A100 rental with crypto: an NVIDIA A100 80GB for $0.99 an hour
The cheapest way to get 80 GB of fast GPU memory by the hour, paid only in crypto. Whole A100 cards, one to eight per machine, with root SSH and a BF16-ready PyTorch image.
A100 rental on Petaflare costs $0.99 per GPU-hour for the 80 GB card, $0.94 on the 500-hour Project block and $0.89 on the 2,000-hour Fleet block. The smallest purchase is 30 GPU-hours for $29.70. You get whole A100 80GB cards, SXM or PCIe, 1 to 8 per machine, with root SSH. The A100 has no FP8, but for BF16 training and serving it gives more memory per dollar than any other card we rent.
A100 80GB specs
Ampere generation, 2020. Its third-generation Tensor Cores handle TF32, BF16, FP16, INT8 and INT4, but not FP8, which arrived with Hopper.
- Architecture
- Ampere
- Memory
- 80 GB HBM2e
- Memory bandwidth
- 2,039 GB/s (SXM) · 1,935 GB/s (PCIe)
- BF16 Tensor (dense)
- 312 TFLOPS
- FP8 Tensor (dense)
- no FP8 support
- Board power
- 400 W (SXM) · 300 W (PCIe)
- GPU link
- NVLink 600 GB/s (SXM)
- Form factor
- SXM or PCIe, shown before you connect
NVIDIA figures from the official A100 80GB page. Tensor figures are without sparsity.
A100 price per hour across clouds
Prices for one A100 80GB, read on each provider's pricing page on 19 November 2026.
| Provider | Price type | A100 version | Per GPU-hour |
|---|---|---|---|
| Petaflare | Fixed, on demand | 80GB, SXM or PCIe | $0.99 · $0.89 on Fleet |
| Thunder Compute | Fixed, on demand | 80GB | $1.09 |
| Massed Compute | Fixed, on demand | SXM4 80GB | $1.38 |
| RunPod Secure Cloud | Fixed, on demand | PCIe · SXM | $2.09 · $2.19 |
| RunPod Community Cloud | Third-party hosts | PCIe · SXM | $1.19 · $1.39 |
| Lambda | Fixed, on demand | 8 × SXM (HGX) only | $3.19 |
| Vast.ai | Marketplace, median of the day | SXM4 · PCIe | $0.76 · $0.51 |
| Silicon Data index | Composite of neo-clouds and marketplaces | A100 | $1.64 |
Sources: Thunder Compute, Massed Compute, RunPod, Lambda (plus sales tax), Vast.ai public price feed (18 November median, cited with attribution), Silicon Data (A100 reading on the same page). Read on 19 November 2026.
Reading those A100 prices
Our $0.99 is the lowest fixed on-demand A100 80GB price in the table, about 40% under the Silicon Data index. Vast.ai's marketplace median is lower, $0.76 for an SXM4 card, because those A100s belong to independent hosts who set their own price and terms, and storage and bandwidth are billed on top. Lambda only rents its 80 GB A100s in 8-card machines.
Thirty A100 hours cost $29.70 and 2,000 hours cost $1,780, or $0.89 an hour. Solana confirms in seconds and USDT on Tron in about a minute, and five other coins work too. There is no card and no ID check, because payment is crypto only, as the no-KYC GPU cloud page explains.
A100 or H100: what one dollar an hour buys
Divide each card's memory and its dense BF16 Tensor throughput by its hourly price. The A100 wins on memory per dollar; the H100 and L40S win on raw math per dollar.
| Card | Price per hour | GB of memory per $1/hour | BF16 TFLOPS per $1/hour | FP8 |
|---|---|---|---|---|
| A100 80GB | $0.99 | 81 | 315 | No |
| H100 80GB (SXM figures) | $2.39 | 33 | 414 | Yes |
| L40S 48GB | $0.87 | 55 | 416 | Yes |
| H200 141GB | $3.29 | 43 | 301 | Yes |
Dense BF16 Tensor figures from NVIDIA's datasheets (H100 SXM 989, A100 312, L40S 362, H200 989 TFLOPS), divided by our hourly prices of 19 November 2026.
Jobs that fit an A100 80GB
Count 2 bytes per parameter for BF16 weights, about 1 byte for INT8 and half a byte for 4-bit weights, then leave room for activations or the KV cache.
- LoRA fine-tuning of a 7B to 13B model in BF16: 14 to 26 GB of weights plus adapters and activations. Fits with room for a decent batch.
- QLoRA fine-tuning of a 70B model: about 40 to 48 GB. Fits on one card.
- Serving a 34B model in BF16: about 68 GB of weights, so short contexts only.
- Serving a 70B model: in 4-bit (AWQ or GPTQ) the weights take about 35 to 40 GB and fit well; in 8-bit they take about 70 GB and leave almost nothing for the cache.
- Training vision or diffusion models from scratch: most fit, and 80 GB lets you raise the batch size.
- Full fine-tuning of an 8B model with Adam: about 128 GB of training state. Use two A100s, or LoRA on one.
The fine-tuning memory guide works through these sums step by step.
PCIe or SXM A100
The two versions share the 80 GB of HBM2e and the same Tensor Core throughput. The SXM module has slightly more memory bandwidth (2,039 GB/s against 1,935 GB/s), a higher power limit (400 W against 300 W) and NVLink at 600 GB/s to the other cards on its board. On one card you will rarely notice the difference; on 4 or 8 cards, NVLink matters for tensor-parallel jobs. You do not choose the version here: the dashboard shows it, with the host's country, once the instance runs.
When to move up to an H100
Choose an H100 instead when your job is limited by compute, not memory: large-batch training, or serving with FP8, which the A100 cannot run natively. The H100 does roughly three times the A100's dense BF16 math for 2.4 times the price, so a job that keeps the card busy finishes cheaper on the H100. A job that mostly waits on memory, or that needs 80 GB for a model that runs in BF16 anyway, costs less on the A100. Every card is on the price list; on Google Cloud the A100 comes in A2 machines, compared on GCP GPU prices.
A100 rental questions
How much does an A100 cost per hour?
$0.99 per GPU-hour for an A100 80GB on Petaflare, $0.94 on the 500-hour block and $0.89 on the 2,000-hour block. Prices for other clouds on 19 November 2026 ranged from $0.51 on a marketplace to $3.19 on Lambda's 8-GPU machines.
What is the difference between the A100 PCIe and the A100 SXM?
Same 80 GB and same Tensor Cores. The SXM version has a little more memory bandwidth, a 400 W power limit instead of 300 W, and NVLink between the cards on its board.
How powerful is an A100?
312 TFLOPS of dense BF16 Tensor math, 156 TFLOPS of TF32, and 2 TB/s of memory bandwidth on 80 GB of HBM2e. It has no FP8.
What is better than an A100?
The H100 (about three times the BF16 throughput, plus FP8), the H200 (141 GB) and the B200 (180 GB, FP4 and FP8). For inference and image work, the L40S also beats it on throughput per dollar, with 48 GB.
Sources
- NVIDIA A100 Tensor Core GPU, specificationsread 19 November 2026
- Thunder Compute pricingread 19 November 2026
- Massed Compute pricingread 19 November 2026
- RunPod pricingread 19 November 2026
- Lambda GPU cloud pricingread 19 November 2026
- Vast.ai public GPU price feedread 19 November 2026