NVIDIA H100 vs H200: which one your workload actually needs.
Same architecture, same software stack, very different memory. The choice is almost always decided by one question, and it is not the one most buyers start with.
The difference in one table.
| Dimension | NVIDIA H100 | NVIDIA H200 |
|---|---|---|
| Architecture | Hopper | Hopper |
| Memory | 80GB HBM3 | 141GB HBM3e |
| Software stack | Identical: CUDA, drivers, frameworks | Identical: CUDA, drivers, frameworks |
| Indicative node price | Around $240,000 refurbished, 8-GPU node | Around $395,000 new, 8-GPU node |
| Indicative rental | Roughly $1.49 to $3.50 per GPU-hour | From roughly $3.72 per GPU-hour |
| Current verified stock | 580 nodes, US located, 7 to 25 days | 200 nodes, EU located, 10 days |
| Wins on | Price per unit of compute, availability, secondary market depth | Memory capacity and bandwidth, long context, larger models per node |
Node pricing from SSP verified inventory, August 2026. Rental ranges from published provider pricing and the GetDeploying GPU Price Index, 2026. Specifications from NVIDIA published documentation.
Does your model fit in 80GB?
If it does, comfortably, with room for the batch sizes and context lengths you actually run, then the H100 is the better buy and it is not close on price. The Hopper software stack is identical, so nothing about your pipeline changes.
If it does not, the H200's extra memory stops you buying more nodes than your compute requires, and that arithmetic usually pays for the premium on its own. Buying additional H100 nodes to work around a memory ceiling is the expensive way to solve a memory problem.
The honest third option: if you are reading this while planning a frontier training run, the real comparison is against B300, not between these two. See B200 vs B300.
Asked first, answered straight.
On memory-bound work, frequently. H200 carries 141GB of HBM3e against the H100's 80GB HBM3, which means larger models and longer contexts fit per node. If your workload sits comfortably inside 80GB, the H100 at current refurbished pricing is very hard to beat.
In our verified channel, refurbished H100 nodes sit near $240,000 per 8-GPU node and new H200 nodes near $395,000. On the rental side, H100 SXM runs roughly $1.49 to $3.50 per GPU-hour on demand and H200 starts near $3.72.
Yes, both are Hopper generation, so CUDA, drivers and frameworks behave the same way. Moving between them is not a migration, which is part of why H200 is an easy upgrade decision when memory is the constraint.
It moves with the cycle. In our current inventory H100 has the deeper pool at 580 nodes US-located, while H200 sits at 200 nodes EU-located with a ten-day lead. EU location matters for European buyers because it removes the import event entirely.
Working through this on a real requirement?
Twenty minutes with the desk, no pitch and no quote at the end of it. Tell us roughly what you need and we will come back within one business day.
Sent. We are on it.
Your enquiry has landed with the desk. Acknowledged within one hour.