An H100 vs H200 search is a renter deciding whether 80 GB failed. Both cards are Hopper. The H200 keeps that compute and replaces the memory system: 141 GB of HBM3e at 4.8 TB/s, against 80 GB of HBM3 at 3.35 TB/s on the H100 SXM.
The numbers below are NovaPod's book on November 5, 2026. A floor is the cheapest offer. A median is the middle of the book. I am not blending them into one hourly rate, and I am not mixing the SXM with the PCIe H100.
H100 vs H200 on one afternoon's book
The H100 SXM started at $1.21 an hour. The median was $1.68. There were 65 offers, each with 80 GB. The H200 started at $2.76, and the median was also $2.76, across 82 offers and 141 GB. The H200 floor and the middle of the book were the same number. The H100 floor was a cheaper host, not the typical offer.
The gap at the floor is $1.55 an hour. That is the price of 61 extra gigabytes and the faster memory, not a new generation of tensor cores. If the job already fits in 80 GB, the H200 hour is unused capacity.
| Card | Memory | Bandwidth | Floor | Median | Offers |
|---|---|---|---|---|---|
| H100 SXM | 80 GB HBM3 | 3.35 TB/s | $1.21/hr | $1.68/hr | 65 |
| H200 | 141 GB HBM3e | 4.8 TB/s | $2.76/hr | $2.76/hr | 82 |
What does not change between them
Both are Hopper. FP8 through the Transformer Engine is on the H100 and still on the H200. NVLink 4 on the SXM parts is the same class of link. A CUDA image that runs on the H100 does not need a new software stack for the H200.
What does change is the memory the decode step can hold. A 70B model in FP16 is about 140 GB of weights before the KV cache. That does not fit on one H100. It fits on one H200, with little room left. FP8 on the H100 is the other way to hold a 70B, if the quality loss is acceptable.
Blackwell is a different comparison. A B200 on the same book started at $4.38 an hour with 192 GB. Do not read an H200 quote as a B200 quote.
When the H200 rental price is the one to pay
Pay it when an H100 job already died on memory. Long context, a KV cache that outgrows 80 GB, or weights that do not quantize cleanly are the usual reasons. The H200 had 82 offers that day, so finding one was not the hard part. The hard part is knowing the extra 61 GB is the failure you hit.
There was also an H200 NVL row at a $2.29 floor, 76 offers, still 141 GB. That is a different board. If a quote says H200, ask whether it is the SXM row at $2.76 or the NVL row. This note's $2.76 is the SXM-class row in the November 5 book.
When the H100 SXM is the rental
Stay on the H100 SXM for FP8 fine-tunes and serving that already fit in 80 GB. The floor was $1.21 and the median was $1.68, so budget the median if you need more than the single cheapest host. Sixty-five offers is a real book and a thinner one than a consumer card.
An A100 80 GB is the cheaper neighbor when you do not need FP8. It is not an H100. A PCIe H100 is not this SXM row either. The comparison here is the SXM at 80 GB against the H200 at 141 GB.
Questions
Is the H200 faster than the H100?
The compute generation is still Hopper. The H200 change is memory: 141 GB of HBM3e at 4.8 TB/s, against 80 GB of HBM3 at 3.35 TB/s on the H100 SXM. Memory-bound jobs feel that. A job that already fit does not.
What was the H100 vs H200 rental price on November 5, 2026?
On NovaPod the H100 SXM started at $1.21 an hour, median $1.68, 65 offers. The H200 started at $2.76, median $2.76, 82 offers. Billing is per second, so the clock stops when the pod stops.
Can one H100 hold a 70B model?
Not in FP16. Those weights are about 140 GB before the KV cache. One H200 can hold them. FP8 on an H100 is the other fit, if you accept the quantization.
Should I skip both and rent a B200?
Only after 141 GB has failed. The B200 on the same book started at $4.38 an hour for 192 GB. That is a Blackwell step, not the next dollar on an H100.
- H100 SXM rental page
- H200 rental page
- H100 rental price on November 4
- H200 and B300 rental prices
- Live GPU marketplace