State of GPU Services

Inside this edition

This early-Q4 briefing uses supplier and GPU-model evidence captured 25–27 October 2026 and a strictly matched Q2-to-October pricing study. It is not a full-quarter market census. Candidate domains, confirmed provider domains, price-bearing domains, advertised configurations and installed capacity are different measurement units.

01 · The auditThe October audit in one view

The audit reviewed 872 candidate domains and excluded 150, leaving 722 active candidates with documented dispositions. Within those, 339 rows were explicitly labelled GPU service providers. A separate count of 216 retained domains had numeric or indicative publicly reported pricing.

Do not read 722 as distinct, confirmed GPU suppliers. The audit includes resellers, marketplaces, managed services, infrastructure-adjacent businesses and access-limited cases. It closed 79 manual-access exceptions without treating them as verified offers.

02 · Model evidenceThe GPUs found in public rate evidence

The largest variant-unspecified groups are H100, A100, RTX PRO 6000, L40S and H200. These counts come from the October native-price ledger, not all 339 confirmed provider domains. Providers can appear in more than one row, and neither a quotation nor a domain count proves stock or installed GPUs.

The July Q3-branded report used a different 588-operator directory population. Its published model counts are historical context, not a Q2-to-October growth series.

03 · PricingPublic pricing is not one price

In the narrow October subset explicitly denominated in native USD per GPU-hour, the observed median was $2.41 for H100, $3.84 for H200, $6.50 for B200 and $1.49 for A100. These small samples are descriptive, not a market-wide price index.

No blanket currency conversion or “price per GPU” normalization is appropriate across unlike products. Compare complete configurations under the same contract, region and term.

04 · Matched historyQ2 to October: what actually changed

The 60-offer private historical ledger recalculated to 41 unchanged prices (68.3%), 15 increases (25.0%) and four decreases (6.7%) across 10 provider domains. Every matched row passed validation and no arithmetic discrepancies were found.

05 · Capacity disciplineGPU quantities are not GPU capacity

The audit can count quantities explicitly stated inside distinct advertised configurations: 270 A100 positions across 75 configurations from 19 domains, and 161 H200 positions across 38 configurations from 20 domains. These are sums of posted configuration sizes, not unique chips.

Physical deployed stock, immediately orderable capacity, firm commitments and announced future GPU capacity are not established. A single underlying machine may support more than one published configuration.

The same 10-domain historical cohort also changed classification: two June “not GPU” entries were classified differently in October. That is a migration in verification method, not evidence of two new suppliers. None of the 60 strict matched-price rows contains a populated region field, so geographic change is not measurable from this cohort.

06 · FindingsFive findings worth retaining

• 01A large candidate list is not a supplier census. The review retained 722 active candidates, but only 339 rows were explicitly labelled GPU service providers.
• 02GPU-model evidence is substantial but uneven. Variant transparency differs, and domain counts do not establish inventory.
• 03Prices require the entire contract. GPU count, sold object, currency, term and billing unit determine comparability.
• 04The matched history is useful because it is narrow. Forty-one of 60 matched rates were unchanged, but the cohort does not describe the full market.
• 05Some important numbers are not measurable yet. Posted configurations are not fleet capacity, and broad market growth remains unestablished.