Coverage
What is stored right now: the coverage matrix by GPU and provider, every unresolved mapping and excluded SKU with its reason, the ingestion schedule and the last successful run.
Updated Markdown
Everything on this page is read from the database on every rebuild, so it is what is actually stored rather than what was true when the text was written. It covers the hyperscaler tier; the neocloud providers are listed with their status under the neocloud index.
Stored right now#
- Observations stored
- 9,582
- Months covered
- 58
- Earliest month
- Dec 2021
Last successful ingestion: AWS 2026-09-06 · AZURE 2026-09-06 · GCP 2026-09-06 · ORACLE 2026-09-06
Coverage matrix#
Availability of the canonical 8-GPU node, with the number of regions in the newest month and the month each provider series begins.
| GPU | AWS | Azure | Google Cloud | Oracle | Earliest usable history |
|---|---|---|---|---|---|
| H100 H100 80GB SXM5, eight-GPU node | Available 13 regions from Jul 2023 | Available 24 regions from Dec 2023 | Available 41 regions from Feb 2024 | Available 45 regions from Aug 2026 | Jul 2023 |
| B200 B200 180GB HGX, eight-GPU node | Available 4 regions from Jun 2025 | Not offered Azure publishes GB200 NVL, a rack-scale Grace-Blackwell system, but no standalone eight-GPU B200 node. Blending the two would compare different machines. | Available 5 regions from Jun 2026 | Available 45 regions from Aug 2026 | Jun 2025 |
Unresolved mappings#
Where a mapping cannot be confirmed against provider documentation, the offering is kept out of every published aggregate and recorded here until it is resolved. Nothing is guessed into a benchmark.
| SKU | Provider | Reason |
|---|---|---|
| Standard_ND128isr_NDR_GB200_v6 | Azure | Grace-Blackwell rack-scale system. Microsoft publishes no accelerator count GPUQuant could verify, and it is not an eight-GPU B200 node in any case. Excluded rather than guessed. |
| Standard_ND128isrf_NDR_GB200_v6 | Azure | A flex variant of the GB200 NVL system. Excluded for the same reason. |
| B109479 | Oracle | L40S is a graphics and inference part, not a tracked training GPU. |
| B109480 | Oracle | A second H100 part at $10.75 in the separate "Compute - GPU - Other" category, with no description and no shape GPUQuant can tie it to. Excluded. Third-party sites quote this as the OCI H100 price; the eight-GPU shape bills against B98415 at $10.00. |
| B109485 | Oracle | AMD Instinct, not NVIDIA. GPUQuant tracks NVIDIA datacenter GPUs only. |
| B110979 | Oracle | Grace-Blackwell rack-scale, not an eight-GPU HGX B200 node. Excluded for the same reason as Azure GB200 NVL. |
| B111758 | Oracle | AMD Instinct, not NVIDIA. |
| B112140 | Oracle | Grace-Blackwell Ultra rack-scale. Not a tracked model. |
| B112237 | Oracle | Blackwell Ultra. Not a tracked model. |
| B112613 | Oracle | Workstation-class Blackwell, not a tracked datacenter GPU. |
| B88517 | Oracle | P100 generation. Predates every tracked model. |
| B88518 | Oracle | P100 generation, single GPU. Predates every tracked model. |
| B93544 | Oracle | A100 40GB generation. Excluded: half the memory of the tracked A100 variant. |
| B89734 | Oracle | V100 generation. Predates every tracked model. |
| B95909 | Oracle | A10 is a graphics and inference part, not a tracked training GPU. |
| B92740 | Oracle | A100 40GB generation. Excluded: half the memory of the tracked A100 variant. |
Excluded SKUs#
Every offering a provider publishes for a tracked GPU that is not the canonical variant, with the reason it is out. The variant rules are on Normalization.
AWS30 excluded
AWS- dl1 Habana GaudiIntel Habana silicon, not NVIDIA.
- dl2q Qualcomm AI 100Qualcomm silicon, not NVIDIA.
- f1 Xilinx FPGAFPGA, not a GPU.
- f2 AMD FPGAFPGA, not a GPU.
- g2 GRID K520Not tracked.
- g3 M60Not tracked.
- g3s M60Not tracked.
- g4ad AMD Radeon Pro V520AMD, not NVIDIA. The AWS GPU column is a count, not a vendor selector.
- g4dn T4 16GBNot tracked.
- g5 A10G 24GBNot tracked.
- g5g T4G 16GBNot tracked.
- g6 L4 24GBInference family, not tracked.
- g6e L40S 48GBL40S. Only AWS publishes it, so it cannot support a cross-provider benchmark.
- g6f L4 fractionalFractional GPU shapes, not tracked.
- g7 GraphicsGraphics family, not tracked.
- g7e Graphics, L40S classGraphics family, not tracked.
- gr6 L4 24GBInference family, not tracked.
- gr6f L4 fractionalFractional GPU shapes, not tracked.
- inf1 AWS InferentiaAWS silicon, not NVIDIA.
- inf2 AWS Inferentia2AWS silicon, not NVIDIA.
- p2 K80Previous generation, not tracked.
- p3 V100 16GBPrevious generation, not tracked.
- p3dn V100 32GBPrevious generation, not tracked.
- p4d A100 40GB SXM4A100 40GB. Excluded: half the memory of the tracked variant.
- p6-b300 B300 HGXBlackwell Ultra, 2144 GiB accelerator memory across eight GPUs. Recorded, not part of the v1 model set.
- p6e-gb200 GB200 NVLGrace-Blackwell rack-scale system, not an eight-GPU B200 node.
- trn1 AWS TrainiumAWS silicon, not NVIDIA.
- trn1n AWS TrainiumAWS silicon, not NVIDIA.
- trn2 AWS Trainium2AWS silicon, not NVIDIA.
- vt1 Xilinx video transcodeVideo transcoding accelerator, not a GPU.
Azure13 excluded
Azure- Standard_NC16as_T4_v3 T4 16GBNot tracked.
- Standard_NC24ads_A100_v4 A100 80GB PCIeNC A100 v4 is the PCIe part with no SXM NVLink fabric. Excluded as a different variant.
- Standard_NC40ads_H100_v5 H100 NVL 94GBNCads H100 v5 is H100 NVL with 94GB. Different capacity, excluded.
- Standard_NC48ads_A100_v4 A100 80GB PCIePCIe A100, excluded.
- Standard_NC4as_T4_v3 T4 16GBNot tracked.
- Standard_NC64as_T4_v3 T4 16GBNot tracked.
- Standard_NC80adis_H100_v5 H100 NVL 94GBH100 NVL 94GB, excluded.
- Standard_NC8as_T4_v3 T4 16GBNot tracked.
- Standard_NC96ads_A100_v4 A100 80GB PCIePCIe A100, excluded.
- Standard_NCC40ads_H100_v5 H100 NVL 94GB confidentialNCCads H100 v5, confidential computing, one H100 NVL 94GB. Excluded.
- Standard_ND128isr_NDR_GB200_v6 GB200 NVLGrace-Blackwell rack-scale system. Microsoft publishes no accelerator count GPUQuant could verify, and it is not an eight-GPU B200 node in any case. Excluded rather than guessed.
- Standard_ND128isrf_NDR_GB200_v6 GB200 NVLA flex variant of the GB200 NVL system. Excluded for the same reason.
- Standard_ND96asr_A100_v4 A100 40GB SXM4ND A100 v4 is the 40GB part. Excluded from the 80GB benchmark.
Google Cloud6 excluded
Google Cloud- A100 40GB A100 40GBHalf the memory of the tracked variant.
- Fractional vGPU Fractional vGPUA fraction of a GPU is not a whole-GPU rental.
- H100 80GB Mega H100 80GB MegaThe a3-megagpu networking tier, priced above standard H100.
- H100 80GB Plus H100 80GB PlusA separately priced H100 tier. Including it would give Google Cloud two H100 quotes in one region.
- Previous and inference generations Previous and inference generationsOutside the v1 model set.
- RTX PRO 6000 RTX PRO 6000Workstation-class part, not tracked.
Oracle0 excluded
OracleNothing excluded.
Refresh and validation#
Ingestion runs daily. AWS and Google Cloud are re-read for the newest month; Azure and Oracle are snapshotted, which is the only way their series can ever grow; the neocloud pages and feeds are read in the same run.
- Raw provider records are stored exactly as published, so any normalized figure can be re-derived from its source row.
- Writes are keyed on a deterministic identifier, so re-running an ingestion rewrites the same values instead of duplicating or destroying history.
- GPU counts are read from provider specification tables, not inferred from a size name. Each offering row links to the document its count came from.
- Every run records rows read, rows written and months covered, and the last successful run for each source is shown above.
- Aggregates carry a calculation version, currently v4, so a change to the maths does not make old rows unexplainable.