Rent the H100 if…
you train or fine-tune transformers at scale, serve models in FP8 with vLLM or TensorRT-LLM, or need 8-GPU NVLink nodes for jobs that last weeks — time saved is money saved.
Deploy an H100 · $1,249/mo
NVIDIA H100 SXM80 GB HBM3 · Hopper
vs
NVIDIA A10080 GB HBM2e · Ampere
NVIDIA H100 vs A100: which GPU should you rent?
Same 80 GB of memory, two generations apart. The H100 SXM brings FP8, 1.6× the memory bandwidth and about 3× the dense tensor throughput; the A100 80 GB costs about half as much per month.
- H100, dedicated$1,249per month · $1.71/h effective
- A100, dedicated$649per month · $0.89/h effective
- Price ratio1.9×the H100 costs 1.9× the A100 per month
The short answer
Which one should you rent?
For training, and for any model that can use FP8, rent the H100: its tensor cores do about three times the dense FP16 work of the A100 and add an FP8 path the A100 does not have, so a compute-bound job finishes well before the price difference catches up. The A100 remains the better buy when the job is bound by memory capacity rather than compute — LoRA and QLoRA fine-tunes, 16-bit inference of models up to about 35B parameters, computer vision, speech — because it offers the same 80 GB for $649 a month instead of $1,249.
- Tensor throughput. About 989 TFLOPS of dense FP16/BF16 on the H100 SXM against 312 on the A100 — 3.2× on paper — plus about 1,979 TFLOPS of FP8, a format Ampere does not support.
- Memory. Both carry 80 GB, but the H100's HBM3 moves 3.35 TB/s against 2 TB/s of HBM2e on the A100: 1.6× the bandwidth for decode-heavy inference.
- Interconnect. NVLink 4 at 900 GB/s per GPU on the H100, NVLink 3 at 600 GB/s on the A100 — it matters for tensor parallelism across an 8-GPU node.
- Price. The H100 costs 1.9× the A100 per month. If your job runs more than about twice as fast on it, it is the cheaper GPU per result.
Rent the A100 if…
your workload runs in BF16, FP16 or TF32 and is limited by the 80 GB rather than by compute — QLoRA, LoRA, vision, speech, embeddings — or you want MIG to split one GPU into isolated instances.
Deploy an A100 · $649/moSpecifications
H100 vs A100: the numbers side by side
NVIDIA's published figures for each part, tensor throughput without sparsity. The last column is how the H100 compares, as a multiple of the A100.
| Specification | NVIDIA H100 SXM | NVIDIA A100 | H100 vs A100 |
|---|---|---|---|
| Architecture | Hopper · 2022 | Ampere · 2020 | — |
| GPU memory | 80 GB HBM3 | 80 GB HBM2e | Same |
| Memory bandwidth | 3.35 TB/s | 2 TB/s | 1.6× |
| FP16 / BF16 tensor (dense) | ~989 TFLOPS | 312 TFLOPS | 3.2× |
| FP8 tensor (dense) | ~1,979 TFLOPS | Not supported | Only the H100 |
| FP4 tensor (dense) | Not supported | Not supported | — |
| Maximum power | Up to 700 W | 400 W | 1.8× |
| Form factor | SXM5 | SXM4 | — |
| GPU-to-GPU link | NVLink 4 · 900 GB/s | NVLink 3 · 600 GB/s | — |
| Multi-Instance GPU (MIG) | Up to 7 instances | Up to 7 instances | — |
| ECC memory | Yes | Yes | — |
| Video engines | 7 NVDEC · 7 NVJPG | 5 NVDEC · no NVENC | — |
| CUDA compute capability | 9.0 | 8.0 | — |
| Dedicated host here | 24 vCPU · 256 GB RAM · 3 TB NVMe | 16 vCPU · 128 GB RAM · 2 TB NVMe | — |
| Price per month here | $1,249 | $649 | 1.9× |
Sources: NVIDIA datasheets and product pages for each GPU. Tensor figures are dense (NVIDIA often prints them doubled “with sparsity”). Prices are this site's flat monthly rates for a dedicated server with the host listed, in USD excluding VAT.
Value
What each dollar buys
Price per month divided into the two resources most jobs run out of first — memory and memory bandwidth — and what the same GPU costs kept on for a month at the market-median hourly rate.
Per GB of GPU memoryH100 $15.61 · A100 $8.11per month, dedicated host included
Per TB/s of bandwidthH100 $373 · A100 $318per month — what decode-heavy inference pays for
Effective hourly rateH100 $1.71/h · A100 $0.89/hflat monthly price over 730 hours
Same GPU on demand, 24/7H100 $2,446 · A100 $1,285market-median hourly rate × 730 h
Model sizes
What fits on one H100 and one A100
The largest model whose weights fit on a single GPU, keeping 10% of the memory for the runtime, activations and a small KV cache. Long contexts and large batches need more — size them with the GPU memory calculator.
| Weights precision | NVIDIA H100 SXM | NVIDIA A100 | Typical use |
|---|---|---|---|
| 16-bit (BF16 / FP16) | up to ~36B parameters | up to ~36B parameters | Full-precision inference; the base weights of a LoRA fine-tune |
| 8-bit (FP8 / INT8) | up to ~72B parameters | up to ~72B parameters | FP8 serving with vLLM or TensorRT-LLM on GPUs that support it; INT8 weights elsewhere |
| 4-bit (GPTQ, AWQ, Q4) | up to ~130B parameters | up to ~130B parameters | Quantised inference on a budget; QLoRA fine-tuning |
Parameters in billions, weights only: 2 bytes per parameter at 16-bit, 1 at FP8 or INT8, about 0.55 at 4-bit (GPTQ, AWQ, GGUF Q4). Mixture-of-experts models need memory for all their experts, not just the active ones.
Prices
H100 and A100 price per month, by term
Per dedicated GPU server, in USD excluding VAT — CPU cores, RAM, NVMe, 20 TB of transfer and 24/7 support included.
| Term | NVIDIA H100 SXM | NVIDIA A100 | Difference per month |
|---|---|---|---|
| Monthly | $1,249 | $649 | $600 |
| 3 months · −5% | $1,187 | $617 | $570 |
| 6 months · −10% | $1,124 | $584 | $540 |
| 12 months · −15% | $1,062 | $552 | $510 |
| Dedicated host | 24 vCPU · 256 GB RAM · 3 TB NVMe · 25 Gbps | 16 vCPU · 128 GB RAM · 2 TB NVMe · 25 Gbps | included |
H100 vs A100: questions answered
Is the H100 three times faster than the A100?
On paper it has 3.2× the dense FP16 tensor throughput and adds FP8, so compute-bound transformer training sees the largest gains. Memory-bound work gains closer to the 1.6× bandwidth difference, and jobs limited by data loading or communication gain less. Benchmark your own model: the monthly price difference here is 1.9×.
Do the H100 and the A100 have the same memory?
Both have 80 GB, so the same models fit. The difference is speed: 3.35 TB/s of HBM3 on the H100 SXM against 2 TB/s of HBM2e on the A100 80 GB.
Which is better for LLM inference?
For FP8 serving and high-throughput batches, the H100: FP8 halves the memory taken by the weights and its bandwidth raises tokens per second. For 16-bit inference of a model that fits comfortably, the A100 serves it for about half the monthly price.
Are the H100 and the A100 in stock?
14 of 64 H100 units were free at the last stock update; 21 of 48 A100 units were free at the last stock update (18 September 2026). Both are provisioned automatically, in under 10 minutes after payment.
Can I rent the H100 or the A100 without KYC?
Yes. Both are rented with an email address and paid from a prepaid balance in Bitcoin, Ether or USDT — $1,249 and $649 a month, dedicated, with no identity check and no card.
Other GPU comparisons
Both are in stock, by the month
Dedicated, single-tenant and live in under 10 minutes: the H100 at $1,249 or the A100 at $649 a month, paid in BTC, ETH or USDT with no KYC.