NVIDIA A100 Pricing
Eli Mernit
Renting an NVIDIA A100 80GB costs $1.19 to $5.07 per GPU-hour on demand across the providers below, and most single-card offers fall between $1.19 and $1.79. The getdeploying median, which also counts 40GB cards, is $1.77 across 40 providers. Beam's A100 80GB SXM4 is $1.36/hr with vCPU, RAM, and NVMe included.
Key takeaways
- RunPod Community Cloud's PCIe card is the cheapest on-demand option in the table at $1.19/hr.
- Beam's $1.36/hr is the lowest SXM4 price listed; Lambda and Spheron sell SXM only in 8-GPU nodes.
- Google Cloud's $5.07/hr VM bundles 12 vCPU, 170 GB of RAM, and a 275 GiB SSD.
A100 80GB cloud prices per hour, by provider
| Provider | SKU | On-demand $/GPU-hr | Notes |
|---|---|---|---|
| RunPod | A100 PCIe 80GB, Community Cloud | $1.19 | Third-party hosts; SXM $1.39; Secure Cloud $1.59 |
| Hyperstack | A100 80GB PCIe | $1.35 | Per-minute billing; SXM $1.60; spot $1.08 |
| Beam | A100 80GB SXM4 | $1.36 | On-demand machine; vCPU, RAM, and NVMe included |
| Spheron | A100 80GB PCIe | $1.43 | 20-minute minimum; spot $1.13; SXM4 only in 8-GPU nodes |
| Runcrate | A100 80GB SXM4 | $1.49 | Per-second billing; same listing on Sesterce |
| Verda | A100 SXM4 80GB | $1.79 | Spot $0.89 |
| Lambda | A100 SXM 80GB | $2.79 | 8-GPU instances only; plus tax |
| Google Cloud | a2-ultragpu-1g | $5.07 | 12 vCPU, 170 GB RAM, 275 GiB SSD; us-central1; spot $3.04 |
| Market median | — | $1.77 | Across 40 providers with an on-demand price (getdeploying); includes 40GB cards |
How these prices were collected
Prices are on-demand, per-GPU list prices in US dollars for single-GPU shapes where offered, with spot excluded from the price column. Lambda's row is the per-GPU share of an 8-GPU instance.
- RunPod's GPU pricing page for Community and Secure Cloud
- Hyperstack's GPU pricing page for on-demand and spot
- Spheron's A100 rental page for rates and billing terms
- Runcrate's pricing page for its cheapest A100 offer
- Verda's pricing page for on-demand and spot rates
- Lambda's pricing page for its 8-GPU A100 instances
- Google Cloud's accelerator-optimized VM pricing for A2 Ultra
- getdeploying's A100 price tracker for the median
A100 vs H100: when the older card wins
The H100's biggest lead is on FP8 work, and the A100 has no FP8 support or Transformer Engine. On Beam, an H100 PCIe machine costs $1.83/hr against $1.36/hr for the A100 80GB SXM4, about 35% more, so the H100 only wins per job if it runs more than 1.35 times faster. Compute-bound training often does. Low-batch LLM inference in BF16 often doesn't, because the H100 PCIe's memory bandwidth is close to the A100 SXM4's. Our H100 PCIe price comparison covers the other side.
Where to rent an A100 80GB for less
If a PCIe card is enough, RunPod Community Cloud ($1.19/hr, third-party hosts) and Hyperstack ($1.35/hr) both undercut Beam. Spot is cheaper if your job survives interruptions: Verda's SXM4 spot rate is $0.89/hr and Hyperstack's is $1.08/hr.
For an on-demand SXM4 card, Beam's $1.36/hr is the lowest in the table, and the machine price covers vCPU, RAM, and NVMe. Lambda's and Spheron's 8-GPU nodes suit multi-GPU training, but you pay for eight cards even when you need one.
What this means for your workload
The A100 80GB suits jobs that need memory but not FP8. At two bytes per parameter, 80 GB holds about 40B parameters of BF16 weights before activations and the KV cache, enough for LoRA and QLoRA fine-tuning of mid-size models and for batch inference. Put multi-GPU training on SXM4 nodes linked by NVLink. If your model fits in 48 GB, renting an L40S costs less per hour.
FAQ
How much does an A100 80GB cost per hour?
On-demand prices in the table run from $1.19 to $5.07 per GPU-hour. The getdeploying median across 40 providers is $1.77, and Beam's A100 80GB SXM4 machine is $1.36/hr.
What's the cheapest way to rent an A100 80GB?
Spot is cheapest, from $0.89/hr for Verda's SXM4, if your job can restart after an interruption. On demand, RunPod Community Cloud's PCIe card costs $1.19/hr, and Beam's $1.36/hr is the lowest SXM4 rate listed.
A100 vs H100: which should I rent?
Rent the H100 if your stack uses FP8 or runs more than 1.35 times faster on it, the break-even at Beam's prices. Otherwise the A100 80GB is the cheaper card per job.
