NVIDIA H200 GPU Cloud: Pricing and Availability in 2026

What Is the H200?

The NVIDIA H200 is the successor to the H100, built on the same Hopper GPU die but paired with next-generation **HBM3e memory**. It represents one of the largest memory bandwidth upgrades in NVIDIA's data-centre history.

Key Specifications

| Spec | H200 SXM | H100 SXM |

|------|----------|----------|

| VRAM | 141 GB | 80 GB |

| Memory type | HBM3e | HBM2e |

| Memory bandwidth | 4.8 TB/s | 3.35 TB/s |

| FP8 TFLOPS | ~2000 | ~1979 |

| TDP | 700 W | 700 W |

| NVLink bandwidth | 900 GB/s | 900 GB/s |

The H200 does not dramatically increase raw compute (FP8 TFLOPS are similar), but the **76% increase in memory capacity** and **43% jump in bandwidth** are transformative for memory-bound workloads.

Cloud Pricing in 2026

| Provider | H200 Price/hr | Notes |

|----------|--------------|-------|

| RunPod | $4.49 | Secure Cloud, SXM |

| Lambda Labs | $4.99 | Cluster available |

| CoreWeave | $4.25–4.75 | Reserved discounts available |

H200 availability is still limited compared to H100 — book early for large cluster runs.

H200 vs H100: When Does It Matter?

H200 Wins

**Very large language models (70B+ parameters)**

At 141 GB, a single H200 can hold a full Llama 3 70B model in FP16 without tensor parallelism. An H100 requires sharding across two GPUs for the same task — doubling the interconnect overhead.

**Long-context inference**

KV cache grows linearly with sequence length. At 128K context, the KV cache for a large model can consume 40–60 GB. The H200's extra VRAM lets you serve longer contexts without aggressive KV cache eviction.

**High-throughput batched inference**

More VRAM means larger batch sizes. Larger batches mean better GPU utilisation and more tokens per dollar at scale.

H100 Still Wins

**Training smaller models (up to 30B)**

For most fine-tuning runs on 7B–30B models, H100 memory is sufficient and the lower price (~$2.49–2.89/hr vs $4.49/hr) means dramatically lower training cost.

**Cost-sensitive experimentation**

At roughly 60% of the H200 price, H100 is the right tool for iterating on ideas before committing to a full training run.

Who Needs the H200?

Who Should Stick with H100?

Availability Outlook

H200 supply is ramping in 2026 but demand from frontier labs is intense. Cloud spot pricing can be significantly cheaper than on-demand when available. Monitor BestGPUCloud for real-time H200 availability across providers.

Find the cheapest H200 GPU cloud pricing → →