H100 vs H200 vs B200: specs, price per hour, and which one to rent.
Three data-center GPUs, two architectures, one question: which hour buys the most work. The specifications are the vendor datasheets, the prices are our catalogue and the market index, and the fit tables are arithmetic on memory. No benchmark of ours is quoted; the ratios are computed from the published figures.
- H100
- $0.55/h spot80 GB · $0.89 on-demand · 96 available
- H200
- $0.75/h spot141 GB · $1.09 on-demand · 32 available
- B200
- $1.69/h spot180 GB · $2.49 on-demand · 16 available
- Market median, H100
- $2.50/h68 providers, on-demand
What the datasheets say.
Memory decides what fits; bandwidth decides how fast tokens come out; tensor throughput decides how fast a training step runs. Everything else is detail.
| H100 SXM | H200 SXM | B200 SXM | |
|---|---|---|---|
| GPU memoryDecides which models fit at all | 80 GB HBM3 | 141 GB HBM3e | 180 GB HBM3e |
| Memory bandwidthSets token throughput on inference | 3,350 GB/s | 4,800 GB/s | 7,700 GB/s |
| FP16 tensor | 990 TFLOPS | 990 TFLOPS | — |
| FP8 tensorHalves memory and doubles throughput where the stack supports it | 1,979 TFLOPS | 1,979 TFLOPS | — |
| Architecture | Hopper (2022) | Hopper (2024) | Blackwell (2025) |
| Board power | 700 W | 700 W | 1000 W |
| GPU-to-GPU link | NVLink 4, 900 GB/s; PCIe Gen5 128 GB/s | NVLink 4, 900 GB/s; PCIe Gen5 128 GB/s | NVLink 5, 1.8 TB/s GPU-to-GPU; PCIe Gen5 128 GB/s |
| Node sizes here | 1×, 2×, 4×, 8× | 1×, 2×, 4×, 8× | 1×, 2×, 4×, 8× |
Sources are linked on each model page: H100 SXM · H200 SXM · B200 SXM. Computed from the datasheets: the B200 has 2.3× the bandwidth and more the FP8 throughput of the H100; the H200 has 1.4× the H100's bandwidth on the same architecture.
Three tiers here, and what the rest of the market charges.
Our prices are per GPU-hour, identical in every region and at every node size, billed per minute. The market columns are the lowest and the median published on-demand rate across the providers in our index for the same GPU.
| GPU | Spot | On-demand | Reserved | 100 h on spot | Market low | Market median | Providers |
|---|---|---|---|---|---|---|---|
| H100 SXMreclaim rate <5% · 30 d | $0.55 | $0.89 | $0.65 | $55.00 | $1.15 | $2.50 | 68 |
| H200 SXMreclaim rate 5–10% · 30 d | $0.75 | $1.09 | $0.79 | $75.00 | $1.39 | $3.89 | 54 |
| B200 SXMreclaim rate 5–10% · 30 d | $1.69 | $2.49 | $1.85 | $169.00 | $3.12 | $5.37 | 39 |
Cost per unit of throughput, H100
$0.45 per PFLOP-hour of FP8 on-demand, $0.28 on spot. $0.27 per TB/s of bandwidth per hour.
Cost per unit of throughput, H200
$0.55 per PFLOP-hour of FP8 on-demand, $0.38 on spot. $0.23 per TB/s of bandwidth per hour.
Cost per unit of throughput, B200
FP8 throughput not published. $0.32 per TB/s of bandwidth per hour.
Price per unit of throughput is the number that decides between them for a job that saturates the card: the GPU with the lowest figure finishes the same work for less, whatever its hourly rate.
The same models on one GPU of each, and the node it takes when one is not enough.
Weights only, at 2 GB per billion parameters in FP16 and 1 GB in FP8, keeping 20% of the memory free for activations and the KV cache. "Fits" means one GPU; otherwise the smallest node size that does.
| Model | Precision | H100 80 GB | H200 141 GB | B200 180 GB |
|---|---|---|---|---|
| Llama 3.1 8B8B parameters | FP16 · 16 GB | Fits · $0.55/h | Fits · $0.75/h | Fits · $1.69/h |
| FP8 · 8 GB | Fits · $0.55/h | Fits · $0.75/h | Fits · $1.69/h | |
| Qwen2.5 32B32B parameters | FP16 · 64 GB | Fits · $0.55/h | Fits · $0.75/h | Fits · $1.69/h |
| FP8 · 32 GB | Fits · $0.55/h | Fits · $0.75/h | Fits · $1.69/h | |
| Llama 3.3 70B70B parameters | FP16 · 140 GB | 4× node · $2.20/h | 2× node · $1.50/h | Fits · $1.69/h |
| FP8 · 70 GB | 2× node · $1.10/h | Fits · $0.75/h | Fits · $1.69/h | |
| Mixtral 8x22B141B parameters | FP16 · 282 GB | 8× node · $4.40/h | 4× node · $3.00/h | 2× node · $3.38/h |
| FP8 · 141 GB | 4× node · $2.20/h | 2× node · $1.50/h | Fits · $1.69/h | |
| Llama 3.1 405B405B parameters | FP16 · 810 GB | more than one node | 8× node · $6.00/h | 8× node · $13.52/h |
| FP8 · 405 GB | 8× node · $4.40/h | 4× node · $3.00/h | 4× node · $6.76/h |
Spot prices, per hour for the whole node. Multi-GPU nodes are priced per GPU at the same rate, so the cost of a 70B model is the cost of the smallest node that holds it: the memory decides, not the hourly rate.
Three rules, from the numbers above.
- Serving a model that fits on an H100
- H100. The cheapest hour of the three, and for a model under about 32B parameters at FP16 the extra bandwidth of the others buys tokens you may not need. Spot at $0.55.
- Serving a 70B model, or long contexts
- H200. One card holds what takes two H100s, and the KV cache of a long context has room to grow. $0.75 an hour on spot against $1.10 for two H100s.
- Training or fine-tuning that saturates the card
- B200. The most throughput for 3.1× the spot price of an H100: the step finishes sooner than the meter runs. Below saturation the advantage shrinks to nothing.
- Anything with a deadline
- On-demand, whichever card. Same nodes, never reclaimed, 99.9% monthly uptime target. Move the disk from spot to on-demand in one click when the deadline appears.
Two at a time
Every pair has its own page, with the full table and what fits on each:
Real throughput depends on the serving engine, batch size, precision and how well the job keeps the card busy. The ratios above are the ceiling; measure your own job for an hour on each, which costs $2.99 on spot.
H100, H200 or B200?
Which is cheaper per hour, the H100, the H200 or the B200?
Here: H100 SXM $0.55 on spot and $0.89 on-demand, H200 SXM $0.75 and $1.09, B200 SXM $1.69 and $2.49, all per GPU-hour and billed per minute. Across the market index the medians are $2.50, $3.89 and $5.37.
Is the B200 worth 180% more than an H100?
For work that scales with tensor throughput, usually yes: the datasheet gives the B200 far more FP8 throughput and 2.3× the memory bandwidth, for 2.8× the on-demand price. If the job finishes proportionally faster, it costs less on the B200 despite the higher rate. For a job that is small enough to leave the card idle, the H100 is the cheaper hour.
When is the H200 the right choice?
When memory is the constraint and Blackwell is not needed: 141 GB against 80 GB on the same Hopper architecture, with 1.4× the bandwidth. A 70B model at FP16 needs about 140 GB for the weights: one H200 holds it with room for the KV cache, an H100 needs two. Inference throughput follows bandwidth, so the H200 also serves more tokens per second per card.
Can I move a job between them?
Yes. All three run the same CUDA stack and the same templates here, and a persistent disk moves between instances. The FP8 code path is the same on Hopper and Blackwell; only FP4 is Blackwell-only. A job written for one H100 runs unchanged on an H200 and gains memory; on a B200 it gains throughput.
How often are they reclaimed on spot?
Over the last 30 days: <5% of spot H100 instances, 5–10% of H200 and 5–10% of B200. The notice is 2 minutes on every model and the disk survives the stop. On-demand is never reclaimed.
More guides: the H100 price index · the cheapest providers for eight GPUs
An hour on each of the three costs $2.99 on spot.
Nothing renews, your disk moves between them, and the meter stops when you do.
Billed per minute from the moment the instance is reachable. Minimum credit $40, no subscription.