Weekly Pricing Pulse

Editor's note

Across the September 21 to September 27 comparison window, no exact matched public on-demand GPU catalog price across the nine providers we track significantly changed. We compared 13,991 of this week’s 13,999 GPU price listings directly with last week’s prices for the same provider, region and GPU configuration. Five new AWS price listings had no prior match. Three Lambda listings no longer included a location, so we checked those GPU configurations separately and confirmed their prices were unchanged.

The visible catalog differences were small. The September 27 AWS capture added five price/SKU observations for the eight-GPU p6-b200.48xlarge in EU (Paris), one Linux entry and four operating-system variants. That is one instance family in one captured market, not five launches. The Linux line was $174.32 per instance-hour, or $21.79 per GPU-hour.

Lambda listed the same 22 GPU configurations at the same prices in both collections. Its availability feed reported fewer locations on September 27. The single-GPU A10, for example, remained listed in California at $1.29 per GPU-hour, while Virginia no longer appeared in its availability list. These are changes in reported availability at collection time, not new products or price changes.

We checked all nine providers in the September 27 catalog for the lowest comparable public on-demand prices: H100 SXM at $3.85 per GPU-hour from Nebius in eu-north1; H200 at $4.29 from Crusoe in Europe (Iceland); B200 at $6.69 from Lambda with region not published; and A100 SXM4 80 GB at $2.30 from Crusoe in US East (New York). These are the lowest observed catalog prices for those GPU groups, not live quotes or guarantees of capacity or availability.

Read of the week: Nebius pricing changes starting October 1

The change to watch is a provider notice, not a realized move in this catalog comparison. Nebius's official compute pricing page, rechecked September 29, lists new USD on-demand rates effective October 1, 2026 for the three tracked families: H100 from $3.85 to $4.50 per GPU-hour, H200 from $4.50 to $5.40, and B200 from $7.15 to $8.50. The September 27 Nebius normalized CSV was byte-identical to September 21, so this issue does not report those announced rates as observed catalog movement. The next Pulse will check whether a later capture reflects the posted change.

OCPI, the transaction-based Ornn Compute Price Index, moved separately between the September 21 and September 28 settlements at 4:00 PM Eastern. H100 SXM moved from $2.79 to $2.56 per GPU-hour, down approximately 8.2%; H200 from $5.07 to $4.84, down approximately 4.5%; A100 SXM4 from $0.99 to $0.91, down approximately 8.1%; and B200 from $7.71 to $8.11, up approximately 5.2%. The public history values are cent-rounded, so the percentage changes are approximate. OCPI is index context, not a provider quote or a public catalog observation.

The AIForge Works Weekly Pricing Pulse is informational and is not a live quote, capacity signal, availability guarantee, workload-specific recommendation or personalized account analysis. For workload-specific options, use Cloud Advisor. To inspect neighboring SKUs, regions or terms, open the Multi-Cloud Pricing Analyzer. To compare your own GPU spend with dated market evidence, use GPU Audit. Verify current price, capacity, availability, orderability, terms, fees and taxes directly with the provider before acting.

Ornn OCPI context

OCPI, the Ornn Compute Price Index, is a transaction-based benchmark, separate from the public on-demand catalog prices tracked by AIForge Works. These are the latest settled daily readings through September 28, 2026 at 4:00 PM Eastern (EDT). Read the API documentation and methodology.

B200 rose over the seven days to September 28; A100 SXM4, H100 SXM and H200 declined. Every percentage below is shown with its dated starting and ending price.

All readings are US dollars per GPU-hour. Seven-day change compares September 21 to 28, 2026; 30-day change compares August 29 to September 28, 2026. Percentage changes are approximate because the public history endpoint returns prices rounded to cents.

Lowest observed catalog price: The lowest comparable public on-demand price found in our September 27 catalog across the nine providers we track. We checked all 6,318,998 pricing rows, verified the matching GPU specifications and normalized eligible prices per GPU-hour. The result describes the saved catalog; it does not guarantee availability or a live quote.

Cohort qualifications: the Azure A100 row uses the 80 GB Standard_ND96amsr_A100_v4 at $4.10/GPU-hour; the 40 GB Standard_ND96asr_A100_v4 is excluded. The AWS H100 source identifies H100 80 GB but does not explicitly establish the SXM form factor, so that row retains a cohort-specific source qualification. In the retained evidence, some Lambda rows lack a published capacity region. The saved CoreWeave A100 label does not establish SXM4, so that entry is not included as a comparable price. OCPI and the observed catalog prices are not live quotes, capacity or orderability proof, or purchase recommendations.

Week at a glance

None of the matched public on-demand GPU prices rose or fell by more than 1%. The September 27 snapshot contains 13,999 normalized public on-demand GPU catalog observations across nine providers. They are public price lines, not capacity or orderability.

The Lambda counts reflect changes in reported regional availability. All 22 GPU configurations kept the same prices. Three configurations no longer had a location listed as available; the A10 continued to list California but no longer listed Virginia. Across all raw feeds, the current snapshot added a net 287 records: GCP +10, OCI +286, Lambda -1 and Vultr -8, while the other providers were flat.

Provider roll-up

*Lambda: The 3 restored and 4 newly absent counts reflect changes in reported locations. All 22 GPU configurations and their prices were unchanged; the A10 remained listed in California.

No reportable exact GPU catalog movement was found for Azure, GCP, OCI, CoreWeave, Crusoe, Nebius or Vultr.

Newly visible catalog observations

Five AWS price lines with regional evidence, representing one instance family in one market, not five launches. AWS P6 product source.

Lambda: prices unchanged, reported availability changed

Both collections used Lambda's live pricing and availability feed. Location details came from its list of regions with available capacity, rather than a separate website check. On September 27, the feed still listed all 22 GPU configurations at unchanged prices:

• A100 40 GB, eight GPUs: $1.99 per GPU-hour; Arizona was listed as available on September 21, with no location listed as available on September 27.
• RTX A6000, four GPUs: $1.09 per GPU-hour; North Texas was listed as available on September 21, with no location listed as available on September 27.
• A100 40 GB, one GPU: $1.99 per GPU-hour; Virginia was listed as available on September 21, with no location listed as available on September 27.
• A10, one GPU: $1.29 per GPU-hour; California remained listed as available, while Virginia no longer appeared.

These are dated availability reports, not product launches or retirements, and they do not establish availability today. Lambda API documentation.

Forward watch: the Nebius September 27 CSV was byte-identical to September 21, so its October 1 pricing notice is not a realized move in this Pulse. The announced changes are H100 $3.85 to $4.50, H200 $4.50 to $5.40 and B200 $7.15 to $8.50 per GPU-hour. AWS G7e Jakarta rows also remained present and unchanged: 786 raw rows and 102 normalized observations in both snapshots, with no event generated.

Evidence and methodology

Comparison dates: September 21 through September 27.

Pricing basis: Public on-demand whole-GPU observations are normalized to dollars per GPU-hour. Reserved, committed, spot, fractional-GPU, tax and provider-specific fee comparisons are outside these claims. Catalog presence does not establish orderability or capacity.

Coverage: All nine configured provider collections completed successfully. The run-bound supplement enumerates 106 AWS EC2 price-catalog locations, 89 GCP GPU zones in 34 parent regions, and 33 Vultr datacenter regions from exact sealed official-source responses. Those are deliberately narrower scopes: they do not claim every AWS or GCP infrastructure location, GPU availability at every Vultr site, current capacity, or orderability. OCI prices are global and use a configured 26-region expansion rather than all 45 official regions; CoreWeave pricing is global rather than availability-zone-specific; Lambda's public capacity response populated one of 14 official regions. Eight retained Azure mapping keys remain source identities, not verified public-region claims: eastus3, greececentral, israelnorthwest, portland, southcentralus2, southeastus, southwestus and swedensouth. No verified numeric egress rates or reserved/committed movement are asserted.

Sources: OCPI responses come from the Ornn public API and are interpreted under the Ornn methodology. AIForge’s current minimum-price check covers all nine saved September 27 provider files and binds the qualifying public on-demand whole-GPU prices to source rows and specifications. Earlier chart points retain their published source lineage. This does not establish provider live-site terms or orderability. The evidence excludes private quotes, negotiated discounts, spot/preemptible, reserved/committed and fractional-GPU pricing.

Next Pulse: Week of 2026-10-05