Where is the cheapest A100 rental right now, and when is the 80GB version worth it?

The cheapest A100 rental is on Vast.ai's marketplace, and the memory size changes the price. In our read of Vast's offer API at 13:30 UTC on Nov 3, 2026, a 40GB A100 started at $0.40/hour on-demand (median $0.53 across 39 offers) and an 80GB A100 at $0.63/hour (median $1.00 across 25 offers). The cheapest 80GB card from a Vast-verified host was $1.00. Managed 80GB A100s listed at $1.35 on Hyperstack and $1.79 on Runpod Secure Cloud. Pay for 80GB when you run a 70B model, gpt-oss-120b or a 70B fine-tune; for 32B-class models and smaller fine-tunes, 40GB does the job for about half the price. Prices move every hour, so check today's live prices before you rent.

Where is the cheapest A100 rental right now?

The table below is one A100, priced per GPU-hour, read on Nov 3, 2026. The Vast.ai rows come from our own pull of Vast's offer API at 13:30 UTC, split by memory size; every other row is the provider's published on-demand price on its own pricing page, read the same day. "From" is the cheapest listing and "median" the middle one.

Two things to know about the cheapest number. First, it's a single host: at 13:30 UTC the $0.63 80GB offer was one machine in Germany, not marked verified by Vast, with a 99.4% reliability score. The next cheapest 80GB cards were around $0.85 to $0.87, and the cheapest verified one was $1.00. Second, Vast's A100 listings mix 40GB and 80GB cards under the same name ("A100 SXM4", "A100 PCIE"), so on Vast, filter by GPU memory before you compare prices. Our live cloud-GPU table groups Vast offers by name, so its A100 rows can show a 40GB price.

Is the 80GB A100 worth twice the price?

Only if your job doesn't fit in 40GB. On Vast, the 80GB median ($1.00) was almost exactly twice the 40GB median ($0.53) on Nov 3. NVIDIA lists the 80GB A100 with HBM2e memory at 1,935 GB/s (PCIe) or 2,039 GB/s (SXM), so the extra money buys room, not a different chip. The table uses our VRAM engine, the same one behind the VRAM calculator: weights at the stated quantization, plus the KV cache at 16-bit and about 0.8 GB of overhead.

The other jobs where 80GB pays off, from the makers' own pages:

• gpt-oss-120b: OpenAI's model card says its MXFP4 weights let "gpt-oss-120b run on a single 80GB GPU". It does not fit on 40GB. Check the live price for it on cost to run gpt-oss-120b.
• Fine-tuning a 70B model: Unsloth's requirements table (which it calls absolute minimums) puts QLoRA (4-bit) on a 70B model at 41 GB, just over a 40GB card, so you need the 80GB one. 16-bit LoRA on a 32B model is 76 GB, which fits 80GB with little room; on a 70B model it's 164 GB, beyond any single A100.
• Where 40GB is enough: the same table puts QLoRA at 26 GB for a 32B model and 30 GB for 40B, and 16-bit LoRA at 33 GB for a 14B model. All of those fit a 40GB A100 at about half the 80GB price.

For any other model, plug it into the VRAM calculator, or open cost to run Llama 3.3 70B, which matches the model to the cheapest live card that fits. More on the 70B case in the cheapest way to run a 70B model.

Managed or marketplace: what do you give up for the low price?

Vast.ai is a marketplace of independent hosts: each offer has its own reliability score, location, disk price and bandwidth price, and the cheapest one may be gone in an hour. Hyperstack, Runpod Secure Cloud, Lambda and CoreWeave each publish one fixed price for their own data centers. On Nov 3 the cheapest managed single 80GB A100 was Hyperstack at $1.35; Runpod's $1.79 includes 117 GB of RAM and 8 vCPUs (PCIe) or 125 GB and 16 vCPUs (SXM).

The big clouds aren't built for one-card rentals at this tier. Lambda's single-GPU A100 is the 40GB one at $1.99; its 80GB A100 appears only in the 8-GPU table, at $2.79 per GPU-hour, so about $22.32 an hour for the machine. CoreWeave lists an 8-GPU A100 80GB machine at $21.60 an hour, which works out to $2.70 per GPU. Those make sense for a multi-GPU training run, not for one model.

Runpod also shows lower Community Cloud prices for the A100 ($1.19 PCIe, $1.39 SXM) in its pricing page's data, but its docs say Runpod is "no longer accepting new hosts for Community Cloud" (existing Community Cloud resources remain available), so a Community A100 may not be there when you look.

Is on-demand or interruptible the better deal for an A100?

Vast's docs describe on-demand as "Fixed pricing, guaranteed resources (high priority)" and interruptible as "Lowest cost, may be paused (low priority)". For 80GB A100s, interruptible bids started at $0.13 an hour on Nov 3 (median $0.56). That's tempting, but an 80GB A100 is usually rented for a long job, a fine-tune or a big model you've spent time loading, and a pause costs you that time. Use interruptible for work that saves checkpoints often; use on-demand for everything else. Hyperstack also lists a spot A100 at $1.08 and reservations from $0.95, if you can commit ahead.

What does an A100 cost for a real job?

Vast's docs say GPU compute is "Charged per second while your instance is running", and Runpod advertises per-second rates too, so you pay for the time the machine is on. At the Nov 3 prices:

• A 4-hour session with a 70B model at 4-bit (needs 80GB): about $4.00 on the cheapest verified Vast host, $5.40 on Hyperstack or $7.16 on Runpod Secure Cloud.
• A 10-hour QLoRA fine-tune of a 70B model (needs 80GB): about $10 on a verified Vast host, $13.50 on Hyperstack or $17.90 on Runpod.
• The same 10 hours for a 32B QLoRA (fits 40GB): about $5.30 at the Vast 40GB median.

Storage and bandwidth come on top. Vast's docs say storage is "Billed continuously while your instance exists, regardless of running state" and add: "Delete instances completely to cease storage billing." Runpod lists volume disk at $0.10/GB a month while running and $0.20/GB a month idle.

Should you rent an A100 or an H100?

Both come with 80GB, so they fit the same models; the H100 is the newer, faster chip and costs more. On Nov 3 Runpod Secure Cloud listed the A100 80GB at $1.79 against $2.89 for an H100 PCIe and $3.99 for an H100 SXM, and Vast's H100 SXM offers started at $1.79 (median $2.27, our 13:20 UTC feed). Rent the A100 when memory is what you need and the job isn't time-critical; rent the H100 when finishing faster matters more than the hourly rate. If you need more than 24GB but less than 80GB, 48GB cards are cheaper still: on Vast an RTX A6000 started at $0.35 and an L40S at $0.43 on Nov 3. For training choices, see the best cloud GPU for fine-tuning.

The short version: rent the A100 on Vast.ai when price matters and you'll check the host, filter for 80GB only when the job needs it, and use Runpod or Hyperstack when you want one fixed price. Check your model with the VRAM calculator first, compare every card on our cloud-GPU compare page, and see Runpod vs Vast vs Lambda and the cloud-GPU pricing pillar for the bigger providers.