What is an AI datacenter?

An AI datacenter is a facility purpose-built or retrofitted for GPU and accelerator workloads — typically at rack densities above 30 kW, with liquid-cooling readiness and high-bandwidth fabric (InfiniBand NDR, NVLink, or Spectrum-X). At least 40% of IT load runs AI training or inference on GPUs such as NVIDIA H100, H200, B200, or GB200 NVL72.

Which datacenter has the most GPUs?

As of 2026, xAI Colossus in Memphis, Tennessee holds the title of the world's largest single-site GPU deployment, with more than 200,000 NVIDIA H100 and H200 GPUs in one InfiniBand domain. Meta Hyperion (Louisiana) and OpenAI Stargate Abilene (Texas) are scaling toward gigawatt-class deployments that will overtake it within the next 18 months.

What is the largest AI cluster?

xAI Colossus is the largest contiguous AI training cluster in 2026, exceeding 200,000 NVIDIA H100/H200 GPUs on a single InfiniBand NDR 400G fabric. Meta's Hyperion build and OpenAI's Stargate program are constructing multi-site fabrics that will exceed one million accelerators across coupled campuses by 2027.

Which GPU cloud is cheapest for H100?

On-demand: Together AI ($1.75/hr) and Voltage Park ($1.89/hr) lead. Reserved 1-year: Voltage Park ($1.20/hr) and Crusoe ($1.39/hr) are typically lowest. AWS p5 sits at the top of the price stack at ~$4.10/hr on-demand.

How is Datacenter.computer different from a datacenter directory?

Datacenter.computer is an agent-readable intelligence layer, not a directory. It normalizes capacity, power, fabric and location across 833 sourced facilities and 40 operators into a single queryable graph for AI agents, traders and procurement teams.

[INT] Intelligence Graph

The living map of global AI infrastructure.

833 sourced facilities · 40 operators · 53 countries · 290 metros. Normalized and exposed as a queryable graph for agents, traders and procurement teams.

Global

graph nodes

Graph

operators · metros · GPUs

Cited

source on every field

Quarterly

refresh cadence

Explore →

[00] Global facility map

Interactive index · US + worldwide

[MAP] Global Datacenter Index

833 facilities · 40 operators · 53 countries · 4,536.77 MW

Showing top 250 of 833 facilities in this view.

[01] Citation graph

Auto-linked node network · 20 anchor posts ↔ hub pages

Every citation node is automatically linked to its cluster hub (providers, benchmarks, marketplace) and to sibling posts via keyword overlap. The graph below is generated live from the post index — no hand-curated links.

25 nodes88 edges68 auto-linked

cluster link auto topic linkhover a node → highlight neighborhood · click to open

[02] Signal stack

Six telemetry streams. One unified graph.

CAPACITY

Documented accelerator capacity

Per-facility accelerator platforms and recorded MW where publicly disclosed, aggregated by operator, metro and country. Undisclosed values stay undisclosed.

POWER

PUE + W/FLOP tracking

Published efficiency metrics per facility, where disclosed. Ranks providers by sustained TFLOPS per watt — the only metric that matters at hyperscale.

PRICE

Spot + reserved curves

Historical price surfaces with 90-day forecast. Predicts H100/H200/B200 spot floor 7 days out within 8% MAPE.

LATENCY

Inter-region mesh

Continuous latency probes across 80+ regions. Identifies optimal multi-region placement for training and inference workloads.

INTERCONNECT

NVLink · InfiniBand · RoCE

Cluster topology graph: which providers offer rail-optimized 400G IB fabrics vs. converged Ethernet. Critical for >256-GPU jobs.

GRID

Power source + carbon

Facility-level grid mix, carbon intensity (gCO₂/kWh), and PPA structure. Enables carbon-aware workload routing.

[03] Illustrative event feed

Example change events

Illustrative examples of the change types tracked in the index (capacity additions, efficiency updates, pricing moves). Not a live telemetry stream — licensed customers receive verified change logs with each update cycle.

00:04alertH100 spot < $1.80/hr — Lambda us-east-3 (24 GPUs live)

00:11infoCrusoe us-central added 384× H200 capacity (+18% region)

00:23warnAWS p5.48xlarge queue depth >2d in us-west-2

00:38alertB200 SXM6 now live · CoreWeave us-east · $5.94/hr

00:51infoNebius EU-FI-01 PUE update: 1.13 (best-in-class)

01:02warnio.net global H100 supply down 22% (24h)

01:14infoTensorWave MI300X price cut: $2.79 → $2.49/hr

Integrate