← Blog
June 8, 2026 · Analysis · Kent Langley

Who Has More Compute? The Frontier Labs, Ranked for Service Delivery

A mid-2026 snapshot of compute capacity across the pure frontier labs, weighted for who can actually deliver top-tier service consistently, not just who can train the biggest model.

The one-sentence answer

If "more compute" means announced, OpenAI wins by a mile (9+ GW). If it means online and serving customers today, Google DeepMind and Anthropic lead, because almost nobody's headline gigawatts are actually plugged in yet.

The gap between those two sentences is the whole story.

9+ GW
OpenAI announced, by 2029
~0.3 GW
OpenAI online now, satellite-confirmed
~20:1
Announced vs online, sector-wide
17 / 8
Claims confirmed vs killed in verification

The honest reframe

You asked who has more compute, scoped to three things that matter for delivering consistent service: data-center power (GW), installed accelerator fleet, and peak inference-serving capacity. Here is what the research kept running into:

Nearly every headline number is contracted, not operational. OpenAI's 9 GW, Anthropic's 5 GW AWS envelope, the 3.5 GW Broadcom deal: all signed, almost none online. The only firmly operational figures the verification gate would accept were OpenAI's ~0.3 GW at Abilene (independently confirmed by satellite imagery) and xAI's ~300 MW Colossus 1 as of mid-2025. Everything else is a press release with a future date attached.

So the ranking that matters for your question (consistent top-level service) is not the announced-capacity leaderboard. It is the much shorter list of who has mature compute online, serving real traffic, with enough platform diversity to absorb a demand spike without melting down.

Reliability-weighted ranking (for consistent service delivery)

RankLabWhyConfidence
1 (tie)Google DeepMindMost vertically integrated: owns the silicon (TPU), the data center, and the network. TPUv7 Ironwood is at near-parity with Nvidia's flagship. Only player proven at coherent multi-datacenter training. Most genuinely-online mature capacity.High
1 (tie)AnthropicMost platform-diversified live fleet (1M+ Trainium2 in use now, 1 GW+ of Google TPU landing in 2026, plus Nvidia GPUs). Diversity is a reliability hedge no single-platform rival has. Caveat below.High
3OpenAILargest future ceiling (9+ GW, 2M+ chips) but the most constrained present: only ~0.3 GW online against the largest user base in the industry. Structurally the most exposed to the compute crunch today.High
4xAIHeld the largest single fully-operational coherent cluster (Colossus 1, mid-2025). But its 2026 expansion claims failed verification, and it is the least diversified on serving.Medium
UnrankedMeta, DeepSeek, MistralTheir compute claims either failed adversarial verification (Meta's Prometheus/Hyperion, DeepSeek's fleet) or never surfaced (Mistral). This is an evidence gap, not a confirmed absence of compute.Low

The reliability caveat that cuts against the obvious read: on June 2, 2026, six days before this report, Claude suffered a major global outage that Anthropic attributed to "unexpected capacity constraints." Anthropic has also run documented usage limits on Claude through 2026. Diversification helps, but demand is outrunning supply even for the best-hedged lab. No lab in this set is immune to capacity-driven throttling right now.

Axis 1: Data-center power (GW)

The bottleneck of the era is not chips, it is power. Here is online-versus-announced, the only distinction that matters.

LabBacker / sourceOnline now (mid-2026)Contracted / announcedReality check
OpenAIMicrosoft + Oracle + Stargate/SoftBank~0.3 GW (Abilene only; 6 other sites at 0 GW)>9 GW by 2029 (≈20M H100-equiv of compute)~97% of the headline is not built. Abilene itself capped ~1.2 GW per March 2026 reporting.
AnthropicAWS (primary) + Google Cloud TPU>1 GW TPU landing in 2026, plus ~1 GW of the AWS envelope by end-20265 GW AWS + up to 1M TPU + 3.5 GW Broadcom (2027+)Best near-term online ratio of the group. ~80% of AWS capacity is still forward.
Google DeepMindGoogle Cloud TPU v7Large mature TPU fleet (no single GW figure disclosed; serves Google-scale traffic today)43-superpod, ~400K-chip fabricLeast announced-hype, most actually running.
xAISelf-built Colossus (Memphis)~0.3 GW (Colossus 1, confirmed mid-2025)Colossus 2 / ~2 GW / 555K GPUsForward 2026 claims failed verification. Confirmed figure is a 2025 snapshot.
MetaSelf-builtUnverifiedPrometheus ~1 GW (2026), Hyperion up to 5 GWBoth cluster claims failed verification.

Reading the gigawatts (sizing-data-centers lens): the "9 GW ≈ 20 million H100s" line is a compute-equivalent measure, not a physical chip count. Do the arithmetic and 9 GW across 20M H100s implies ~450W per chip, well below an H100's real ~1kW all-in draw. The reconciliation: Blackwell and Rubin parts deliver far more FLOPs per watt, so "20M H100-equivalents" of compute fits into far fewer, far more powerful physical chips. Translate GW to fleet at a realistic ~1.4 kW/accelerator (chip plus PUE overhead of ~1.25-1.35 for these liquid-cooled builds) and 1 GW of IT load is roughly 700K-1M modern accelerators. That is the unit to keep straight when a lab quotes you a gigawatt number.

Axis 2: Installed fleet and largest coherent cluster

A single coherent cluster (one fabric, low-latency, trains one model) is worth more than the same chip count scattered across regions. Here is what survived verification.

LabLargest coherent cluster (verified)Fleet notes
xAIColossus 1: ~200K H100/H200 + ~30K GB200, ~300 MWLargest fully operational, single-coherent cluster at the mid-2025 snapshot.
AnthropicProject Rainier: ~500K Trainium2 across multiple US data centers1M+ Trainium2 chips in use now to train and serve Claude; plus up to 1M TPUv7 Ironwood incoming.
Google DeepMind9,216-TPU coherent superpod (1.77 PB shared HBM, 9.6 Tb/s interconnect); scales to ~400K chips across 43 superpods via Jupiter networkingThe only player demonstrated at coherent multi-datacenter training.
OpenAIAbilene GB200 buildout (capped ~1.2 GW)2M+ chips under development across the Stargate program.
MetaUnverifiedCluster claims failed verification.

On raw per-chip maturity, Google's TPUv7 Ironwood lands within ~10% of Nvidia's GB200/Blackwell on peak FLOPs, memory, and bandwidth (4.6 PFLOPS dense FP8, 192 GB HBM3e, 7.4 TB/s). That matters because Google's TPU fleet underpins two labs' serving at once: its own (DeepMind) and Anthropic's.

Axis 3: Peak inference-serving capacity (the axis that maps to your purpose)

This is the soft axis, and the research is honest about why: no measured serving metric survived verification. Not tokens/sec, not concurrent-user ceilings, not p99 latency, not a documented throttling incident, for any lab. The serving-reliability comparison therefore rests on three proxies:

  1. Capacity online in 2026 (not announced for 2029). Favors Google DeepMind and Anthropic.
  2. Platform diversification (can you reroute around a failed platform?). Strongly favors Anthropic (Trainium + TPU + Nvidia). Google is single-platform but owns that platform end to end.
  3. Known incidents. The confirmed signal here is the June 2, 2026 Claude outage and Anthropic's 2026 usage limits, both pointing to demand outrunning supply industry-wide. Reporting also flags a broader 2026 compute crunch with rationing and rising GPU prices across the sector.

Net read for consistent service: Google DeepMind's vertical integration (silicon to serving) and Anthropic's platform diversity are the two structurally strongest positions. OpenAI carries the most serving risk today because it pairs the largest demand with the smallest online footprint. None of them is currently serving without strain.

Per-lab attribution (who is actually behind each "lab")

Since these are pure labs, the compute is really their backers' compute:

What did not survive verification (and why that matters)

Eight claims were killed in 3-vote adversarial verification. Treat these as unconfirmed, not disproven, but do not build a ranking on them:

The pattern: self-built and Chinese-lab capacity is the hardest to substantiate. Meta may well hold serious compute; the evidence set just could not confirm it.

What to watch (the open questions that would change the ranking)

  1. Real serving metrics. The entire reliability axis is built on proxies. The first lab to publish honest tokens/sec, concurrent-user ceilings, and p99 under load changes this analysis.
  2. Meta's true buildout. If Prometheus (~1 GW) and Hyperion (up to 5 GW) are real and on schedule, Meta jumps into the top tier.
  3. xAI's 2026 reality. Colossus 1 is confirmed; everything after it is not. Satellite-imagery analysts already dispute the Colossus 2 "1 GW" framing (Tom's Hardware put on-site cooling at ~350 MW).
  4. The training-versus-inference split. None of the contracted capacity is broken out by how much is reserved for training versus available for serving, or, for Google TPU, how it is split between DeepMind and external tenants like Anthropic. That split is what actually determines service consistency.

Bottom line

For consistent, top-level service delivery specifically, the order is Google DeepMind and Anthropic at the top (one by vertical integration, the other by diversification), OpenAI third (biggest ceiling, tightest present), xAI fourth (strong cluster, thin and unverified forward capacity), and Meta, DeepSeek, and Mistral unranked for lack of verifiable data.

The dominant caveat

But hold the dominant caveat in front of all of it: this is a mid-2026 snapshot of a market where the announced numbers are roughly 20-to-1 against the online ones, and where even the best-positioned lab took a capacity-driven global outage last week. The leaderboard you actually care about is "online and serving," and on that board, almost everyone is smaller than their headlines.

Sources

Independent / third-party (strongest):

First-party (directionally reliable, promotional, unaudited):

Secondary / contested (used with caution):

Research method: deep-research harness, 6 search angles, 25 sources fetched, 113 claims extracted, top 25 verified via 3-vote adversarial verification (17 confirmed, 8 killed). 108 agents, ~3.4M tokens. GW-to-fleet reconciliation applied via the sizing-data-centers skill.

Subscribe

Notes like this one also go out through factually, my newsletter. Subscribe at news.kentlangley.com, or point your reader at the RSS feed.

Founder OS · Published 2026-06-08 · Instance: factual · Project: content-engine/ai-lab-compute
Skills applied: designing-fos, writing-copy, navigating-skills, deep-research, sizing-data-centers
fos.kentlangley.com