-optimized fabrics), and power and cooling envelopes engineered for sustained high-utilization training workloads rather than diversified cloud serving.

This page covers the named superclusters currently operating or in commissioning. Each entry tracks operator, location, accelerator count and generation, interconnect, power class, and the frontier training programs the cluster has hosted. The companion Frontier Training Runs page covers the events that ran on these clusters - the training jobs themselves with their compute, parameters, and outcomes. Read the two together to map infrastructure to events.

Master ranking

By interconnect topology

Interconnect choice shapes what training jobs a supercluster can run efficiently. Three architectures dominate at frontier scale: NVIDIA NVLink Switch domains for tightly-coupled multi-GPU coherence, InfiniBand for traditional HPC-style high-bandwidth fabrics, and NVIDIA Spectrum-X Ethernet for cost-optimized AI fabrics at extreme scale.

By accelerator generation

Most operating frontier-tier superclusters were built around NVIDIA H100. Blackwell (B200, GB200) deployment is now the dominant new-build generation, with Rubin reference designs entering customer hands in 2026. Google operates a parallel TPU track. AWS Trainium and a small number of internal hyperscaler accelerators (Microsoft Maia, Meta MTIA) are also at scale though typically not for the largest frontier training runs.

What makes a supercluster distinct from a hyperscaler region

A traditional hyperscaler region is optimized for diversified cloud workloads with N+1 redundancy, multiple availability zones, and customer mix across compute, storage, networking, and PaaS services. A training supercluster is optimized differently. The compute is bursty and high-utilization rather than diversified. The fabric is engineered for tens of thousands of synchronized GPU updates per second across the full cluster, not for inter-tenant isolation. The power envelope is concentrated at the rack level (50-130 kW per Blackwell rack vs single-digit kW for traditional racks). The cooling is direct-to-chip liquid in most new builds, not air. The result is a fundamentally different facility class even when located within or adjacent to a hyperscaler campus.

The DX Types pillar has a dedicated AI Factory child that covers the broader category of facilities optimized for AI training and inference at scale. AI training superclusters are the densest specific instances of the AI Factory category - the maximum-density configuration where the architectural lessons get learned first.

Where this fits

This page covers infrastructure (named superclusters with stable attributes). The Frontier Training Runs page covers events (specific training jobs that consumed compute on these clusters). Reading the two together maps the AI infrastructure to the AI events. Cross-pillar references run through Types:AI Factory for the facility class, Sites for the specific named campuses, Bottleneck Atlas for the supply chain dependencies, and SX:NVIDIA Spotlight for the silicon side.

Related coverage

Frontier Training Runs | AI Factory | Sites | xAI Colossus | Stargate | Meta Hyperion | Tesla Dojo | Bottleneck Atlas | SX:NVIDIA Spotlight | SX:HBM | SX:CoWoS