NVIDIA H200 GPU prices
H200 adds memory capacity and bandwidth for memory-bound training and inference. Compare the delivered node, commitment, and region alongside the normalized rate.
Dataset checked 2026-09-08. Prices and availability may change.
Current public offers
Source-linked records with the published billing context preserved.
| Provider and node | Region | Term | Normalized price | Source |
|---|---|---|---|---|
| FluidCore8x H200 141GB bare-metal node | Not published | hourly | $2.51/GPU-hr20.08 | Provider source ↗Checked 2026-08-23 |
| Clore.ai8x H200 Colorado bare-metal node | Colorado, USA | 360 days | $2.59/GPU-hr179020.80 | Provider source ↗Checked 2026-08-23 |
| Leaseweb1x H200 (HP DL385 G11) | London | term not stated | $2.80/GPU-hr£1,503.02/mo | Provider source ↗Checked 2026-08-20 |
| Taiga Cloud8x H200 SXM | Chester, UK | hourly | $3.00/GPU-hrFrom $3.00/GPU/hr (bare-metal tier) | Provider source ↗Checked 2026-09-04 |
| Qubrid AI8x H200 141GB bare-metal server | Not published | 12 months | $3.00/GPU-hr210240.00 | Provider source ↗Checked 2026-08-23 |
| DataPacket8x H200 NVL | Not published | 3–6-mo prepay typ. | $3.06/GPU-hr15368 | Provider source ↗Checked 2026-09-08 |
| Gcore8x H200 | Not published | hourly | $3.17/GPU-hr21.84 | Provider source ↗Checked 2026-09-04 |
| Hostrunway2x H200 (GPU Montreal 9) | Montreal | no lock-in | $4.45/GPU-hr$6,490/mo | Provider source ↗Checked 2026-08-20 |
| Cirrascale8x H200 | US | 1-mo min | $4.54/GPU-hr$26,499/mo (annual commit $21,199/mo) | Provider source ↗Checked 2026-08-20 |
| Hostrunway1x H200 dedicated GPU server (dual EPYC 9224) | Montreal | no lock-in; 15% discount at 12 months | $4.89/GPU-hr3572.00 | Provider source ↗Checked 2026-08-29 |
| CoreWeave8x H200 bare-metal GPU node | North America / Europe | on-demand | $6.31/GPU-hr50.44 | Provider source ↗Checked 2026-08-23 |
| Oracle Cloud8x H200 (BM.GPU.H200.8) | Select OCI regions | no commitment | $10.00/GPU-hr$10.00/GPU/hr (pay as you go) | Provider source ↗Checked 2026-08-20 |
| OpenMetal1x H200 NVL bare-metal server (starting configuration) | Ashburn, Virginia | monthly / sales quote | Not publishedmonthly / sales quote | Provider source ↗Checked 2026-09-08 |
| Impossible Cloud8x H200 bare-metal node | European Union | sales quote | Not publishedsales quote | Provider source ↗Checked 2026-08-23 |
| LeaderGPU1x H200 141GB bare-metal server | Netherlands | 1 day | Not published2986.9 | Provider source ↗Checked 2026-08-23 |
| Massed Compute8x H200 bare-metal cluster node | Not published | custom quote | Not publishedcustom quote | Provider source ↗Checked 2026-08-23 |
| DigitalOcean8x HGX H200 141GB SXM5 | NYC1 / AMS3 | sales contract | Not publishedcontract-based | Provider source ↗Checked 2026-09-08 |
| GlassPearl8x H200 whole bare-metal server | Not published | quote-only fixed whole-server contract | Not publishedquote-only fixed whole-server contract | Provider source ↗Checked 2026-08-23 |
| Sakura InternetKoukaryoku PHY: 8x H200 SXM 141GB | Ishikari, Japan | monthly / sales consultation | Not publishedmonthly / sales consultation | Provider source ↗Checked 2026-08-23 |
| NAVER Cloud Platform8x H200 141GB bare-metal server | South Korea | quote | Not publishedquote | Provider source ↗Checked 2026-08-23 |
When H200 changes the economics
A higher hourly rate can still reduce delivered workload cost when additional memory avoids model partitioning, lowers communication overhead, or supports larger inference batches. Benchmark the intended model and serving stack before committing.
What to verify in an H200 quote
Record the exact GPU and node form factor, GPU count, host platform, interconnect, network, billing basis, region, and delivery date. For confidential workloads, request the supported CPU/GPU attestation path and the provider operating procedure.
Questions buyers ask
Is H200 always more cost-effective than H100?
No. The result depends on model memory, batch size, utilization, topology, and provider terms. Compare delivered throughput or completed-job cost.
Are H200 prices directly comparable?
Only after normalizing GPU count and time while preserving the original commitment, billing basis, region, and configuration.