Weekly Pricing Pulse
Editor's note
Nebius put through the on-demand increases it had posted for October 1. Between our September 27 and October 4 catalog captures, those were the only verified public on-demand GPU price changes of more than 1% across the nine providers we track: 20 matched Nebius listings spanning single-GPU and eight-GPU configurations.
- H100: $3.85 to $4.50 per GPU-hour (eu-north1).
- H200: $4.50 to $5.40 per GPU-hour (eu-north1, eu-north2, eu-west1, us-central1).
- B200: $7.15 to $8.50 per GPU-hour (me-west1, us-central1).
- B300: $7.85 to $9.50 per GPU-hour (eu-west2, uk-south1, us-north1).
The increase changed the lowest observed H100 price across the nine providers we track. On October 4 it was Crusoe at $3.90 per GPU-hour, including US East (Virginia), compared with Nebius at $3.85 on September 27. The other three lowest observed prices were unchanged:
- H200: $4.29 per GPU-hour (Crusoe, Europe (Iceland)).
- B200: $6.69 per GPU-hour (Lambda, region not published).
- A100 SXM4 80 GB: $2.30 per GPU-hour (Crusoe, US East (Virginia)).
Of this week’s 14,001 GPU price listings, we compared 13,996 directly with September 27 prices for the same provider, region and instance. The other five were returning Lambda regional listings. These reflect reported availability at collection time, not new GPU products. No other verified price changes exceeded our 1% reporting threshold; the methodology section explains 17 Azure listings that are outside the per-GPU comparison.
Read of the week: Nebius’s October 1 rates appear in the catalog
Nebius’s published price schedule lists the four increases above as effective October 1. Our October 4 capture now reflects those rates. This follows last week’s watch item, which described the increases as announced but not yet present in the September 27 catalog.
OCPI, the transaction-based Ornn Compute Price Index, moved separately between the September 28 and October 4 settlements: H200 moved from $4.84 to $5.57 per GPU-hour (+15.1%); H100 SXM moved from $2.56 to $2.59 per GPU-hour (+1.2%); B200 moved from $8.11 to $7.71 per GPU-hour (-4.9%); A100 SXM4 moved from $0.91 to $0.88 per GPU-hour (-3.3%). This is a six-day index comparison. The public history values are rounded to cents, so percentages are approximate. OCPI is index context, not a provider quote or a catalog price.
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 settled daily readings through October 4, 2026.
| GPU family | September 28 | October 4 | 6-day change | September 4 | 30-day change |
|---|---|---|---|---|---|
| H100 SXM | $2.56 | $2.59 | +1.2% | $2.89 | -10.4% |
| H200 | $4.84 | $5.57 | +15.1% | $4.63 | +20.3% |
| B200 | $8.11 | $7.71 | -4.9% | $6.40 | +20.5% |
| A100 SXM4 | $0.91 | $0.88 | -3.3% | $1.06 | -17.0% |
All readings are US dollars per GPU-hour. 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 October 4 catalog across the nine providers we track.
OCPI and Lowest Observed Catalog Prices
H100 SXM
Public on-demand catalog prices for the same GPU group, normalized per GPU-hour. The lowest price may come from a different region or instance size than the offer you need.
OCPI-H100 SXM
Transaction-based - Ornn x ICE - settled October 4, 2026
$2.59/ GPU / hrLowest observed catalog price
Crusoe - US East (Virginia) · October 4 catalog
$3.90/ GPU / hrAbove OCPI
Catalog checked
Oct 4All 9 provider files scanned
Spread = catalog price minus the October 4 OCPI of $2.59/GPU-hour; percentage = spread divided by OCPI. Orange is above OCPI; green is below. Percentages are approximate because OCPI is rounded to cents. This compares catalog prices with an index, not region-matched quotes.
OCPI vs Lowest Observed Catalog Prices by Week
OCPI index prices alongside the lowest observed catalog price for each GPU. Historical points retain their dated comparison basis.
H100 SXM
H200
B200
A100 SXM4
The October 4 point pairs that day’s OCPI settlement with the October 4 catalog minimum. Earlier points retain their published evidence.
Cohort qualifications: The A100 comparison uses the 80 GB SXM4 cohort. The Azure A100 listing is 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 listing retains a cohort-specific source qualification. Some Lambda listings 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 observed catalog prices are not live quotes, proof of capacity or orderability, or purchase recommendations.
Region-label correction: Our September 28 issue labeled Crusoe’s US East region as New York. Crusoe’s location documentation identifies it as Virginia. Only the location label was corrected; the H100 SXM price remains $3.90 per GPU-hour and A100 SXM4 80 GB remains $2.30.
Week at a glance
We compared 13,996 of this week’s 14,001 GPU price listings directly with last week’s prices for the same provider, region and instance. 5 listings did not have an exact prior match; the details below explain those differences. 20 verified per-GPU prices changed by more than 1%.
Provider roll-up
| Provider | Class | Newly visible | Restored | Newly absent | Price moves >1% |
|---|---|---|---|---|---|
| Lambda | neocloud | 0 | 5 | 3 | 0 |
| Nebius | neocloud | 0 | 0 | 0 | 20 |
| Totals | 0 | 5 | 3 | 20 |
Lambda: the five restored listings reflect returning reported locations. The three newly absent entries previously had no published location; their A100 configurations now have named locations. These counts do not mean five product launches or three retirements.
No reportable exact GPU catalog movement was found for AWS, GCP, OCI, CoreWeave, Crusoe or Vultr.
17 Azure RTX PRO 6000 (NCv6) preview-size listings are outside the per-GPU comparison; see methodology.
Catalog changes in detail
Lambda: prices unchanged, reported availability changed
Both collections used Lambda’s live pricing and availability feed. All 22 GPU configurations kept identical prices and specifications. In the October 4 capture, Lambda reported regional availability for four configurations in six configuration-and-location combinations, up from one combination on September 27. Five regional listings returned:
- Lambda: A100 40 GB, one GPU: Arizona and Virginia returned at $1.99 per GPU-hour.
- Lambda: A100 40 GB, eight GPUs: Arizona returned at $1.99 per GPU-hour.
- Lambda: A100 80 GB, eight GPUs: Illinois returned at $2.79 per GPU-hour.
- Lambda: A10, one GPU: Virginia returned at $1.29 per GPU-hour.
These are dated availability reports, not product launches or retirements, and they do not establish availability today. Lambda API documentation.
Nebius: matched prices changed by more than 1%
- Nebius:
gpu-b300-sxm/1gpu-24vcpu-346gb(B300 288 GB, 1 GPU) in eu-west2: $7.85 to $9.50 per GPU-hour. - Nebius:
gpu-b300-sxm/8gpu-192vcpu-2768gb(B300 288 GB, 8 GPUs) in eu-west2: $7.85 to $9.50 per GPU-hour. - Nebius:
gpu-b300-sxm/1gpu-24vcpu-346gb(B300 288 GB, 1 GPU) in uk-south1: $7.85 to $9.50 per GPU-hour. - Nebius:
gpu-b300-sxm/8gpu-192vcpu-2768gb(B300 288 GB, 8 GPUs) in uk-south1: $7.85 to $9.50 per GPU-hour. - Nebius:
gpu-b300-sxm/1gpu-24vcpu-346gb(B300 288 GB, 1 GPU) in us-north1: $7.85 to $9.50 per GPU-hour. - Nebius:
gpu-b300-sxm/8gpu-192vcpu-2768gb(B300 288 GB, 8 GPUs) in us-north1: $7.85 to $9.50 per GPU-hour. - Nebius:
gpu-h200-sxm/1gpu-16vcpu-200gb(H200 141 GB, 1 GPU) in eu-north1: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-h200-sxm/8gpu-128vcpu-1600gb(H200 141 GB, 8 GPUs) in eu-north1: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-h200-sxm/1gpu-16vcpu-200gb(H200 141 GB, 1 GPU) in eu-north2: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-h200-sxm/8gpu-128vcpu-1600gb(H200 141 GB, 8 GPUs) in eu-north2: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-h200-sxm/1gpu-16vcpu-200gb(H200 141 GB, 1 GPU) in eu-west1: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-h200-sxm/8gpu-128vcpu-1600gb(H200 141 GB, 8 GPUs) in eu-west1: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-h200-sxm/1gpu-16vcpu-200gb(H200 141 GB, 1 GPU) in us-central1: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-h200-sxm/8gpu-128vcpu-1600gb(H200 141 GB, 8 GPUs) in us-central1: $4.50 to $5.40 per GPU-hour. - Nebius:
gpu-b200-sxm-a/1gpu-20vcpu-224gb(B200 180 GB, 1 GPU) in me-west1: $7.15 to $8.50 per GPU-hour. - Nebius:
gpu-b200-sxm-a/8gpu-160vcpu-1792gb(B200 180 GB, 8 GPUs) in me-west1: $7.15 to $8.50 per GPU-hour. - Nebius:
gpu-b200-sxm/1gpu-20vcpu-224gb(B200 180 GB, 1 GPU) in us-central1: $7.15 to $8.50 per GPU-hour. - Nebius:
gpu-b200-sxm/8gpu-160vcpu-1792gb(B200 180 GB, 8 GPUs) in us-central1: $7.15 to $8.50 per GPU-hour. - Nebius:
gpu-h100-sxm/1gpu-16vcpu-200gb(H100 80 GB, 1 GPU) in eu-north1: $3.85 to $4.50 per GPU-hour. - Nebius:
gpu-h100-sxm/8gpu-128vcpu-1600gb(H100 80 GB, 8 GPUs) in eu-north1: $3.85 to $4.50 per GPU-hour.
Forward watch
The next comparison will check whether Nebius’s October 1 rates persist and whether Lambda’s reported locations change again. We will add Microsoft’s current RTX PRO 6000 (NCv6) sizes to the comparison once our collection recognizes them, including lower North Europe list prices seen on October 4.
Evidence and methodology
Comparison dates: September 27 through October 4, 2026.
Pricing basis: Public on-demand prices for the same provider, region and GPU configuration are normalized to dollars per GPU-hour and rounded to cents. Spot/preemptible, reserved/committed, fractional-GPU, private quotes, negotiated discounts, taxes and provider-specific fees are outside these comparisons. Catalog presence does not establish orderability or capacity.
Vultr comparisons: In the October 4 capture, Vultr’s public plan catalog marks its H100, B200 and A100 SXM plans as not available on demand, so they are outside our on-demand comparisons. Its on-demand A100 plans do not state the SXM4 form factor, and no H200 plan appears in that capture. These are the reasons for “N/A” in all four comparisons. Missing location information is a separate limitation.
Collection methods and the prior-week recount
GCP: The October 4 collection correctly labels 3,748 pricing records as Spot and prevents 4,184 invalid commitment-price records from being synthesized from Spot prices. These are counts across the full GCP feed, not GPU price changes. In a controlled replay of the September 27 source, the correction relabeled 3,745 records and removed 4,178 synthesized commitment records, taking the GCP total from 32,209 to 28,031. On that corrected basis, October 4’s 28,078 records are an increase of 47; the raw-count decline of 4,131 below includes the method correction.
Azure: The October 4 collection separately labels 376,060 Windows and Cloud Services pricing records across on-demand, Spot, LowPriority and DevTest pricing. This includes 80,398 on-demand records (75,511 Windows and 4,887 Cloud Services) that are kept separate from the 81,368 base on-demand records. The correction deletes no Azure records and does not rewrite prices. A controlled replay of the September 27 source relabeled 350,230 records while preserving all 650,959 original records. These classification totals cover the entire feed, not just GPUs.
Last week’s count was recounted: We processed both saved weeks through the current Pulse comparison engine, with the Azure Linux-pricing review bound separately to each snapshot. September 27 still contains 13,999 comparable GPU price listings. The saved source files were not rewritten or recollected; their original raw-record totals are retained below. Recounting the listings is separate from the controlled collection-method replays described above.
Raw pricing records by provider
Raw records increased from 6,318,998 to 6,366,004, a net gain of 47,006. These counts cover all collected pricing records before the GPU comparison filters; they are not counts of GPU products or launches.
- AWS: 5,616,647 to 5,616,719 (+72).
- Azure: 650,959 to 702,049 (+51,090). This increase includes 20,318 records for Taiwan North (8,899), Taiwan Northwest (6,242) and Southeast Asia 3 in Malaysia (5,177), none present in last week’s snapshot. The remaining 30,772 is net growth elsewhere in the feed. The classification correction added no records.
- GCP: 32,209 to 28,078 (-4,131).
- OCI: 16,068 to 16,068 (unchanged).
- CoreWeave: 9 to 9 (unchanged).
- Crusoe: 24 to 24 (unchanged).
- Lambda: 22 to 24 (+2).
- Nebius: 144 to 144 (unchanged).
- Vultr: 2,916 to 2,889 (-27).
Coverage: All nine provider files were checked. Coverage describes the pricing catalogs captured for this issue, not every infrastructure location or current capacity. OCI prices use a configured regional expansion; CoreWeave lists global prices; Lambda’s reported locations come from its capacity-availability feed at collection time.
Sources: OCPI responses come from the Ornn public API and are interpreted under the Ornn methodology. The current minimum-price check scans all nine saved October 4 provider files and checks the public on-demand whole-GPU prices against specifications. Earlier chart points retain their dated evidence. August 3 is omitted because no verified catalog comparison was available that week.
Azure NCv6 sources: Microsoft’s preview-to-GA FAQ and current GPU size tables.
17 Azure listings are excluded from the per-GPU comparison. They are RTX PRO 6000 (NCv6) sizes from Microsoft’s preview. Nine are preview sizes Microsoft has replaced and no longer supports for new deployments; eight are smaller preview sizes priced the same as Microsoft’s fractional ¼- and ½-GPU sizes. Non-orderable and fractional-GPU listings are outside our per-GPU comparisons.