Weekly Pricing Pulse

About Pulse

The Weekly Pricing Pulse tracks changes in published GPU cloud pricing across AWS, Azure, GCP, OCI, CoreWeave, Lambda, Vultr, Nebius, and Crusoe. AIForge Works compares successive harvests 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

Public GPU prices held; regional catalog visibility shifted. The transaction benchmark was again moving this week.

Ornn's H100 SXM index saw the largest seven-day percentage move among the four tracked GPU families, settling on September 14 at $2.75 per GPU-hour, approximately 13.2% below its September 7 reading. B200 moved the other way: $7.19 per GPU-hour, approximately 6.5% higher over the same period. A100 SXM4 slipped approximately 2.9%, while H200 edged up by just under 1%.

Across 13,873 matched public on-demand GPU observations, we found no price changes above the 1% reporting threshold. The four tracked catalog floors remain H100 SXM at $3.85, H200 at $4.29, B200 at $6.69, and A100 SXM4 at $2.30 per GPU-hour. Azure's complete saved-offer review also found no comparable price movement.

The six newly observed catalog entries tell a smaller story than the count might suggest: five operating-system variants of AWS's existing B200 instance in one newly observed region, Stockholm, plus one Lambda H100 SKU with no published region. These are six catalog entries, not six new GPU products.

Elsewhere, GCP's existing A3 High model appeared in three additional regions. Lambda's eight-GPU A100 40 GB offer returned in Arizona and California but was no longer observed in Virginia. Its different eight-GPU A100 80 GB SXM4 offer remained in Virginia while disappearing from the Illinois observations. Those are different configurations, not contradictory findings. Across all five newly absent regional observations, every affected model remained visible elsewhere or without a published region; none establishes a product retirement.

Following last week's B300 coverage, this week's saved-data comparison found no B300 additions, newly absent observations or comparable price changes.

For buyers, compare the GPU family and configuration the workload needs, then confirm region, purchase terms and availability. Lambda's B200 floor includes an offer whose region is not published. These are catalog leads to verify, not executable quotes; neither index movement nor catalog presence proves available capacity.

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 14, 2026 at 4:00 PM Eastern (EDT), retrieved September 14 at approximately 10:27 PM Eastern (EDT). Ornn methodology and API documentation.

H100 SXM and A100 SXM4 declined over the seven days to September 14, while B200 and H200 rose. Those transaction-index movements occurred alongside unchanged comparable public-catalog prices in this week's analysis.

All readings are in US dollars per GPU-hour. Both dates are September 2026 settlements at 4:00 PM Eastern (EDT). The 7-day OCPI change is (September 14 OCPI ÷ September 7 OCPI − 1) × 100, not OCPI versus a catalog floor. Percentage changes are approximate because the public history endpoint rounds readings to cents; full-precision latest values are retained in the source record.

Comparison dates: The September 7 issue displayed September 8 OCPI readings. This table uses the September 7 readings for a true seven-day comparison, so its percentages are not changes from the readings displayed in that prior issue.

Did our catalog floors change? No. The complete saved-data recalculation, including Azure, produced the same four tracked floor prices and winning providers as last week. Stable floors do not mean stable catalog coverage.

The current floors are recalculated from the saved September 14 files, not copied from the prior issue. B200's Lambda region is not fully published. Floor comparison summary.

Comparison coverage. All 27 displayed provider/cohort prices match their September 7 recorded values. Azure's complete exact-cohort review covers 24 H100 SXM regions, 31 H200 regions and seven A100 SXM4 regions, with respective floors of $12.29, $10.60 and $3.3996 per GPU-hour. No otherwise-eligible Azure record remains excluded for missing capture-time evidence. The established B200 chart excludes GB200 systems and has no Azure B200 target; this is an explicit cohort boundary, not a missing Azure import. Azure evidence summary.

AWS H100 retains the source limitation that its hardware description does not literally state SXM; Lambda offers with incomplete region identity are not global regional floors. Historical chart points remain unchanged, with August 3 omitted as in last week's graph. The final point pairs September 14 catalog data with the September 14 OCPI settlement; last week's frozen point pairs September 7 catalog data with September 8 OCPI. These charts are not an exhaustive market search, performance comparison, executable quote, or availability guarantee.

Sources: Ornn latest daily index, OCPI methodology, and the OCPI source summary, including captured-response checksums and comparison dates.

Week at a glance

• Market benchmark: H100 SXM fell approximately 13.2% over seven days; B200 rose approximately 6.5%. These OCPI movements are separate from public-catalog movements.
• On-demand GPU prices: No matched observations moved above 1%. All four tracked catalog floors are unchanged.
• Catalog visibility: 15 observations became visible and five exact observations became newly absent. The standard lookback categories divide the additions into six newly observed entries and nine restored / expanded entries—not six new GPU technologies.
• Azure: All 1,256 qualified public on-demand GPU observations match the prior snapshot, with no additions, absences or reportable price changes.

AWS, GCP, Lambda and Vultr account for this week's catalog changes. All nine providers have comparable GPU observations in both snapshots.

Provider roll-up

No reportable exact catalog movement was found for Azure, OCI, CoreWeave, Crusoe or Nebius. Azure is included in this conclusion, not withheld.

The standard New Silicon column names observations not seen under the provider/SKU lookback rule. Here it comprises five AWS operating-system variants in Stockholm and one Lambda catalog identity with no published region. These are not new accelerator launches. Five of the 15 additions have Region not published; they remain in the counts but do not support geographic-expansion claims.

The underlying source feeds grew by 1,690 rows: AWS +1,385, OCI +312, GCP +3, Lambda +3 and Vultr −13; the other provider row counts were unchanged. These raw-feed changes are not price changes.

Catalog changes this week

Catalog changes track exact provider + region + instance/SKU observations between harvests. They are not confirmed capacity launches or retirements unless provider-published evidence establishes that separately.

New Silicon

Six entries fall in the standard newly observed provider/SKU category: AWS (5) and Lambda (1). AWS's p6-b200.48xlarge was already an observed model; its five Stockholm catalog identities were not present in the September 7, August 30 or August 24 exact-region comparisons. Lambda's gpu_1x_h100_sxm5 entry has no published region and makes no geographic-expansion claim.

Restored / expanded visibility

Nine entries belong to the established restored / expanded visibility category: GCP (3), Lambda (5) and Vultr (1).

GCP's a3-highgpu-8g appears in asia-northeast3, australia-southeast2 and europe-west9. The model existed elsewhere; the older snapshots do not establish these exact regions as restorations. Lambda's Arizona and California gpu_8x_a100 observations are supported exact regional restorations. The remaining four entries in this category have no published region and are catalog-visibility observations only.

Newly absent exact catalog observations

Five exact observations became newly absent: Lambda (3) and Vultr (2). Every affected model remains in the current catalog in another region or with its region unpublished. None is described as a product retirement.

Top on-demand movers by provider

No comparable public on-demand GPU observations met the greater-than-1% reporting threshold. The finer matched-price check also found no changes at its numerical tolerance of $0.000000001 per GPU-hour. This conclusion covers the matched saved observations, not every possible provider quote or purchase term.

Methodology

Evidence through. The catalog snapshot and OCPI settlement are dated September 14, 2026. This is the September 14 issue, assembled and published on September 15; the issue date does not imply it was published Monday.

What we compared. The September 14 and September 7 saved Harvest snapshots cover the same nine providers. August 30 and August 24 material is used only to distinguish newly observed entries from restored or expanded visibility—not to calculate this week's price moves. The OCPI table separately compares the September 7 and September 14 settled readings.

How catalog prices are compared fairly. The standard comparison divides the instance hourly price by its verified GPU count, preserving provider, region, configuration, and comparable purchase terms. The public-catalog reporting threshold is a move greater than 1%. Providers without qualified comparable offers in both snapshots must be excluded explicitly, not assigned a zero-change result.

Commercial scope. As in the September 7 issue, government observations remain in the source totals but are excluded from buyer-facing commercial changes. There are 656 such observations in each week: 594 AWS GovCloud and 62 Azure Government. Commercial totals, including entries with unpublished regions, are 13,232 current and 13,222 prior. Unpublished-region entries increased from 18 to 23; these are not claims of geographic expansion or global availability.

Scope and limits. All OCPI prices are US dollars per GPU-hour. Latest levels use the higher-precision daily response; approximate percentage changes use cent-rounded public history consistently at both endpoints. Paired historical chart points retain the OCPI and catalog snapshot dates in their frozen issue records. Index values, catalog presence, and listed prices are not live quotes or availability guarantees.

This week's coverage. The saved snapshot contains 6,316,366 raw provider rows. After the existing GPU/spec and purchase-term rules, the comparison contains 13,888 current public on-demand GPU observations, against 13,878 prior; 13,873 are matched. The application separately reports 925,553 normalized pricing rows and 113,296 all-in catalog instance rows—those are not GPU-only counts. Existing unresolved RTX PRO 6000 server-edition labels remain excluded from catalog-change rendering; none of this week's additions or absences is affected. Comparison evidence summary.

Azure evidence recovery. All 652 saved successful response pages were examined. Of 2,507 GPU candidates, 1,256 qualified as public base-compute offers; 1,251 Windows/Win variants were excluded. For 180 qualified offers, an authentic acquisition interval supplies effective-date evidence. The original parent per-page capture timestamps and successful raw seal remain absent; none was invented. All candidate decisions are resolved for this saved-data analysis, without rerunning Harvest. This recovered analytical evidence is not represented as an original successful run seal. Recovery and classification record.

Next Pulse: Week of September 21, 2026.