{ "spec_version": "0.1", "generated_from": "2026-09-10", "entry_count": 70, "providers": [ "AWS", "CoreWeave", "Crusoe", "DataCrunch (rebranded to Verda, verda.com)", "FluidStack", "Google Cloud", "Hyperstack", "Lambda", "Microsoft Azure", "Nebius", "OVHcloud", "Oracle Cloud (OCI)", "Scaleway", "Together AI (GPU Clusters)", "Vast.ai (live API)", "Voltage Park" ], "entries": [ { "sku": "p5.48xlarge", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink/NVSwitch, 900 GB/s NVSwitch, 3.6 TB/s bisectional bandwidth per instance (AWS P5 page)", "inter_node_fabric": "EFA, 3200 Gbps per instance (AWS P5 page)", "max_contiguous_gpus": 20000, "regions": [ "us-east-1" ], "price_usd_per_gpu_hr": 6.88, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20(N.%20Virginia)/Linux/index.json", "spec_confidence": "published", "notes": "Per-instance $55.04/hr (Linux, us-east-1, official AWS on-demand pricing feed that backs aws.amazon.com/ec2/pricing/on-demand/) divided by 8. AWS P5 page: UltraClusters 'enable scaling up to 20,000 H100 or H200 GPUs'.", "provider": "AWS", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink/NVSwitch, 900 GB/s NVSwitch, 3.6 TB/s bisectional bandwidth per instance (AWS P5 page))", "scale-out ~400 Gbps/GPU", "contiguity 20000 GPUs" ] }, { "sku": "p5en.48xlarge", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink/NVSwitch, 900 GB/s NVSwitch (AWS P5 page); 1128 GB HBM3e total", "inter_node_fabric": "EFAv3 (Nitro v5), 3200 Gbps per instance, 'up to 35% improvement in latency' vs earlier EFA (AWS P5 page)", "max_contiguous_gpus": 20000, "regions": [ "us-east-1" ], "price_usd_per_gpu_hr": 7.912, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20(N.%20Virginia)/Linux/index.json", "spec_confidence": "published", "notes": "Per-instance $63.296/hr divided by 8. p5e.48xlarge has no on-demand entry in the us-east-1 feed; p5en is the H200 SKU with on-demand list pricing.", "provider": "AWS", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink/NVSwitch, 900 GB/s NVSwitch (AWS P5 page); 1128 GB HBM3e total)", "scale-out ~400 Gbps/GPU", "contiguity 20000 GPUs" ] }, { "sku": "p6-b200.48xlarge", "accelerator": "B200", "vram_gb": 179, "form_factor": "SXM (NVLink baseboard)", "gpus_per_node": 8, "intra_node": "NVLink, 'up to 14.4 TB/s of total bidirectional NVLink bandwidth' (AWS P6 page)", "inter_node_fabric": "EFAv4, 3.2 Tbps per instance (AWS P6 page)", "max_contiguous_gpus": null, "regions": [ "us-east-1" ], "price_usd_per_gpu_hr": 14.2416, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20(N.%20Virginia)/Linux/index.json", "spec_confidence": "published", "notes": "Per-instance $113.9328/hr divided by 8. vram_gb = AWS-published 1,432 GB HBM3e total / 8 = 179 GB. UltraCluster claim is non-numeric ('tens of thousands of GPUs').", "provider": "AWS", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink, 'up to 14.4 TB/s of total bidirectional NVLink bandwidth' (AWS P6 page))", "scale-out ~400 Gbps/GPU", "cluster product claim (contiguity unstated)" ] }, { "sku": "p6-b300.48xlarge", "accelerator": "B300 (Blackwell Ultra)", "vram_gb": 268, "form_factor": "SXM (NVLink baseboard)", "gpus_per_node": 8, "intra_node": "NVLink (Blackwell generation, per AWS P6 page)", "inter_node_fabric": "EFAv4, 6.4 Tbps per instance (AWS P6 page)", "max_contiguous_gpus": null, "regions": [ "us-east-1" ], "price_usd_per_gpu_hr": 17.802, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://b0.p.awsstatic.com/pricing/2.0/meteredUnitMaps/ec2/USD/current/ec2-ondemand-without-sec-sel/US%20East%20(N.%20Virginia)/Linux/index.json", "spec_confidence": "published", "notes": "Per-instance $142.416/hr divided by 8. vram_gb = AWS-published 2,144 GB HBM3e total / 8 = 268 GB.", "provider": "AWS", "as_of": "2026-09-10", "accelerator_canonical": "B300", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~800 Gbps/GPU", "contiguity unstated" ] }, { "sku": "a3-highgpu-8g", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink (nvidia-h100-80gb, 640 GB total per Google machine table)", "inter_node_fabric": "gVNIC with GPUDirect-TCPX, 1,000 Gbps max per VM, 5 NICs (Google accelerator-optimized machines doc table)", "max_contiguous_gpus": null, "regions": [ "us-central1", "us-east4" ], "price_usd_per_gpu_hr": 11.0613, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://gcloud-compute.com/a3-highgpu-8g.html", "spec_confidence": "secondary", "notes": "Per-VM $88.49/hr us-central1 divided by 8 ($88.5558 us-east4). Price from gcloud-compute.com (mirror of Google Cloud Billing data, updated 2026-09-06) because Google's official pricing pages are client-side rendered and returned no figures. Networking/GPU specs are published in Google docs. On-demand availability for A3 High 8-GPU confirmed in Google's consumption-options table.", "provider": "Google Cloud", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "a3-megagpu-8g", "accelerator": "H100-SXM (nvidia-h100-mega-80gb)", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink (640 GB total GPU memory per Google machine table)", "inter_node_fabric": "gVNIC with GPUDirect-TCPXO, 1,800 Gbps max per VM, 9 NICs — 'doubles the network bandwidth when compared to the A3 High' (Google doc)", "max_contiguous_gpus": null, "regions": [ "us-east4" ], "price_usd_per_gpu_hr": 11.6172, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://gcloud-compute.com/a3-megagpu-8g.html", "spec_confidence": "secondary", "notes": "Per-VM $92.9376/hr us-east4 divided by 8 (us-central1 not listed for this type). Price from gcloud-compute.com Billing-data mirror; specs from Google docs. On-demand availability confirmed in Google's consumption-options table.", "provider": "Google Cloud", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "a3-ultragpu-8g", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink (1,128 GB total GPU memory per Google machine table)", "inter_node_fabric": "RDMA over Converged Ethernet (RoCE) via ConnectX-7, 3,600 Gbps max per VM, 10 NICs (Google doc table)", "max_contiguous_gpus": null, "regions": [], "price_usd_per_gpu_hr": null, "pricing_model": null, "min_commit": null, "notice_period": null, "source_url": "https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines", "spec_confidence": "published", "notes": "No on-demand price exists: Google's consumption-options table marks A3 Ultra (H200) On-demand = No; available only via Spot, Flex-start, and reservation modes. Specs are published; offering recorded without price per hard rules.", "provider": "Google Cloud", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "a4-highgpu-8g", "accelerator": "B200", "vram_gb": 180, "form_factor": "SXM (NVLink baseboard)", "gpus_per_node": 8, "intra_node": "NVLink 1,800 GBps bidirectional per GPU, up to 14.4 TBps aggregate all-to-all (Google doc); 1,440 GB total GPU memory", "inter_node_fabric": "RoCE via NVIDIA ConnectX-7 NICs on Google's four-way rail-aligned network, 3,600 Gbps max per VM, 10 NICs (Google doc)", "max_contiguous_gpus": null, "regions": [], "price_usd_per_gpu_hr": null, "pricing_model": null, "min_commit": null, "notice_period": null, "source_url": "https://docs.cloud.google.com/compute/docs/accelerator-optimized-machines", "spec_confidence": "published", "notes": "No on-demand price exists: Google's consumption-options table marks A4 (B200) On-demand = No; Spot/Flex-start/reservations only. A4X (GB200, a4x-highgpu-4g, 4 GPUs, 2,000 Gbps RoCE CX-7) is further restricted to future reservations in AI Hypercomputer only.", "provider": "Google Cloud", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "Standard_ND96isr_H100_v5", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink 4.0 (Microsoft Learn ND H100 v5 page); 640 GB total GPU memory", "inter_node_fabric": "InfiniBand NDR: 'dedicated, topology-agnostic 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand connection' per GPU, 3.2 Tbps per VM, GPUDirect RDMA (Microsoft Learn)", "max_contiguous_gpus": null, "regions": [ "eastus" ], "price_usd_per_gpu_hr": 12.29, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://prices.azure.com/api/retail/prices?$filter=priceType eq 'Consumption' and armSkuName eq 'Standard_ND96isr_H100_v5' and armRegionName eq 'eastus'", "spec_confidence": "published", "notes": "Per-VM $98.32/hr (Linux Consumption meter, eastus, Azure Retail Prices API) divided by 8. Scale claim is non-numeric: 'can scale up to thousands of GPUs'.", "provider": "Microsoft Azure", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink 4.0 (Microsoft Learn ND H100 v5 page); 640 GB total GPU memory)", "scale-out ~400 Gbps/GPU", "cluster product claim (contiguity unstated)" ] }, { "sku": "Standard_ND96isr_H200_v5", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink, 900 GB/s (Microsoft Learn ND H200 v5 page); 1,128 GB total GPU memory", "inter_node_fabric": "InfiniBand NDR: dedicated 400 Gbps NVIDIA Quantum-2 CX7 InfiniBand per GPU, 3.2 Tb/s per VM, GPUDirect RDMA (Microsoft Learn)", "max_contiguous_gpus": null, "regions": [ "eastus2" ], "price_usd_per_gpu_hr": 10.6, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://prices.azure.com/api/retail/prices?$filter=priceType eq 'Consumption' and contains(armSkuName,'H200')", "spec_confidence": "published", "notes": "Per-VM $84.80/hr (Linux Consumption, eastus2/westus2/westus3; centralus is $95.824) divided by 8. Not listed in eastus in the API at collection time. H200 currently lists below H100 in Azure's retail API.", "provider": "Microsoft Azure", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~400 Gbps/GPU", "contiguity unstated" ] }, { "sku": "Standard_ND128isr_NDR_GB200_v6", "accelerator": "GB200 (Blackwell)", "vram_gb": 192, "form_factor": "SXM/superchip (2x Grace CPU + 4x Blackwell GPU per VM)", "gpus_per_node": 4, "intra_node": "5th-gen NVLink, 4x 1.8 TB/s per VM; NVL72 rack domain of 72 GPUs across 18 VMs with 130 TB/s cross-sectional NVLink (Microsoft Learn)", "inter_node_fabric": "InfiniBand NDR: 4x 400 Gb/s NVIDIA Quantum-2 CX7 InfiniBand connections per VM (one 400 Gb/s NIC per GPU); 28.8 Tb/s scale-out networking per 72-GPU rack (Microsoft Learn)", "max_contiguous_gpus": 100000, "regions": [ "eastus", "eastus2" ], "price_usd_per_gpu_hr": 27.04, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://prices.azure.com/api/retail/prices?$filter=priceType eq 'Consumption' and contains(armSkuName,'GB200')", "spec_confidence": "published", "notes": "Per-VM $108.16/hr (Linux Consumption, eastus/eastus2/westus2/westus3) divided by 4 GPUs (this VM is a quarter of a GB200 NVL72 rack). max_contiguous_gpus from Learn network-table note: 'Bandwidth is per NIC to up to 100,000 GPUs through InfiniBand'.", "provider": "Microsoft Azure", "as_of": "2026-09-10", "accelerator_canonical": "GB200-NVL72", "grade": "A", "grade_reasons": [ "scale-up domain (5th-gen NVLink, 4x 1.8 TB/s per VM; NVL72 rack domain of 72 GPUs across 18 VMs with 130 TB/s cross-sectional NVLink (Microsoft Learn))", "scale-out ~28800 Gbps/GPU", "contiguity 100000 GPUs" ] }, { "sku": "BM.GPU.H100.8", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM (bare metal)", "gpus_per_node": 8, "intra_node": "NVLink (8x NVIDIA H100 80GB Tensor Core per Oracle shape table)", "inter_node_fabric": "RDMA cluster network: '8x2x200 Gb/sec RDMA' = 3,200 Gbps per node (Oracle price-list shape table); 'up to 3,200 Gb/sec of RDMA cluster network bandwidth' (Oracle GPU page)", "max_contiguous_gpus": 16384, "regions": [], "price_usd_per_gpu_hr": 10, "pricing_model": "on-demand (PAY_AS_YOU_GO)", "min_commit": null, "notice_period": null, "source_url": "https://apexapps.oracle.com/pls/apex/cetools/api/v1/products/?partNumber=B98415", "spec_confidence": "published", "notes": "Oracle prices per GPU-hour directly: part B98415 'Compute - GPU - H100', $10.00/GPU/hr PAY_AS_YOU_GO ($80/node). Oracle publishes one global list price; regions field left empty. Supercluster max: 16,384 H100 GPUs (Oracle GPU page footnote).", "provider": "Oracle Cloud (OCI)", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~200 Gbps/GPU", "contiguity 16384 GPUs" ] }, { "sku": "BM.GPU.H200.8", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM (bare metal)", "gpus_per_node": 8, "intra_node": "NVLink (8x NVIDIA H200 141GB Tensor Core per Oracle shape table)", "inter_node_fabric": "RDMA cluster network: '8x1x400 Gb/sec RDMA' = 3,200 Gbps per node (Oracle price-list shape table)", "max_contiguous_gpus": 65536, "regions": [], "price_usd_per_gpu_hr": 10, "pricing_model": "on-demand (PAY_AS_YOU_GO)", "min_commit": null, "notice_period": null, "source_url": "https://apexapps.oracle.com/pls/apex/cetools/api/v1/products/?partNumber=B110519", "spec_confidence": "published", "notes": "Part B110519 'OCI - Compute - GPU - H200', $10.00/GPU/hr PAY_AS_YOU_GO — same per-GPU rate as H100. Supercluster max: 65,536 H200 GPUs (Oracle GPU page footnote).", "provider": "Oracle Cloud (OCI)", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink (8x NVIDIA H200 141GB Tensor Core per Oracle shape table))", "scale-out ~400 Gbps/GPU", "contiguity 65536 GPUs" ] }, { "sku": "BM.GPU.B200.8", "accelerator": "B200", "vram_gb": 180, "form_factor": "SXM (bare metal)", "gpus_per_node": 8, "intra_node": "NVLink (8x Nvidia B200 180GB per Oracle shape table)", "inter_node_fabric": "RDMA cluster network: '8x400 Gb/sec' = 3,200 Gbps per node (Oracle price-list shape table; fabric protocol not named for this row)", "max_contiguous_gpus": 131072, "regions": [], "price_usd_per_gpu_hr": 14, "pricing_model": "on-demand (PAY_AS_YOU_GO)", "min_commit": null, "notice_period": null, "source_url": "https://apexapps.oracle.com/pls/apex/cetools/api/v1/products/?partNumber=B110978", "spec_confidence": "published", "notes": "Part B110978 'OCI - Compute - GPU - B200', $14.00/GPU/hr PAY_AS_YOU_GO ($112/node). Supercluster max: 131,072 B200 GPUs (Oracle GPU page footnote).", "provider": "Oracle Cloud (OCI)", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "BM.GPU.GB200.4", "accelerator": "GB200 (4x B200 189GB in NVL72)", "vram_gb": 189, "form_factor": "Superchip/NVL72 (bare metal, Grace + Blackwell)", "gpus_per_node": 4, "intra_node": "NVLink NVL72 rack domain ('4x Nvidia B200 189GB (NVL72)' per Oracle shape table)", "inter_node_fabric": "'4x400 Gb/sec InfiniBand' = 1,600 Gbps per node (Oracle price-list shape table)", "max_contiguous_gpus": 131072, "regions": [], "price_usd_per_gpu_hr": 16, "pricing_model": "on-demand (PAY_AS_YOU_GO)", "min_commit": null, "notice_period": null, "source_url": "https://apexapps.oracle.com/pls/apex/cetools/api/v1/products/?partNumber=B110979", "spec_confidence": "published", "notes": "Part B110979 'OCI - Compute - GPU - GB200', $16.00/GPU/hr PAY_AS_YOU_GO. Oracle GPU page footnote: Supercluster scales to '131,072 NVIDIA Blackwell B200 GPUs in NVIDIA Grace Blackwell GB200 Superchips'; page also says GB200 superclusters 'scale to more than 100,000 GPUs'.", "provider": "Oracle Cloud (OCI)", "as_of": "2026-09-10", "accelerator_canonical": "GB200-NVL72", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink NVL72 rack domain ('4x Nvidia B200 189GB (NVL72)' per Oracle shape table))", "scale-out ~400 Gbps/GPU", "contiguity 131072 GPUs" ] }, { "sku": "NVIDIA B200 SXM", "accelerator": "B200", "vram_gb": 192, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": [ "CANADA-1" ], "price_usd_per_gpu_hr": 6, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.hyperstack.cloud/gpu-pricing", "spec_confidence": "published", "notes": "Reserved rate also published at $5.10/GPU/hr. Docs flavor page lists 8x-only config, 252 vCPU / 2048 GB RAM, default network bandwidth only (no SR-IOV noted for B200). NVLink not explicitly stated on fetched pages.", "provider": "Hyperstack", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "NVIDIA H200 SXM", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "SR-IOV Ethernet up to 350 Gbps inter-VM (default up to 16 Gbps)", "max_contiguous_gpus": null, "regions": [ "CANADA-1" ], "price_usd_per_gpu_hr": 3.99, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.hyperstack.cloud/gpu-pricing", "spec_confidence": "published", "notes": "Reserved rate also published at $2.79/GPU/hr. 8x-only flavor per docs. NVLink not explicitly stated in fetched flavor table despite SXM form factor.", "provider": "Hyperstack", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "NVIDIA H100 SXM", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "SR-IOV Ethernet up to 350 Gbps inter-VM (default up to 16 Gbps)", "max_contiguous_gpus": null, "regions": [ "CANADA-1", "US-1" ], "price_usd_per_gpu_hr": 3.2, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.hyperstack.cloud/gpu-pricing", "spec_confidence": "published", "notes": "8x-only flavor, 192 vCPU, 1800 GB RAM (CANADA-1) / 960 GB (US-1) per docs. NVLink not explicitly stated in fetched flavor table.", "provider": "Hyperstack", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "NVIDIA H100 SXM (reserved)", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "SR-IOV Ethernet up to 350 Gbps inter-VM (default up to 16 Gbps)", "max_contiguous_gpus": null, "regions": [ "CANADA-1", "US-1" ], "price_usd_per_gpu_hr": 2.72, "pricing_model": "reserved", "min_commit": null, "notice_period": null, "source_url": "https://www.hyperstack.cloud/gpu-pricing", "spec_confidence": "published", "notes": "Pricing page states reserved contracts are invoiced monthly but does not publish minimum commitment periods.", "provider": "Hyperstack", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "NVIDIA H100 PCIe", "accelerator": "H100-PCIe", "vram_gb": 80, "form_factor": "PCIe", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "SR-IOV Ethernet up to 350 Gbps inter-VM (default up to 16 Gbps)", "max_contiguous_gpus": null, "regions": [ "CANADA-1" ], "price_usd_per_gpu_hr": 2.5, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.hyperstack.cloud/gpu-pricing", "spec_confidence": "published", "notes": "1x/2x/4x/8x configs per docs (8 is max). Reserved rate also published at $1.75/GPU/hr. Separate 'H100 NVLink' SKU (PCIe + NVLink, 8x flavor) published at $2.60/GPU/hr on-demand.", "provider": "Hyperstack", "as_of": "2026-09-10", "accelerator_canonical": "H100-PCIe", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "NVIDIA A100 PCIe 80GB", "accelerator": "A100", "vram_gb": 80, "form_factor": "PCIe", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": [ "CANADA-1" ], "price_usd_per_gpu_hr": 1.35, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.hyperstack.cloud/gpu-pricing", "spec_confidence": "published", "notes": "1x/2x/4x/8x configs (8 max). Related published SKUs: A100 SXM (US-1) $1.60/hr, A100 NVLink $1.40/hr on-demand; A100 reserved from $0.95/hr (variant not specified on pricing page).", "provider": "Hyperstack", "as_of": "2026-09-10", "accelerator_canonical": "A100-PCIe", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "B300 SXM6", "accelerator": "B300", "vram_gb": 268, "form_factor": "SXM", "gpus_per_node": null, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 7.5, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://verda.com/pricing", "spec_confidence": "published", "notes": "Spot price $3.75/hr also published. Pricing page lists VRAM as 268GB. Long-term self-service discounts published: 2% (1-mo), 3% (3-mo), 4% (6-mo), 8% (1-yr), 25% (2-yr) off on-demand.", "provider": "DataCrunch (rebranded to Verda, verda.com)", "as_of": "2026-09-10", "accelerator_canonical": "B300", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "B200 SXM6", "accelerator": "B200", "vram_gb": 180, "form_factor": "SXM", "gpus_per_node": null, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 6.11, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://verda.com/pricing", "spec_confidence": "published", "notes": "Spot price $3.06/hr also published. Pricing page lists VRAM as 180GB (not 192GB). Same long-term discount schedule applies (up to 25% for 2-yr).", "provider": "DataCrunch (rebranded to Verda, verda.com)", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "H200 SXM5", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": null, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 4, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://verda.com/pricing", "spec_confidence": "published", "notes": "Spot price $2.00/hr also published. Per-GPU pricing shown with 44 vCPU / 170GB RAM per GPU.", "provider": "DataCrunch (rebranded to Verda, verda.com)", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "H100 SXM5", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": null, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 3.25, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://verda.com/pricing", "spec_confidence": "published", "notes": "Spot price $1.63/hr also published.", "provider": "DataCrunch (rebranded to Verda, verda.com)", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "A100 SXM4 80GB", "accelerator": "A100", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": null, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 1.79, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://verda.com/pricing", "spec_confidence": "published", "notes": "Spot price $0.895/hr also published. A100 SXM4 40GB variant also listed at $1.29/hr OD ($0.645 spot).", "provider": "DataCrunch (rebranded to Verda, verda.com)", "as_of": "2026-09-10", "accelerator_canonical": "A100-SXM", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "L40S", "accelerator": "L40S", "vram_gb": 48, "form_factor": null, "gpus_per_node": null, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 1.37, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://verda.com/pricing", "spec_confidence": "published", "notes": "Spot price $0.685/hr also published.", "provider": "DataCrunch (rebranded to Verda, verda.com)", "as_of": "2026-09-10", "accelerator_canonical": "L40S", "grade": "U", "grade_reasons": [ "no interconnect or form-factor information published" ] }, { "sku": "H100-SXM (H100-SXM-*-80G)", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": null, "intra_node": "NVLink", "inter_node_fabric": "Up to 20 Gbps Ethernet", "max_contiguous_gpus": null, "regions": [ "EU (Paris PAR-2, DC5)" ], "price_usd_per_gpu_hr": 3.61, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://gpuperhour.com/providers/scaleway", "spec_confidence": "secondary", "notes": "Hardware specs (80GB HBM3 3.35TB/s per GPU, NVLink, up to 20 Gbps network, PAR-2, 99.5% SLA) are published on scaleway.com/en/h100-sxm/. Price is from third-party tracker gpuperhour.com (USD as listed) because Scaleway's official pricing page renders PAR-2 EUR prices client-side and they could not be extracted.", "provider": "Scaleway", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "H100-1-80G", "accelerator": "H100-PCIe", "vram_gb": 80, "form_factor": "PCIe", "gpus_per_node": 2, "intra_node": null, "inter_node_fabric": "Up to 20 Gbps Ethernet", "max_contiguous_gpus": null, "regions": [ "EU (Paris PAR-2, Warsaw WAW-2)" ], "price_usd_per_gpu_hr": 3.34, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://gpuperhour.com/providers/scaleway", "spec_confidence": "secondary", "notes": "gpus_per_node = 2 is the max published config (H100-1-80G and H100-2-80G) per scaleway.com/en/h100/. Specs (80GB HBM2e 2TB/s, PCIe, 20 Gbps) published on official product page; price from third-party tracker (USD as listed).", "provider": "Scaleway", "as_of": "2026-09-10", "accelerator_canonical": "H100-PCIe", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "L40S (L40S-*-48G)", "accelerator": "L40S", "vram_gb": 48, "form_factor": null, "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": [ "EU (Paris PAR-2)" ], "price_usd_per_gpu_hr": 1.68, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://gpuperhour.com/providers/scaleway", "spec_confidence": "secondary", "notes": "Official pricing page lists L40S-1-48G through L40S-8-48G (so up to 8 GPUs/instance) available in PAR-2, but EUR rates are client-rendered and not extractable; price from third-party tracker (USD as listed).", "provider": "Scaleway", "as_of": "2026-09-10", "accelerator_canonical": "L40S", "grade": "U", "grade_reasons": [ "no interconnect or form-factor information published" ] }, { "sku": "L4-1-24G", "accelerator": "L4", "vram_gb": 24, "form_factor": null, "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "2.5-20 Gbps Ethernet", "max_contiguous_gpus": null, "regions": [ "EU (Paris PAR-1)" ], "price_usd_per_gpu_hr": 0.87, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.scaleway.com/en/pricing/gpu/", "spec_confidence": "published", "notes": "Official published price EUR 0.79/GPU/hr, converted at ~1.10 USD/EUR = $0.87. Scales linearly to L4-8-24G (8 GPUs, EUR 6.30/hr total). Included as the one fully first-party-priced Scaleway offering; L4 is below datacenter-class focus.", "provider": "Scaleway", "as_of": "2026-09-10", "accelerator_canonical": "L4", "grade": "D", "grade_reasons": [ "scale-out fabric only, no scale-up domain" ] }, { "sku": "B300-SXM-8-288G", "accelerator": "B300", "vram_gb": 288, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": [ "EU (Paris PAR-2)" ], "price_usd_per_gpu_hr": 8.74, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://gpuperhour.com/providers/scaleway", "spec_confidence": "secondary", "notes": "Instance name B300-SXM-8-288G (8 GPUs, 288G per GPU per Scaleway naming convention) confirmed on official pricing page; official EUR rate not extractable. Tracker price $8.74/GPU/hr USD; tracker lists VRAM as 262GB, conflicting with the official 288G instance name — VRAM recorded from official name.", "provider": "Scaleway", "as_of": "2026-09-10", "accelerator_canonical": "B300", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "Nvidia H200 (141 GB HBM3e, NVLink, 8-GPU instance)", "accelerator": "H200", "vram_gb": 141, "form_factor": null, "gpus_per_node": 8, "intra_node": "NVLink", "inter_node_fabric": "Up to 25 Gbps Ethernet (public/private network)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 6.2, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.ovhcloud.com/en/public-cloud/gpu/", "spec_confidence": "inferred", "notes": "Page publishes 'From $49.56/hour' with footnote '*Estimated pricing for 8 GPUs' — per-GPU rate of $6.195 (~$6.20) is inferred by dividing by 8, not stated directly. 8 GPUs/instance and 141 GB HBM3e per GPU published on the H200 product page. Exact instance name not published on fetched pages.", "provider": "OVHcloud", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "Nvidia H100 (80 GB HBM2e, PCIe 5.0)", "accelerator": "H100-PCIe", "vram_gb": 80, "form_factor": "PCIe", "gpus_per_node": 4, "intra_node": null, "inter_node_fabric": "Up to 25 Gbps Ethernet (public/private network)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 2.99, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.ovhcloud.com/en/public-cloud/gpu/", "spec_confidence": "published", "notes": "'From $2.99/hour'. Page states '1-4 GPUs per instance' generally for GPU instances (4 recorded as max configurable).", "provider": "OVHcloud", "as_of": "2026-09-10", "accelerator_canonical": "H100-PCIe", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "Nvidia L40S (48 GB GDDR6, PCIe 4.0)", "accelerator": "L40S", "vram_gb": 48, "form_factor": "PCIe", "gpus_per_node": 4, "intra_node": null, "inter_node_fabric": "Up to 25 Gbps Ethernet (public/private network)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 1.8, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.ovhcloud.com/en/public-cloud/gpu/", "spec_confidence": "published", "notes": "'From $1.80/hour'. '1-4 GPUs per instance' general statement (4 recorded as max configurable).", "provider": "OVHcloud", "as_of": "2026-09-10", "accelerator_canonical": "L40S", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "Nvidia L4 (24 GB GDDR6, PCIe 4.0)", "accelerator": "L4", "vram_gb": 24, "form_factor": "PCIe", "gpus_per_node": 4, "intra_node": null, "inter_node_fabric": "Up to 25 Gbps Ethernet (public/private network)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 1, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.ovhcloud.com/en/public-cloud/gpu/", "spec_confidence": "published", "notes": "'From $1.00/hour'. Below datacenter-class focus; included for coverage since OVH no longer lists A100.", "provider": "OVHcloud", "as_of": "2026-09-10", "accelerator_canonical": "L4", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "HGX B200", "accelerator": "B200", "vram_gb": 180, "form_factor": null, "gpus_per_node": null, "intra_node": null, "inter_node_fabric": "High-speed InfiniBand", "max_contiguous_gpus": 1000, "regions": [ "US", "EU", "Asia", "Middle East" ], "price_usd_per_gpu_hr": 8.19, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.together.ai/gpu-clusters", "spec_confidence": "published", "notes": "Page lists B200 VRAM as 180 GB. Published scale: 256 to 1,000+ GPUs (max_contiguous recorded as 1000, page says '1,000+'). Managed Kubernetes or Slurm on Kubernetes.", "provider": "Together AI (GPU Clusters)", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "B", "grade_reasons": [ "scale-up domain present", "fabric stated, bandwidth unquantified", "contiguity 1000 GPUs" ] }, { "sku": "HGX B200 (reserved)", "accelerator": "B200", "vram_gb": 180, "form_factor": null, "gpus_per_node": null, "intra_node": null, "inter_node_fabric": "High-speed InfiniBand", "max_contiguous_gpus": 1000, "regions": [ "US", "EU", "Asia", "Middle East" ], "price_usd_per_gpu_hr": 6.79, "pricing_model": "reserved", "min_commit": null, "notice_period": null, "source_url": "https://www.together.ai/gpu-clusters", "spec_confidence": "published", "notes": "'Starting at $6.79/hr'. Page states reserved options up to 6 months with upfront payment; exact minimum term not published.", "provider": "Together AI (GPU Clusters)", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "B", "grade_reasons": [ "scale-up domain present", "fabric stated, bandwidth unquantified", "contiguity 1000 GPUs" ] }, { "sku": "H200", "accelerator": "H200", "vram_gb": 140, "form_factor": null, "gpus_per_node": null, "intra_node": null, "inter_node_fabric": "High-speed InfiniBand", "max_contiguous_gpus": 1000, "regions": [ "US", "EU", "Asia", "Middle East" ], "price_usd_per_gpu_hr": 5.99, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.together.ai/gpu-clusters", "spec_confidence": "published", "notes": "Page lists H200 VRAM as 140 GB (vendor spec is 141 GB; recorded as published). Published scale: 256 to 1,000 GPUs.", "provider": "Together AI (GPU Clusters)", "as_of": "2026-09-10", "accelerator_canonical": "H200", "grade": "D", "grade_reasons": [ "scale-out fabric only, no scale-up domain" ] }, { "sku": "H200 (reserved)", "accelerator": "H200", "vram_gb": 140, "form_factor": null, "gpus_per_node": null, "intra_node": null, "inter_node_fabric": "High-speed InfiniBand", "max_contiguous_gpus": 1000, "regions": [ "US", "EU", "Asia", "Middle East" ], "price_usd_per_gpu_hr": 3.99, "pricing_model": "reserved", "min_commit": null, "notice_period": null, "source_url": "https://www.together.ai/gpu-clusters", "spec_confidence": "published", "notes": "'Starting at $3.99/hr'. Reserved options up to 6 months with upfront payment.", "provider": "Together AI (GPU Clusters)", "as_of": "2026-09-10", "accelerator_canonical": "H200", "grade": "D", "grade_reasons": [ "scale-out fabric only, no scale-up domain" ] }, { "sku": "H100", "accelerator": "H100", "vram_gb": 80, "form_factor": null, "gpus_per_node": null, "intra_node": null, "inter_node_fabric": "High-speed InfiniBand", "max_contiguous_gpus": 256, "regions": [ "US", "EU", "Asia", "Middle East" ], "price_usd_per_gpu_hr": 3.99, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.together.ai/gpu-clusters", "spec_confidence": "published", "notes": "Published scale: 8 to 256 GPUs. SXM vs PCIe not distinguished on the page.", "provider": "Together AI (GPU Clusters)", "as_of": "2026-09-10", "accelerator_canonical": "H100", "grade": "D", "grade_reasons": [ "scale-out fabric only, no scale-up domain" ] }, { "sku": "H100 (reserved)", "accelerator": "H100", "vram_gb": 80, "form_factor": null, "gpus_per_node": null, "intra_node": null, "inter_node_fabric": "High-speed InfiniBand", "max_contiguous_gpus": 256, "regions": [ "US", "EU", "Asia", "Middle East" ], "price_usd_per_gpu_hr": 3.19, "pricing_model": "reserved", "min_commit": null, "notice_period": null, "source_url": "https://www.together.ai/gpu-clusters", "spec_confidence": "published", "notes": "'Starting at $3.19/hr'. Reserved options up to 6 months with upfront payment.", "provider": "Together AI (GPU Clusters)", "as_of": "2026-09-10", "accelerator_canonical": "H100", "grade": "D", "grade_reasons": [ "scale-out fabric only, no scale-up domain" ] }, { "sku": "NVIDIA GB200 NVL72 (gb200-4x)", "accelerator": "GB200 (Blackwell, Grace Blackwell Superchip)", "vram_gb": 186, "form_factor": "superchip", "gpus_per_node": 4, "intra_node": "NVLink Switch System (per NVL72 docs)", "inter_node_fabric": "400 Gb/s NDR InfiniBand, NVIDIA Quantum-2 fabric", "max_contiguous_gpus": null, "regions": [ "North America", "Europe" ], "price_usd_per_gpu_hr": 10.5, "pricing_model": "on-demand", "min_commit": "docs state NVL72 must be deployed as full racks of 18 nodes", "notice_period": null, "source_url": "https://www.coreweave.com/pricing", "spec_confidence": "published", "notes": "Priced per instance: $42.00/hr for a 4-GPU instance (2 Superchips x 2 Blackwell GPUs); divided by 4. Separate 'inference' rate of $10.50/hr also listed. NVLink domain size (72 GPUs) not stated on fetched pages.", "provider": "CoreWeave", "as_of": "2026-09-10", "accelerator_canonical": "GB200-NVL72", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~400 Gbps/GPU", "contiguity unstated" ] }, { "sku": "NVIDIA HGX B200 (8x)", "accelerator": "B200-SXM", "vram_gb": 180, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "8x NVIDIA ConnectX-7 InfiniBand HCAs, 400G NDR non-blocking NVIDIA Quantum-2 fabric, plus BlueField-3 DPU (per B200 release notes)", "max_contiguous_gpus": null, "regions": [ "North America", "Europe" ], "price_usd_per_gpu_hr": 8.6, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.coreweave.com/pricing", "spec_confidence": "published", "notes": "Priced per instance: $68.80/hr for 8 GPUs; divided by 8. Spot also published: $34.11/instance-hr NA ($4.26/GPU), $34.87 Europe. 180 GB HBM3e per GPU per docs changelog.", "provider": "CoreWeave", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~400 Gbps/GPU", "contiguity unstated" ] }, { "sku": "NVIDIA HGX H200 (gd-8xh200ib-i128)", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink, up to 900 GB/s GPU-to-GPU (stated as 'Up to 900 Gbps' on product page)", "inter_node_fabric": "InfiniBand (instance ID carries 'ib'; CoreWeave HPC interconnect docs state Quantum-2/Quantum-X800 non-blocking fat-tree fabrics, but per-node bandwidth for H200 not stated on fetched pages)", "max_contiguous_gpus": null, "regions": [ "North America", "Europe" ], "price_usd_per_gpu_hr": 6.31, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.coreweave.com/pricing", "spec_confidence": "published", "notes": "Priced per instance: $50.44/hr for 8 GPUs; $6.305/GPU rounded to 6.31. Spot published: $20.93/instance NA, $20.64 Europe. Instance: 128 vCPU / 2048 GB RAM.", "provider": "CoreWeave", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "fabric stated, bandwidth unquantified", "contiguity unstated" ] }, { "sku": "NVIDIA HGX H100 (gd-8xh100ib-i128)", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand (instance ID carries 'ib'; fabric generation per HPC interconnect docs is Quantum-2/Quantum-X800 non-blocking fat-tree, but per-node bandwidth for H100 not stated on fetched pages)", "max_contiguous_gpus": null, "regions": [ "North America", "Europe" ], "price_usd_per_gpu_hr": 6.16, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.coreweave.com/pricing", "spec_confidence": "published", "notes": "Priced per instance: $49.24/hr for 8 GPUs; $6.155/GPU rounded to 6.16. Instance: 128 vCPU / 2048 GB RAM.", "provider": "CoreWeave", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "fabric stated, bandwidth unquantified", "contiguity unstated" ] }, { "sku": "NVIDIA HGX H100 (gd-8xh100ib-i128) — Spot", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand (same instance as on-demand H100)", "max_contiguous_gpus": null, "regions": [ "North America", "Europe" ], "price_usd_per_gpu_hr": 2.46, "pricing_model": "spot", "min_commit": "none", "notice_period": null, "source_url": "https://www.coreweave.com/pricing", "spec_confidence": "published", "notes": "Spot per instance: $19.71/hr NA ($2.46/GPU), $19.51/hr Europe ($2.44/GPU).", "provider": "CoreWeave", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "fabric stated, bandwidth unquantified", "contiguity unstated" ] }, { "sku": "NVIDIA A100 (gd-8xa100-i128)", "accelerator": "A100 80GB", "vram_gb": 80, "form_factor": null, "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": null, "regions": [ "North America", "Europe" ], "price_usd_per_gpu_hr": 2.7, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://www.coreweave.com/pricing", "spec_confidence": "published", "notes": "Priced per instance: $21.60/hr for 8 GPUs. Spot published: $9.65/instance NA, $9.51 Europe. Docs instance ID has no 'ib' suffix; no fabric stated.", "provider": "CoreWeave", "as_of": "2026-09-10", "accelerator_canonical": "A100", "grade": "U", "grade_reasons": [ "no interconnect or form-factor information published" ] }, { "sku": "8x NVIDIA B200 SXM6 (on-demand instance)", "accelerator": "B200-SXM6", "vram_gb": 180, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": null, "price_usd_per_gpu_hr": 6.69, "pricing_model": "on-demand", "min_commit": "none", "notice_period": null, "source_url": "https://lambda.ai/instances", "spec_confidence": "published", "notes": "Priced per GPU per hour. Smaller configs cost slightly more: 4x $6.79, 2x $6.89, 1x $6.99. No egress fees. Max contiguous for on-demand instances is the 8-GPU node; larger scale goes via 1-Click Clusters.", "provider": "Lambda", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "8x NVIDIA H100 SXM (on-demand instance)", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": null, "price_usd_per_gpu_hr": 3.99, "pricing_model": "on-demand", "min_commit": "none", "notice_period": null, "source_url": "https://lambda.ai/instances", "spec_confidence": "published", "notes": "Per GPU per hour. 4x $4.09, 2x $4.19, 1x $4.29.", "provider": "Lambda", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "1x NVIDIA H100 PCIe (on-demand instance)", "accelerator": "H100-PCIe", "vram_gb": 80, "form_factor": "PCIe", "gpus_per_node": 1, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 1, "regions": null, "price_usd_per_gpu_hr": 3.29, "pricing_model": "on-demand", "min_commit": "none", "notice_period": null, "source_url": "https://lambda.ai/instances", "spec_confidence": "published", "notes": "Only listed as 1x configuration.", "provider": "Lambda", "as_of": "2026-09-10", "accelerator_canonical": "H100-PCIe", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "8x NVIDIA A100 SXM 80GB (on-demand instance)", "accelerator": "A100-SXM 80GB", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": null, "price_usd_per_gpu_hr": 2.79, "pricing_model": "on-demand", "min_commit": "none", "notice_period": null, "source_url": "https://lambda.ai/instances", "spec_confidence": "published", "notes": "A100 SXM 40GB also listed at $1.99/GPU/hr (8x).", "provider": "Lambda", "as_of": "2026-09-10", "accelerator_canonical": "A100-SXM", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "1-Click Cluster — NVIDIA HGX B200", "accelerator": "B200-SXM6", "vram_gb": 180, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "NVIDIA Quantum-2 InfiniBand with SHARP acceleration (per-node bandwidth not stated on fetched pages)", "max_contiguous_gpus": 2000, "regions": null, "price_usd_per_gpu_hr": 9.86, "pricing_model": "reserved", "min_commit": "16 GPUs, 2-week minimum term", "notice_period": null, "source_url": "https://lambda.ai/1-click-clusters", "spec_confidence": "published", "notes": "Tiered by size for 2wk-1yr terms: 16 GPUs $9.86, 64 GPUs $9.36, 256+ GPUs $8.87 per GPU/hr; 1yr+ is custom pricing. Page states clusters 'from 16 to 2,000+ NVIDIA GPUs' (2000 recorded as stated floor of the top range). Prices exclude sales tax/VAT/GST.", "provider": "Lambda", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "B", "grade_reasons": [ "scale-up domain present", "fabric stated, bandwidth unquantified", "contiguity 2000 GPUs" ] }, { "sku": "1-Click Cluster — NVIDIA H100", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "NVIDIA Quantum-2 InfiniBand with SHARP acceleration (per-node bandwidth not stated on fetched pages)", "max_contiguous_gpus": 2000, "regions": null, "price_usd_per_gpu_hr": 6.16, "pricing_model": "reserved", "min_commit": "16 GPUs, 2-week minimum term", "notice_period": null, "source_url": "https://lambda.ai/1-click-clusters", "spec_confidence": "published", "notes": "Tiered: 16 GPUs $6.16, 64 GPUs $5.85, 256 GPUs $5.54 per GPU/hr; 1yr+ custom. H100 form factor for 1CC nodes not explicitly stated on fetched page (B200 nodes explicitly 'HGX B200 SXM6').", "provider": "Lambda", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "fabric stated, bandwidth unquantified", "contiguity 2000 GPUs" ] }, { "sku": "h200-141gb-sxm-ib.8x", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand, 3200 Gbps per 8-GPU instance (per docs VM overview)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 4.29, "pricing_model": "on-demand", "min_commit": "none stated", "notice_period": null, "source_url": "https://www.crusoe.ai/cloud/pricing", "spec_confidence": "published", "notes": "Priced per GPU per hour. Instance: 176 vCPU / 2000 GB RAM. InfiniBand included in cost of IB-enabled VMs; no ingress/egress charges.", "provider": "Crusoe", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~3200 Gbps/GPU", "contiguity unstated" ] }, { "sku": "h100-80gb-sxm-ib.8x", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand, 3200 Gbps per 8-GPU instance (per docs VM overview)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 3.9, "pricing_model": "on-demand", "min_commit": "none stated", "notice_period": null, "source_url": "https://www.crusoe.ai/cloud/pricing", "spec_confidence": "published", "notes": "Instance: 176 vCPU / 960 GB RAM. Self-serve inference tier for H100 listed separately at $5.50/hr.", "provider": "Crusoe", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~3200 Gbps/GPU", "contiguity unstated" ] }, { "sku": "mi300x-192gb-ib.8x", "accelerator": "AMD MI300X", "vram_gb": 192, "form_factor": null, "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand, 3200 Gbps per 8-GPU instance (per docs VM overview)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 3.45, "pricing_model": "on-demand", "min_commit": "none stated", "notice_period": null, "source_url": "https://www.crusoe.ai/cloud/pricing", "spec_confidence": "published", "notes": "Instance: 240 vCPU / 2000 GB RAM. MI355X (288GB, mi355x-288gb-roce.8x, 3200 Gbps RoCE) exists in docs but is contact-sales, no public price.", "provider": "Crusoe", "as_of": "2026-09-10", "accelerator_canonical": "MI300X", "grade": "D", "grade_reasons": [ "scale-out fabric only, no scale-up domain" ] }, { "sku": "a100-80gb-sxm-ib.8x", "accelerator": "A100-SXM 80GB", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand, 1600 Gbps per 8-GPU instance (per docs VM overview)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 2.3, "pricing_model": "on-demand", "min_commit": "none stated", "notice_period": null, "source_url": "https://www.crusoe.ai/cloud/pricing", "spec_confidence": "published", "notes": "Pricing page lists 'A100 (SXM) 80GB $2.30' and 'A100 (PCIe) 80GB $2.00'. Non-IB A100 slices (1x/2x/4x/8x) also exist per docs.", "provider": "Crusoe", "as_of": "2026-09-10", "accelerator_canonical": "A100-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~1600 Gbps/GPU", "contiguity unstated" ] }, { "sku": "NVIDIA HGX B200 (gpu-b200-sxm platform)", "accelerator": "B200-SXM", "vram_gb": 180, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink (platform listed as 'B200 NVLink' in cluster docs; topology details not stated)", "inter_node_fabric": "InfiniBand, 400 Gbps per GPU / 3.2 Tbps per 8-GPU node, GPUDirect RDMA (per docs.nebius.com)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 7.15, "pricing_model": "on-demand", "min_commit": "none", "notice_period": null, "source_url": "https://nebius.com/prices", "spec_confidence": "published", "notes": "Priced per GPU per hour (20 vCPU / 224 GB RAM listed per GPU). vram_gb is the standard SKU value; Nebius pricing page does not state VRAM.", "provider": "Nebius", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink (platform listed as 'B200 NVLink' in cluster docs; topology details not stated))", "scale-out ~3200 Gbps/GPU", "cluster product claim (contiguity unstated)" ] }, { "sku": "NVIDIA HGX B200 — preemptible", "accelerator": "B200-SXM", "vram_gb": 180, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand, 400 Gbps per GPU / 3.2 Tbps per node (per docs.nebius.com)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 3.95, "pricing_model": "spot", "min_commit": "none", "notice_period": null, "source_url": "https://nebius.com/prices", "spec_confidence": "published", "notes": "Preemptible tier; can be interrupted.", "provider": "Nebius", "as_of": "2026-09-10", "accelerator_canonical": "B200", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~400 Gbps/GPU", "contiguity unstated" ] }, { "sku": "NVIDIA HGX H200 (gpu-h200-sxm platform)", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink (platform listed as 'H200 NVLink' in cluster docs)", "inter_node_fabric": "InfiniBand, 400 Gbps per GPU / 3.2 Tbps per 8-GPU node, GPUDirect RDMA (per docs.nebius.com)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 4.5, "pricing_model": "on-demand", "min_commit": "none", "notice_period": null, "source_url": "https://nebius.com/prices", "spec_confidence": "published", "notes": "Per GPU per hour (16 vCPU / 200 GB RAM per GPU). vram_gb is the standard SKU value; not stated on pricing page.", "provider": "Nebius", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink (platform listed as 'H200 NVLink' in cluster docs))", "scale-out ~3200 Gbps/GPU", "cluster product claim (contiguity unstated)" ] }, { "sku": "NVIDIA HGX H200 — preemptible", "accelerator": "H200-SXM", "vram_gb": 141, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand, 400 Gbps per GPU / 3.2 Tbps per node (per docs.nebius.com)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 2.45, "pricing_model": "spot", "min_commit": "none", "notice_period": null, "source_url": "https://nebius.com/prices", "spec_confidence": "published", "notes": "Preemptible tier.", "provider": "Nebius", "as_of": "2026-09-10", "accelerator_canonical": "H200-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~400 Gbps/GPU", "contiguity unstated" ] }, { "sku": "NVIDIA HGX H100 (gpu-h100-sxm platform)", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink (platform listed as 'H100 NVLink' in cluster docs)", "inter_node_fabric": "InfiniBand, 400 Gbps per GPU / 3.2 Tbps per 8-GPU node, GPUDirect RDMA (per docs.nebius.com)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 3.85, "pricing_model": "on-demand", "min_commit": "none", "notice_period": null, "source_url": "https://nebius.com/prices", "spec_confidence": "published", "notes": "Per GPU per hour (16 vCPU / 200 GB RAM per GPU). vram_gb is standard SKU value; not stated on pricing page.", "provider": "Nebius", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "A", "grade_reasons": [ "scale-up domain (NVLink (platform listed as 'H100 NVLink' in cluster docs))", "scale-out ~3200 Gbps/GPU", "cluster product claim (contiguity unstated)" ] }, { "sku": "NVIDIA HGX H100 — preemptible", "accelerator": "H100-SXM", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": "InfiniBand, 400 Gbps per GPU / 3.2 Tbps per node (per docs.nebius.com)", "max_contiguous_gpus": null, "regions": null, "price_usd_per_gpu_hr": 2.15, "pricing_model": "spot", "min_commit": "none", "notice_period": null, "source_url": "https://nebius.com/prices", "spec_confidence": "published", "notes": "Preemptible tier.", "provider": "Nebius", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "B", "grade_reasons": [ "scale-up domain present", "scale-out ~400 Gbps/GPU", "contiguity unstated" ] }, { "sku": "HGX H100 on-demand", "accelerator": "H100-SXM (HGX)", "vram_gb": 80, "form_factor": "SXM", "gpus_per_node": null, "intra_node": null, "inter_node_fabric": "3,200 Gbps NVIDIA Quantum-2 InfiniBand available (per homepage); Ethernet tier also offered; pricing page does not state which fabric the $1.99 rate applies to", "max_contiguous_gpus": 1016, "regions": null, "price_usd_per_gpu_hr": 1.99, "pricing_model": "on-demand", "min_commit": "none ('rent one H100 GPU for 1 hour starting at $1.99 without a contract')", "notice_period": null, "source_url": "https://voltagepark.com/pricing", "spec_confidence": "published", "notes": "This is a 'starting at' price — the full rate card (Ethernet vs InfiniBand tiers) is quote-based ('Contact for pricing'). Ethernet tier: 1-1016 GPUs; InfiniBand tier: 8-1016 GPUs; both self-serve, up within 15 minutes. Servers are Dell PowerEdge XE9680 with HGX H100 per homepage (GPU-per-node count not explicitly stated). vram_gb is standard SKU value; not stated on fetched pages.", "provider": "Voltage Park", "as_of": "2026-09-10", "accelerator_canonical": "H100-SXM", "grade": "C", "grade_reasons": [ "scale-out fabric below training threshold or contiguity < 64" ] }, { "sku": "H200 8x (marketplace, cheapest of 1)", "accelerator": "H200", "vram_gb": 140, "form_factor": null, "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": [ ", US" ], "price_usd_per_gpu_hr": 4.606, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://vast.ai/pricing", "spec_confidence": "marketplace", "notes": "median $4.61/GPU-hr across 1 listings of >=8 GPUs; interconnect not exposed by marketplace", "provider": "Vast.ai (live API)", "as_of": "2026-09-10", "accelerator_canonical": "H200", "grade": "U", "grade_reasons": [ "no interconnect or form-factor information published" ] }, { "sku": "H100 PCIE 8x (marketplace, cheapest of 3)", "accelerator": "H100 PCIE", "vram_gb": 80, "form_factor": "PCIe", "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": [ "United Arab Emirates, AE" ], "price_usd_per_gpu_hr": 2.4, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://vast.ai/pricing", "spec_confidence": "marketplace", "notes": "median $3.00/GPU-hr across 3 listings of >=8 GPUs; interconnect not exposed by marketplace", "provider": "Vast.ai (live API)", "as_of": "2026-09-10", "accelerator_canonical": "H100-PCIe", "grade": "D", "grade_reasons": [ "PCIe form factor — no scale-up domain" ] }, { "sku": "A100 SXM4 8x (marketplace, cheapest of 2)", "accelerator": "A100 SXM4", "vram_gb": 40, "form_factor": "SXM", "gpus_per_node": 8, "intra_node": "NVLink (implied by SKU name)", "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": [ "Mississippi, US" ], "price_usd_per_gpu_hr": 0.907, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://vast.ai/pricing", "spec_confidence": "marketplace", "notes": "median $1.12/GPU-hr across 2 listings of >=8 GPUs; interconnect not exposed by marketplace", "provider": "Vast.ai (live API)", "as_of": "2026-09-10", "accelerator_canonical": "A100-SXM", "grade": "C", "grade_reasons": [ "no verifiable scale-out fabric — single-node ceiling" ] }, { "sku": "RTX 5090 8x (marketplace, cheapest of 44)", "accelerator": "RTX 5090", "vram_gb": 32, "form_factor": null, "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": [ "United States, US" ], "price_usd_per_gpu_hr": 0.401, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://vast.ai/pricing", "spec_confidence": "marketplace", "notes": "median $0.61/GPU-hr across 44 listings of >=8 GPUs; interconnect not exposed by marketplace", "provider": "Vast.ai (live API)", "as_of": "2026-09-10", "accelerator_canonical": "RTX5090", "grade": "U", "grade_reasons": [ "no interconnect or form-factor information published" ] }, { "sku": "RTX 4090 8x (marketplace, cheapest of 13)", "accelerator": "RTX 4090", "vram_gb": 24, "form_factor": null, "gpus_per_node": 8, "intra_node": null, "inter_node_fabric": null, "max_contiguous_gpus": 8, "regions": [ "Bulgaria, BG" ], "price_usd_per_gpu_hr": 0.374, "pricing_model": "on-demand", "min_commit": null, "notice_period": null, "source_url": "https://vast.ai/pricing", "spec_confidence": "marketplace", "notes": "median $0.40/GPU-hr across 13 listings of >=8 GPUs; interconnect not exposed by marketplace", "provider": "Vast.ai (live API)", "as_of": "2026-09-10", "accelerator_canonical": "RTX4090", "grade": "U", "grade_reasons": [ "no interconnect or form-factor information published" ] } ], "provider_notes": [ { "provider": "AWS", "fabric_notes": "AWS P5 family: 3200 Gbps EFA per instance (p5=EFA, p5en=EFAv3 on Nitro v5), 900 GB/s NVSwitch intra-node; EC2 UltraClusters published claim of scaling up to 20,000 H100/H200 GPUs with up to 20 exaflops aggregate. P6: EFAv4 at 3.2 Tbps (B200) and 6.4 Tbps (B300) per instance; P6e-GB200 UltraServers publish 72 GPUs per NVLink domain (u-p6e-gb200x72) with 28,800 Gbps EFA per UltraServer and 'up to 130 TB/s' NVLink; UltraCluster scale for P6 stated only as 'tens of thousands of GPUs within a petabit-scale nonblocking network'.", "unverifiable": [ "p5e.48xlarge on-demand price (no entry in the us-east-1 on-demand pricing feed; only p5/p5en/p6-b200/p6-b300 listed)", "P6e-GB200 (GB200 NVL72) on-demand list price — UltraServers not in the us-east-1 on-demand feed; sold via Capacity Blocks/UltraServer channels", "Numeric UltraCluster maximum for P6 family (page says 'tens of thousands of GPUs')" ] }, { "provider": "Google Cloud", "fabric_notes": "Google publishes per-VM network bandwidth in the accelerator-optimized machines doc: a3-highgpu-8g 1,000 Gbps (5 NICs, gVNIC/GPUDirect-TCPX), a3-megagpu-8g 1,800 Gbps (9 NICs, GPUDirect-TCPXO), a3-ultragpu-8g 3,600 Gbps (10 NICs, RoCE CX-7), a4-highgpu-8g 3,600 Gbps (10 NICs, RoCE CX-7, four-way rail-aligned network), a4x-highgpu-4g 2,000 Gbps (6 NICs, RoCE CX-7) and A4X Max (GB300) up to 3,200 Gbps via CX-8 SuperNICs with eight-way rail-alignment. No numeric max-cluster GPU count published on the fetched pages.", "unverifiable": [ "On-demand prices for a3-ultragpu-8g (H200) and a4-highgpu-8g (B200) — Google's docs state these have no on-demand consumption option (Spot/Flex-start/reservations only)", "Official Google pricing-page dollar figures — cloud.google.com pricing pages are client-side rendered and returned no prices to fetch; a3-high/a3-mega prices are from the gcloud-compute.com Billing-data mirror (secondary)", "Maximum contiguous GPU cluster scale per machine series (not stated on fetched pages)" ] }, { "provider": "Microsoft Azure", "fabric_notes": "Azure publishes per-GPU InfiniBand bandwidth on Microsoft Learn: ND H100 v5 and ND H200 v5 each give every GPU a dedicated 400 Gbps NVIDIA Quantum-2 ConnectX-7 InfiniBand link (3.2 Tbps per 8-GPU VM) with GPUDirect RDMA, scaling 'up to thousands of GPUs'. ND GB200 v6 (4 GPUs/VM) has 4x 400 Gb/s Quantum-2 CX7 IB per VM, NVL72 NVLink domains of 72 GPUs (18 VMs/rack, 28.8 Tb/s rack scale-out), and a stated InfiniBand fabric reach of up to 100,000 GPUs. Prices from the official Azure Retail Prices API (Consumption meters, Linux).", "unverifiable": [ "Numeric maximum cluster size for ND H100 v5 / ND H200 v5 (Microsoft states only 'thousands of GPUs')", "ND96isr_H200_v5 price in eastus specifically (API listed eastus2/centralus and others, not eastus, at collection time)" ] }, { "provider": "Oracle Cloud (OCI)", "fabric_notes": "OCI publishes per-node RDMA cluster-network bandwidth in its price-list shape table: BM.GPU.H100.8 = 8x2x200 Gb/s RDMA (3,200 Gbps), BM.GPU.H200.8 = 8x1x400 Gb/s RDMA (3,200 Gbps), BM.GPU.B200.8 = 8x400 Gb/s (3,200 Gbps), BM.GPU.GB200.4 = 4x400 Gb/s InfiniBand, BM.GPU.B300.8 = 8x800 Gb/s, BM.GPU.GB300.4 = 4x800 Gb/s. GPU page claims: up to 3,200 Gb/s RDMA cluster network bandwidth, microsecond latency; OCI Supercluster max scale footnote: 131,072 B200 (incl. GB200 Superchips), 65,536 H200, 32,768 A100, 16,384 H100, 16,384 MI300X. Prices are per-GPU-hour from Oracle's official cetools price API (PAY_AS_YOU_GO); Oracle lists a single global USD list price, no regional table.", "unverifiable": [ "Region-specific price variation (Oracle's price API returns one global USD list price; no per-region GPU pricing published)", "Fabric protocol name (RoCE vs InfiniBand) for BM.GPU.B200.8 — Oracle's table gives only '8x400 Gb/sec' without naming the protocol", "Newer B300/GB300 per-GPU prices were not extracted (shapes appear in the price-list table; only 4 offerings recorded per instructions)" ] }, { "provider": "Hyperstack", "fabric_notes": "Docs (flavors page, fetched) state default inter-VM bandwidth up to 16 Gbps, with SR-IOV-capable network-optimised flavors reaching up to 350 Gbps Ethernet inter-VM for H100 SXM, H100 PCIe and H200 SXM; B200 and L40 flavors listed with default bandwidth only. No InfiniBand mentioned. B300 (288 GB, $7.40/hr OD) announced as coming (on-demand August, reserved private clusters Q4). No MI300X or L40S listed; Hyperstack sells L40 48GB at $1.00/hr OD.", "unverifiable": [ "max_contiguous_gpus", "reserved min_commit and notice_period", "NVLink presence on SXM flavors (not explicit in fetched flavor table)", "B200 inter-node fabric", "intra_node interconnect details" ] }, { "provider": "DataCrunch (rebranded to Verda, verda.com)", "fabric_notes": "datacrunch.io now 301-redirects to verda.com (confirmed via redirect during fetch), i.e. DataCrunch has rebranded as Verda. Pricing page (USD, EUR toggle available) publishes on-demand and ~50%-off spot rates plus a long-term discount schedule (2%/3%/4%/8%/25% for 1-mo/3-mo/6-mo/1-yr/2-yr) rather than absolute reserved prices. GB300 NVL72 also listed ($8.62/hr OD, NVLink v5 mentioned) and RTX PRO 6000 ($1.89/hr). No networking/fabric bandwidth, region, or node-size details were published on the fetched pages. No MI300X listed.", "unverifiable": [ "gpus_per_node options", "intra_node interconnect (except GB300 'NVLink v5')", "inter_node_fabric", "max_contiguous_gpus", "regions", "min_commit", "notice_period" ] }, { "provider": "FluidStack", "fabric_notes": "No public pricing available. https://www.fluidstack.io/pricing returns HTTP 404, and the homepage (fetched 2026-09-10) contains no per-GPU rates, GPU model list, cluster specs, or networking details — it is a mission/company page with no pricing section, consistent with a sales-led/contact model. Third-party trackers (e.g. gpucost.org, via search) cite roughly $2.10/hr for H100 and $2.30/hr for H200, but these are not first-party sources and are not recorded as offerings per registry rules.", "unverifiable": [ "all pricing", "all hardware specs", "fabric", "NVLink", "max_contiguous_gpus", "regions", "commitments" ] }, { "provider": "Scaleway", "fabric_notes": "Official pages publish 'up to 20 Gbps' network bandwidth for H100/H100-SXM instances and 2.5-20 Gbps for L4; NVLink is published for H100-SXM. No InfiniBand or cluster-fabric details published on fetched pages. GPU instances are single-node VMs (max 8 GPUs); no contiguous multi-node cluster size published. Official prices are EUR before tax and zone-dependent (PAR-1/PAR-2/WAW-2); the pricing page only rendered PAR-1 (L4) prices in static fetch, so PAR-2 GPU prices were taken from third-party tracker gpuperhour.com and marked secondary. No A100, H200, B200 or MI300X listed.", "unverifiable": [ "official EUR rates for H100/H100-SXM/L40S/B300 in PAR-2 (client-rendered)", "H100-SXM gpus-per-instance options", "inter_node_fabric beyond 20 Gbps Ethernet", "max_contiguous_gpus", "reserved pricing", "min_commit", "notice_period" ] }, { "provider": "OVHcloud", "fabric_notes": "Public Cloud GPU page (USD, Worldwide site) publishes 'Up to 25 Gbps included' for public and private networks and '1-4 GPUs per instance' as general specs; H200 is the exception at 8 GPUs/instance with NVLink. No InfiniBand or multi-node cluster fabric published on fetched pages, and no contiguous cluster product described. A100 no longer appears on the GPU lineup page (only H200, H100, L40S, L4, A10 $1.00/hr, V100S $0.88/hr, Quadro RTX 5000 $0.60/hr). No B200, MI300X. The main /public-cloud/prices/ price list is client-rendered and did not expose GPU rows to static fetch.", "unverifiable": [ "H200 exact instance name and per-GPU list price", "regions per GPU type", "max_contiguous_gpus", "reserved/savings-plan pricing", "min_commit", "notice_period", "H100/L40S NVLink details" ] }, { "provider": "Together AI (GPU Clusters)", "fabric_notes": "GPU Clusters page publishes 'high-speed InfiniBand' interconnect for all clusters, managed Kubernetes or Slurm-on-Kubernetes orchestration, on-demand with no commitment, and reserved terms up to 6 months with upfront payment. Capacity described as 25+ cities: USA (600MW), Europe (150MW+), Asia and Middle East. Also listed reserved-only with contact-sales pricing: HGX B300 (270 GB), GB200 NVL72 (186 GB, 512 to 1,000+ GPUs), GB300 NVL72 (288 GB) — not recorded as offerings since no public price. No A100, L40S, or MI300X on the clusters page. InfiniBand generation/bandwidth (e.g. NDR 3.2Tbps) not specified on fetched page.", "unverifiable": [ "gpus_per_node", "intra_node NVLink (not explicitly stated for B200/H200/H100)", "InfiniBand generation/bandwidth", "form_factor", "exact reserved minimum term", "regions per SKU" ] }, { "provider": "CoreWeave", "fabric_notes": "CoreWeave prices all GPUs as fixed-size instances (per node), not per GPU; per-GPU figures above are derived by division and noted. HPC interconnect docs (fetched) state two RDMA fabrics — NVIDIA Quantum-2 and Quantum-X800 InfiniBand — in non-blocking fat-tree architecture, plus Spectrum-X RoCE on GB300 instances (800 Gbps options for GB300). B200 release notes give the most explicit per-instance networking (8x ConnectX-7 400G NDR + BlueField-3). GB200 NVL72 instances: 4 GPUs each, deployed only as full 18-node racks, 400 Gb/s NDR IB, Quantum-2. GB300 NVL72, HGX B300 on-demand, and RTX PRO 6000 (standard) are 'Contact sales'. Reserved/committed pricing is not published as a rate card (marketing claims of up to ~60% off exist in secondary sources only).", "unverifiable": [ "max_contiguous_gpus (any SKU)", "per-node InfiniBand bandwidth for H100/H200/A100 instances", "intra-node NVLink/NVSwitch topology per instance", "reserved/committed rate card", "NVL72 NVLink domain size (72 GPUs not stated on fetched pages)" ] }, { "provider": "Lambda", "fabric_notes": "Lambda prices per GPU per hour directly. 1-Click Clusters: 16 to 2,000+ GPUs on NVIDIA Quantum-2 InfiniBand with SHARP; per-node InfiniBand bandwidth (e.g. 3.2 Tbps) is NOT stated on the fetched pricing/1CC pages. No H200 on-demand instances listed as of fetch date; GH200 96GB available at $2.29/hr (1x). NVLink/NVSwitch intra-node topology not stated. No egress fees. Regions not published on fetched pages.", "unverifiable": [ "per-node InfiniBand bandwidth for 1-Click Clusters", "intra-node NVLink/NVSwitch details", "regions", "1yr+ reserved rate card (custom pricing)", "exact upper bound of cluster size ('2,000+')" ] }, { "provider": "Crusoe", "fabric_notes": "Crusoe prices per GPU per hour. Docs VM overview (fetched) publishes exact per-instance InfiniBand bandwidth: H100/H200/B200/MI300X 8x at 3200 Gbps, A100-SXM-IB and GB200 (gb200-186gb-nvl-4x, 4 GPUs) at 1600 Gbps, B300 8x at 6400 Gbps, MI355X 8x at 3200 Gbps RoCE. However, B200 HGX (180GB), GB200 NVL72 (186GB), and MI355X have NO public prices — contact sales only. Networking docs state InfiniBand is the basis of GPU-to-GPU communication, is included in IB-enabled VM cost, and there are no ingress/egress charges. Fabric generation (Quantum-2 etc.), max contiguous cluster size, reserved/committed pricing, and regions are not published on fetched pages.", "unverifiable": [ "max_contiguous_gpus", "InfiniBand fabric generation (Quantum-2/NDR not named on fetched pages)", "intra-node NVLink/NVSwitch details", "regions per SKU", "reserved/committed pricing", "B200, GB200 NVL72, MI355X prices (contact sales)" ] }, { "provider": "Nebius", "fabric_notes": "Nebius prices per GPU per hour with both on-demand and preemptible tiers published. Cluster docs (fetched) publish the clearest fabric spec of the five providers: each GPU has a 400 Gbps NIC, 8-GPU VMs give 3.2 Tbps per node, GPUDirect RDMA, InfiniBand traffic isolated per cluster; 13 named physical fabrics are listed but max cluster size is not published. Cluster-capable platforms: gpu-h100-sxm, gpu-h200-sxm, gpu-b200-sxm, gpu-b300-sxm. HGX B300 also priced ($7.85 on-demand / $4.30 preemptible). GB200 NVL72 and GB300 NVL72 are listed but 'Contact us' — no public price. Reserved/committed pricing exists only as a generic 'up to 35% less' claim — no rate card published. Fabric generation name (Quantum-2) appears on marketing pages not fetched; recorded fabric details are from docs only.", "unverifiable": [ "max_contiguous_gpus / max cluster size", "InfiniBand fabric generation on fetched docs page", "VRAM per GPU (not stated on Nebius pages; standard SKU values used)", "reserved/committed rate card (only 'up to 35% less' claim)", "GB200 NVL72 and GB300 NVL72 prices (contact sales)", "region-to-SKU mapping" ] }, { "provider": "Voltage Park", "fabric_notes": "Voltage Park publishes structure but almost no rate card: three tiers on the pricing page — (1) On-Demand Ethernet, 1-1016 HGX H100 GPUs; (2) On-Demand 3200 Gbps InfiniBand, 8-1016 HGX H100 GPUs; (3) Long-Term Reserve, 32-8000+ H100 GPUs on 3200 Gbps InfiniBand with 6+ month contracts — all 'Contact for pricing' except the blanket 'starting at $1.99/hr' on-demand claim. Homepage states 3,200 Gbps NVIDIA Quantum-2 InfiniBand, a fleet of 36,000 NVIDIA Hopper/Blackwell/Grace Blackwell GPUs, and Dell PowerEdge XE9680 HGX H100 servers. Cluster-size figures are inconsistent across their own pages: pricing page caps on-demand at 1,016 GPUs, product page says 64-4,064 GPUs per cluster, homepage references clusters 'up to 8,176'. Blackwell and Grace Blackwell GPUs are mentioned as available with no SKUs or pricing. Note: third-party reporting says Voltage Park merged into Lightning AI (Jan 2026) with the brand still operating — not confirmed on fetched Voltage Park pages.", "unverifiable": [ "full on-demand rate card (Ethernet vs InfiniBand price split)", "reserved pricing", "GPUs per node (Dell XE9680 HGX H100 stated, count not stated)", "intra-node NVLink details", "regions/data center locations", "consistent max cluster size (1,016 vs 4,064 vs 8,176 across their own pages)", "Blackwell/GB200 SKUs and pricing" ] }, { "provider": "Vast.ai (live API)", "fabric_notes": "Public bundles API exposes price, GPU count, VRAM, geolocation and PCIe details per listing, but nothing about inter-node fabric or NVLink topology. Multi-node contiguity is not a marketplace concept here: each listing is a single host.", "unverifiable": [ "inter_node_fabric", "oversubscription", "max_contiguous_gpus beyond single host" ] } ] }