The H200 is an H100 with 141GB instead of 80GB. It rents from ~$2.63/hr (Vast spot) to $6.31 (CoreWeave). What it costs and when the memory is worth paying for.
This post contains affiliate links. If you buy through them, Aliteq may earn a commission — at no extra cost to you. Prices verified at publish time.
Share
The H200 is the memory upgrade to the H100 — same Hopper speed, but 141GB of faster HBM3e instead of 80GB — and it rents for a premium. As of 24 Sep 2026 it starts around $2.63/hour on Vast.ai spot and $3.59/hour on Runpod on-demand, with the enterprise neoclouds charging $4.29–$6.31 per published rates. Here's what the H200 costs to rent, where, and when its extra memory is worth paying for. Part of our cloud-GPU pricing pillar.
The H200 is an H100 with more, faster memory — 141GB of HBM3e versus 80GB — and roughly the same compute. So the premium buys capacity, not speed: it earns its keep when a model or a long context won't fit an 80GB H100 on a single card (avoiding a two-GPU setup), or when memory bandwidth is your bottleneck for high-throughput inference. If your workload fits in 80GB, rent an H100 and pocket the difference. For the generation above, see H100 vs B200.
Referral link
Rent the cheapest H200 on Vast.ai
The lowest H200 rate is on Vast.ai spot — from about $2.63/hr as of 24 Sep 2026, well under the neoclouds' $4.29–$6.31. Check availability and use a verified host.
Referral link — we may earn a commission at no cost to you. Prices on our compare page are the provider's live figures, cheapest first; this never changes the ranking.
Frequently asked
How much does it cost to rent an H200?
As of 24 September 2026, from about $2.63/hour on Vast.ai spot and $3.59/hour on Runpod on-demand, rising to $4.29–$6.31 on enterprise neoclouds (Crusoe, Nebius, CoreWeave) per their published rates. H200 spot supply is limited, so the cheapest rates depend on availability.
Is the H200 worth it over an H100?
Only if you need the memory. The H200 has 141GB versus the H100's 80GB but similar compute, so it's for models or long contexts that won't fit 80GB on one card, or memory-bandwidth-bound inference. If your workload fits in 80GB, an H100 is cheaper for the same speed.
What can an H200 run that an H100 can't?
Larger models on a single card without splitting across two GPUs, and longer context windows, thanks to its 141GB. Its extra HBM3e bandwidth also helps high-throughput serving. For pure compute the two are close; the difference is memory.
Referral link
Managed pods + serverless
Want a stable, managed H200? Runpod on-demand is about $3.59/hr (24 Sep 2026).
