Weekly Pricing Pulse

About Pulse

The Weekly Pricing Pulse tracks what changed in published GPU cloud pricing across AWS, Azure, GCP, OCI, CoreWeave, Lambda, Vultr, Nebius, and Crusoe over the last week. AIForge Works compares each new harvest with the prior verified harvest and highlights buyer-relevant movement: comparable on-demand price changes, New Silicon, restored / expanded visibility, newly absent exact provider + region + instance/SKU observations, and likely price/SKU rolls.

This is not a live quote, workload-specific recommendation, or personalized account analysis. For workload-specific recommendations, use Cloud Advisor. To inspect neighboring SKUs, regions, or commitment terms, open the Multi-Cloud Pricing Analyzer. To compare your own GPU spend against a comparable floor, use GPU Audit.

Editor's note

Comparable public on-demand per-GPU rates were effectively flat again this week. The transaction benchmark was not.

Across the nine providers we track, no directly comparable public on-demand per-GPU rate moved more than our 1% reporting threshold. The Ornn × ICE transaction-price index (OCPI) rose across all four tracked families over the seven days through September 8: H200 increased 8.9% to $4.91 per GPU-hour, H100 SXM increased 8.8% to $3.08, B200 increased 5.0% to $6.72, and A100 SXM4 increased 2.0% to $1.02. These OCPI readings are one day newer than the September 7 catalog snapshot and are separate from the catalog-movement result.

For H200, the index now sits above the lowest eligible public offer in our reviewed set, at $4.91 versus Crusoe's $4.29 in Europe (Iceland). B200 is roughly level, at $6.72 versus Lambda's $6.69 listing. Lambda does not publish a complete region for that offer, so these lower catalog rates are leads to verify, not executable quotes.

Catalog changes were visibility changes, not comparable week-over-week repricing above 1%. Twelve observations became newly visible, five returned across Lambda and Vultr, and six exact observations across those same providers were no longer visible. AWS accounted for ten of the newly visible observations: five operating-system variants each of the eight-GPU p6-b200.48xlarge in Hyderabad, from $20.08 per GPU-hour, and p6-b300.48xlarge in Jakarta, from $28.87. Azure supplied the other two: one- and two-GPU NCads H100 v5 configurations in UK West, both normalized to $9.14 per GPU-hour. The same NCads H100 configurations remained listed in East US at $6.98 per GPU-hour in the same snapshot. UK West was therefore $2.16, or 30.9%, higher per GPU-hour. This is a same-snapshot regional price difference, not a weekly price move.

These catalog observations do not prove a launch, added capacity, retirement, orderability, or current availability. Verify the exact GPU, region, purchase terms, and availability before treating any listing as a buying option.

Ornn OCPI context

The broader transaction benchmark moved up while the comparable public on-demand catalog cohort showed no movement over 1%. Over the seven days from September 1 through September 8, all four tracked OCPI families rose; H200 and H100 SXM posted the largest changes.

OCPI, the Ornn Compute Price Index, is a transaction-based benchmark that is separate from the public on-demand catalog prices tracked elsewhere in this Pulse. The latest settled daily readings below run through September 8, 2026 at 20:00 UTC and were captured from Ornn on September 9. The paired catalog floors use the separately certified September 7 Harvest snapshot, so the benchmark is one day newer than the catalog evidence. Ornn methodology and API documentation.

For every provider row, the displayed catalog value is the lowest eligible reviewed public offer found for that exact OCPI GPU cohort across the regions covered by the September 7 harvest. It is not a floor across GPU families, an availability guarantee, or an exhaustive market search. Missing evidence does not mean a provider lacks that GPU.

The comparison includes AWS and Azure only where exact-cohort evidence qualifies. Azure is absent from B200 when its reviewed evidence is GB200 rather than an exact B200 match. The AWS H100 source identifies H100 80 GB hardware but does not literally state the SXM form factor, so that row retains a source note. Lambda offers without complete region identity are not presented as global regional floors.

The final trend point pairs the September 7 catalog snapshot with the September 8 settled OCPI close; exact source dates remain recorded in the frozen artifact and visible in the comparison cards.

Week at a glance

No provider had comparable on-demand offers over the 1% movement threshold. AWS, Azure, Lambda, and Vultr changed catalog visibility through New Silicon, restored / expanded visibility, and newly absent exact catalog observations, shown below.

These movement counts apply only to comparable public-catalog on-demand prices. OCPI is a separate transaction-based market index and is shown independently.

• On-demand pricing was quiet this week. We did not find any comparable public on-demand GPU price changes above the 1% reporting threshold; the visible movement is catalog visibility.

Provider roll-up

No reportable exact catalog movement was certified this week for GCP, OCI, CoreWeave, Crusoe, or Nebius.

Underlying provider feeds also shifted: Azure added 648 raw catalog records and four normalized GPU offers. Two were in Azure Government and remain in the source evidence, but they are excluded from this buyer-facing commercial comparison because access eligibility and pricing are separate from Azure's commercial catalog. The two commercial additions are the UK West H100 observations described above. After normalizing to comparable GPU offers, purchase terms, and verified hardware specs, those feed-level changes did not produce a public on-demand GPU price move above 1% in this Pulse.

Offers seen in earlier verified harvests returned to view this week: Lambda (4) and Vultr (1). Because these provider and SKU combinations appeared in earlier verified harvests, we treat them as restored catalog visibility rather than brand-new silicon launches.

Catalog changes this week

Catalog change counts summarize week-over-week visibility changes in the compared harvests. They show provider + region + instance/SKU observations that became newly visible or no longer visible between harvests. We do not describe these as confirmed capacity launches or retirements unless provider-published sources say so.

New Silicon

12 New Silicon GPU catalog entries were detected across AWS (10) and Azure (2). Entries are grouped by provider so high-volume catalog churn does not hide smaller-provider changes.

New Silicon means a provider + region + instance/SKU observation newly visible in this harvest that is not paired with a same-provider, same-geo, same-instance no-longer-visible observation. It still does not by itself prove net-new capacity or a brand-new accelerator family.

Restored / expanded visibility

5 previously seen public GPU catalog observations were visible again this week across Lambda (4), Vultr (1).

Each provider and SKU had appeared in an earlier verified comparison; the region shown is where the offer was visible this week. We classify these as restored or expanded catalog visibility, not a new hardware launch, proof of added capacity, or confirmation that the offer is currently orderable.

Newly absent exact catalog observations

6 newly absent exact provider + region + instance/SKU GPU catalog entries were detected across Lambda (4), Vultr (2). Entries are grouped by provider so high-volume catalog churn does not hide smaller-provider changes.

A newly absent exact catalog observation means the provider + region + instance/SKU combination was present in the prior harvest but absent from this one in this week's public-catalog comparison, even when a same-instance replacement looks likely. Multiple regional observations can refer to the same SKU string; absence here does not prove the SKU is unavailable to buy.

Top on-demand movers by provider

No comparable offers met the 1% movement threshold for this section.

Methodology

Evidence through. Public-catalog pricing evidence in this issue runs through September 7, 2026. The separately captured OCPI context runs through September 8 at 20:00 UTC. These evidence dates are recorded separately from the publication date shown in the article header.

What we compared. This Pulse compares the September 7 harvest with the prior harvest collected August 30 for the August 31 Pulse cycle. Both harvests cover the same 9 cloud providers. The August 24 archive is used only to distinguish genuinely new entries from offers that disappeared and later returned; it is not used to calculate the weekly price movement.

How we compare prices fairly. Different providers package GPUs differently: some sell single-GPU instances, others sell 8-GPU instances. To compare apples-to-apples, we divide every instance price by the number of GPUs on that instance. All movement percentages in the public-catalog sections are based on that per-GPU hourly rate, not the raw instance price.

What counts as a real move. We only flag comparable public on-demand price changes greater than 1%. Anything smaller is treated as rounding noise and excluded. Providers with zero comparable GPU offers in either harvest are excluded from GPU price-movement conclusions. New Silicon means a provider + region + instance/SKU observation newly visible in this harvest that is not paired with a same-provider, same-geo, same-instance no-longer-visible observation and has not been seen in the lookback window. Restored / expanded visibility means the provider and instance/SKU appeared in an earlier verified harvest and is visible again, sometimes in additional regions. A newly absent exact catalog observation means the exact provider + region + instance/SKU combination was present in the prior harvest but absent from this one; multiple regional observations can share one SKU string.

Scope and limits. The generated Pulse diff certifies on-demand movement only. It does not certify reserved or committed-term movement. This issue's buyer-facing catalog-change section excludes government and sovereign cloud regions because access eligibility and pricing are separate from commercial provider catalogs. Those records remain in the source evidence and underlying normalized-offer totals. All prices are in US dollars. Catalog entries missing a clean price or verified GPU specs are dropped before comparison. This Pulse compiles publicly published prices and is not a live quote, availability guarantee, or workload-specific advice.

This week's coverage. The September 7, 2026 harvest completed across 9 providers. After normalizing 6,314,676 raw catalog price observations against verified hardware specs, this Pulse identified 13,878 current public on-demand GPU offers. The two Azure Government observations remain in that underlying total but are excluded from the 12 buyer-facing New Silicon observations. Of the current offers, 13,859 exact provider + region + instance/SKU observations were present in both harvests and eligible for direct movement comparison. The separate August 24 history review verified that both newly visible commercial Azure entries are new rather than returning.

Next Pulse: Week of 2026-09-14.