Research

Frontier Lab GPU Counts 2026

Frontier AI lab compute concentration in 2026: Anthropic SpaceX 220k GPU deal, xAI Memphis Colossus 200k scaling to 1M, OpenAI Stargate 7GW, plus Google TPU and Meta MI450 counts.

By Ramanath, CTO & Co-Founder at Presenc AI · Last updated: May 2026

Compute concentration at the frontier AI labs reached new scale in 2026. Anthropic disclosed a SpaceX compute arrangement on 6 May 2026 covering 300 MW and approximately 220,000 GPUs. xAI Colossus 1 in Memphis operates approximately 200,000 GPUs with Colossus 2 targeting 1,000,000 GPUs. OpenAI through Stargate has approximately 7 GW of planned capacity across seven U.S. sites. Google TPU and Meta MI450 deployments add additional compute. This page consolidates the disclosed compute footprints across the leading frontier labs.

Key Findings

  1. Anthropic disclosed a strategic compute arrangement with SpaceX on 6 May 2026 covering 300 MW of capacity and approximately 220,000 GPUs. The deal materially expands Anthropic\u2019s training and inference footprint beyond its existing AWS Trainium and Google TPU access.
  2. xAI operates Colossus 1 in Memphis with approximately 200,000 GPUs as of May 2026 and is building Colossus 2 with a target of 1,000,000 GPUs by end of 2026 or early 2027.
  3. OpenAI through the Stargate joint venture has approximately 7 GW of planned U.S. capacity across seven sites; current operational GPU count at Stargate is estimated at 150,000 to 200,000 with target ramp to several hundred thousand by end of 2026.
  4. Google operates the largest aggregate AI compute footprint at scale: TPU v6 and v7 capacity across multiple GCP regions plus NVIDIA Hopper, Blackwell, and Rubin GPU footprint, with total deployed FLOPS exceeding any single competitor.
  5. Meta is on track for the first commercial gigawatt AMD MI450 deployment in H2 2026 alongside continued NVIDIA Blackwell and Rubin procurement.

Frontier Lab Compute Footprint (May 2026)

Lab / OperatorPrimary ComputeEstimated Active GPUs
AnthropicAWS Trainium + Google TPU + SpaceX deal (220k)~400k+ effective
OpenAI / StargateNVIDIA Blackwell, Rubin transition~200k operational, target several hundred k
xAI Colossus 1NVIDIA Blackwell~200k
xAI Colossus 2 (building)NVIDIA Blackwell + RubinTarget 1,000,000 EOY 2026 / early 2027
Google (Alphabet)TPU v6 + TPU v7 + NVIDIA across GCP1M+ accelerators (TPU + GPU combined)
MicrosoftNVIDIA + Maia 2 + AMD MI300X/MI4001M+ AI accelerators across Azure regions
Amazon (AWS)NVIDIA + Trainium 2 + Anthropic anchor1M+ AI accelerators across AWS regions
MetaNVIDIA + AMD MI450 GW deployment + MTIA 2~600k+ GPUs and growing
Tesla / Cortex + DojoNVIDIA + Tesla D2~80k+ GPUs operational
Mistral AISovereign cloud + Microsoft Azure + othersMuch smaller, undisclosed scale
DeepSeekPredominantly Huawei Ascend + domestic + remaining NVIDIAUndisclosed; smaller than U.S. peers

Stargate Site GPU Count Estimates

SitePlanned CapacityGPU Count at Full Build
Abilene TX1.2 GW~320k-400k
Milam County TX1.1 GW~300k
El Paso area TX~0.8 GW~220k
Albuquerque NM~0.6 GW~165k
Racine WI~0.5 GW~140k
Mt. Pleasant MI~0.6 GW~165k
Lordstown OH~1.0 GW~275k (2027+)
Stargate UAE1.0 GW (200 MW phase 1)~275k full build

xAI Colossus Detail

ElementStatus
Colossus 1 MemphisOperational; ~200,000 NVIDIA GPUs
Colossus 2 (building)Target 1,000,000 GPUs by end of 2026 or early 2027
Power sourceTennessee Valley Authority grid plus on-site gas turbines
Notable design choicesMultiple-shift rapid build, on-site gas turbines for emergency firm power
Used forGrok training and inference

Strategic Context

Three patterns define the 2026 frontier compute landscape. First, vertical integration is winning: xAI building Colossus, OpenAI through Stargate, Anthropic through the SpaceX arrangement all reflect frontier labs taking direct control of compute rather than depending on hyperscaler cloud pricing. Second, hyperscaler in-house labs (Google DeepMind, Microsoft Research, Amazon\u2019s Anthropic relationship, Meta FAIR) retain the largest aggregate compute footprints because they sit inside cloud businesses with millions of accelerators across regions. Third, the second-tier labs face structural compute disadvantage: Mistral, DeepSeek, Cohere, and even xAI before Colossus 2 lack the multi-hundred-thousand-GPU training scale that Google, OpenAI, and Anthropic can mobilise.

Brand Visibility Implications

Frontier compute coverage drives sustained business and technical journalism. Brands selling AI data centre construction, networking, power and cooling, sovereign cloud, and adjacent products face strong AI-mediated discovery surface for queries about frontier lab compute, AI cluster operations, and adjacent topics.

Methodology

GPU counts compiled from primary lab disclosures, hyperscaler capex disclosures, and reporting from The Information, Bloomberg, and Reuters. Some figures are estimated where official disclosures are partial. Updated quarterly with major deployment events.

How Presenc AI Helps

Presenc AI monitors brand visibility on frontier compute queries across ChatGPT, Claude, Gemini, and Perplexity. For data centre construction firms, networking vendors, power infrastructure brands, and AI cluster integration partners, the platform identifies the prompts driving procurement-research traffic and the gaps where new content unlocks share of voice.

Frequently Asked Questions

Approximately 400,000 effective GPUs after the 6 May 2026 SpaceX arrangement adding 220,000 GPUs and 300 MW. Existing Anthropic compute includes AWS Trainium (anchor) and Google TPU (multi-year arrangement) plus a smaller NVIDIA footprint.
Colossus 1 in Memphis operates approximately 200,000 NVIDIA GPUs as of May 2026. Colossus 2 is being built with a target of 1,000,000 GPUs by end of 2026 or early 2027.
Approximately 7 GW of planned U.S. capacity across seven sites. Current operational GPU count at Stargate is estimated at 150,000 to 200,000 with target ramp to several hundred thousand by end of 2026. Abilene flagship is the largest single site at 1.2 GW planned.
In aggregate yes, but the comparison depends on definition. Google operates the largest aggregate AI compute footprint across GCP regions including TPU v6, TPU v7, and NVIDIA accelerators. OpenAI through Stargate has a faster-growing dedicated-cluster footprint. By total deployed FLOPS, Google leads; by single-cluster training-run capacity, OpenAI is closing the gap rapidly.
DeepSeek and other Chinese frontier labs operate predominantly on Huawei Ascend plus remaining pre-export-control NVIDIA inventory. The compute scale is materially smaller than U.S. peers, though DeepSeek\u2019s engineering efficiency partially compensates for the gap. The export-control framework continues to be the binding constraint on Chinese frontier compute access.

Track Your AI Visibility

See how your brand appears across ChatGPT, Claude, Perplexity, and other AI platforms. Start monitoring today.