NVIDIA H200 Rental

141GB of HBM3e per GPU.

HGX H200 brings Hopper compute together with 141GB of HBM3e per GPU and 4.8 TB/s memory bandwidth. The right rental when KV cache, context length, or batch size — not compute — is your bottleneck.

1.1 TB HBM3e per node

8 × 141GB HBM3e — fit a 405B Llama 3.1 in FP8 on a single server with room for a real KV cache.

4.8 TB/s per-GPU

Memory bandwidth jumps 43% vs H100. Token-generation throughput on long context scales nearly linearly.

Hopper compute

Same 3,958 TFLOPS FP8 sparse as H100 — your training and inference code runs unchanged.

Specifications

GPU (HGX)
8 × NVIDIA H200 SXM5
141GB HBM3e · 4.8 TB/s each
CPU
2 × Intel Xeon Platinum 8480+
112 cores total
Memory
2 TB DDR5
32 × 64GB @ 4800 MT/s
Storage
Up to 60 TB NVMe
Configurable scratch tier
Fabric
8 × NDR 400G IB
Rail-optimized multi-node
Power
6 × 3,000W PSU
208V/415V redundant
Best fit: long-context inference

32K, 128K, 1M-token contexts where KV cache size dominates. H200 doubles serveable concurrency vs H100 at the same latency.

Best fit: MoE serving

Mixture-of-experts models (Mixtral, DeepSeek, Llama 4) benefit directly from more HBM and bandwidth — fewer cards, simpler topology.

Drop-in for H100 code

CUDA, cuDNN, NCCL, vLLM, TensorRT-LLM — no changes. Just more memory and more bandwidth.

Long-context LLM inference

Modern serving is memory-bound, not compute-bound. Once context windows pass 32K tokens, KV cache — not matmul — decides how many concurrent users a node can serve. H200's 141GB of HBM3e per GPU and 4.8 TB/s bandwidth let you keep more sessions resident and feed the tensor cores without stalling.

128K–1M token contexts

Llama 3.1 405B at 128K, Claude-style document QA, and full-codebase agents stay resident in HBM. No paging to host memory, no tail-latency cliff.

vLLM, TensorRT-LLM, SGLang

Drop-in with H100 stacks. PagedAttention, continuous batching, and FP8 KV cache compression all benefit directly from the extra memory headroom.

MoE serving (Mixtral, DeepSeek)

Mixture-of-experts weights fit in fewer cards. Fewer GPUs per replica means simpler topology, lower NCCL overhead, and better $/token in production.

Higher batch, lower $/token

43% more memory bandwidth than H100 translates almost linearly to token-generation throughput at the same latency SLA.

Training, fine-tuning, and RAG pipelines

H200 keeps Hopper's 3,958 TFLOPS FP8 sparse compute and adds the memory you need for large-batch SFT, RLHF, and embedding pipelines. Code that ran on H100 runs unchanged — you just stop hitting OOMs.

SFT, DPO, RLHF at scale

Larger effective batch sizes without gradient accumulation. Reward-model passes and PPO rollouts share HBM with the policy model on a single node.

Embedding & RAG indexing

Bulk-embed corpora 1.4–2x faster than H100. Pair with NVMe scratch for end-to-end vector index builds without spilling to host RAM.

405B in FP8 on one node

Fit Llama 3.1 405B in FP8 on an HGX H200 with real KV cache headroom — no cross-node tensor parallelism required for inference.

Mixed train + serve

Fine-tune by day, serve by night on the same hardware. No re-quantization surprises, no migration between SKUs.

Privacy, security, and compliance

Long-context inference often means sensitive data — full medical records, internal codebases, legal discovery, customer chat history. H200 nodes ship as bare metal so the only software with access to that HBM is yours.

Single-tenant bare metal

No hypervisor, no shared GPU, no neighbor tenant. Confidential Computing on Hopper available for attested workloads.

HIPAA, SOC 2, regulated AI

Customer-held FDE keys, NIST 800-88 drive sanitization between tenants, and BAAs available for healthcare deployments.

On-prem & air-gapped

Identical H200 builds deployable in your facility or a dedicated cage — useful for ITAR, FedRAMP-track, or contract-isolated inference.

Private fabric by default

No public egress unless you ask for it. Direct Connect / ExpressRoute peering keeps prompts and completions off the open internet.

Frequently asked

How is H200 different from H100?+

Same Hopper GPU compute, but with 141GB of HBM3e per GPU (vs 80GB HBM3 on H100) and 4.8 TB/s memory bandwidth (vs 3.35 TB/s). For memory-bound workloads — long-context inference, larger batch sizes, KV cache — H200 delivers a real 1.4–2x speedup over H100 with no code changes.

When can I get an H200?+

We have HGX H200 allocations landing on a rolling basis. Typical lead time today is 6–10 weeks. Reserve a slot early — H200 supply is tighter than H100 was at launch.

Is H200 worth the premium over H100?+

For inference and fine-tuning of long-context models (32K+ tokens) and large MoE architectures, yes. For pure compute-bound pretraining of dense models under 32K context, H100 SXM5 still wins on price/performance.

H200 allocation is tight.

Reserve early — we'll hold a slot while you finalize specs.

Reserve an H200