NVIDIA H200 GPU Cloud Pricing 2026: Every Provider Compared

NVIDIA H200 Cloud Pricing — Live Table (April 2026)

GridStackHub tracks NVIDIA H200 pricing across 10 cloud providers. The H200 SXM is available on-demand from specialist providers starting at $2.99/hr, while hyperscalers (AWS, Google Cloud, Azure) primarily offer H200 via reserved capacity. Here is every provider we track:

*RunPod H200 NVL community cloud pricing reflects spot/shared-node availability and may not be consistently available. SXM5 instances are the standard for production workloads. Data sourced from GridStackHub's live pricing database, April 21, 2026. VERIFIED = confirmed via live provider API. Prices subject to change — check provider directly before committing.

NVIDIA H200 vs H100: Full Specification Comparison

The H200 is not a new architecture — it shares the Hopper GPU with the H100 but adds a dramatically upgraded memory system. Here is what changed and what stayed the same:

The story in two sentences: H200 and H100 have identical compute throughput — same TFLOPS, same Hopper architecture, same CUDA code runs unchanged. The difference is entirely in memory: 141GB HBM3e at 4.8 TB/s versus 80GB HBM3 at 3.35 TB/s. That memory delta determines when to choose each GPU.

When H200 Beats H100: Use Case Guide

Both GPUs run the same CUDA code. The decision is about memory: does your workload benefit from more VRAM, faster bandwidth, or both?

Choose H200 when:

• You're serving Llama 3.1 70B or similar models in BF16 precision. Llama 3.1 70B requires approximately 140GB at BF16. A single H200 (141GB) fits it with room for a KV cache. On H100, you'd need two GPUs with tensor parallelism, doubling cost and adding 15–25% latency from cross-GPU communication.
• Inference throughput is your primary constraint. For any autoregressive decoding workload, the H200's 4.8 TB/s bandwidth versus H100's 3.35 TB/s delivers roughly 43% faster token generation — directly improving latency and increasing tokens-per-dollar at any batch size.
• You're running long-context inference (32K+ token windows). KV cache memory scales linearly with context length. At 128K context windows, KV cache alone can exceed 40GB for a 70B model — eliminating the H100's working memory almost entirely and making the H200 the only option for single-GPU operation.
• You're building RAG pipelines with large retrieval contexts. Each retrieved document added to context increases KV cache memory. H200's 141GB gives meaningful headroom over H100's 80GB for retrieval-augmented workloads at scale.
• You want to deploy future models on existing hardware. Next-generation open models (Llama 4, future Mixtral variants) will continue growing in parameter count. H200's 141GB provides a larger runway before you need multi-GPU inference infrastructure.

Choose H100 when:

• You're training or fine-tuning at scale. H100 and H200 have identical compute throughput for training. H100 is cheaper per hour and has broader multi-node cluster availability. For long training runs, H100 is often the more cost-effective choice.
• Your model fits in 80GB and you're compute-bound. Models under 40B parameters at BF16 fit comfortably in H100's 80GB. If your bottleneck is FLOPS rather than memory bandwidth, H100's lower cost per hour wins on economics.
• You need the widest provider choice and maximum availability. H100 is available from 15+ providers with better reserved-instance options and more competitive spot pricing. If availability SLAs matter, H100 has a deeper supply pool.
• Cost per hour is your primary constraint and model fits H100. Lambda H200 at $2.99/hr vs H100 at ~$1.74/hr is a 72% premium. If you don't need the extra memory or bandwidth, that premium doesn't pay back.

What Models Fit on a Single NVIDIA H200 (141GB)?

141GB HBM3e is a meaningful jump from H100's 80GB. Here's a practical reference for which models fit at various precision levels, including KV cache headroom:

*Weight estimates assume 2 bytes/parameter for BF16 and ~0.5 bytes/parameter for 4-bit. Production inference requires additional memory for KV cache — estimate 20–40% overhead for typical serving configurations. Larger batches or longer context windows require proportionally more KV cache memory.

H200 Availability in 2026: What to Expect

H200 supply is constrained by NVIDIA production capacity and aggressive allocation by AI hyperscalers and foundation model labs. Here is the realistic availability picture by provider category:

Ask GridStackHub About H200 Pricing

Get answers from live pricing data — compare providers, estimate monthly costs, or evaluate H200 vs alternatives for your specific workload.

Frequently Asked Questions

Track H200 Prices and Get Alerts

H200 pricing is moving in 2026 as production scales and providers compete for AI inference workloads. GridStackHub tracks every change daily — here is how to stay ahead: