Skip to content

NVIDIA · 80 GB VRAM

Rent NVIDIA A100 GPUs, by the second

Rent A100 GPUs from a prepaid balance. No waitlist, no sales call, no contract: launch, work, stop, pay for the seconds you used.

Live stock is on the standard GPUs page.

Standard A100 · available

For work that isn't sensitive

Ordinary GPU machines, no enclave. Templates or your own Docker image, SSH, volumes, the API and the CLI. If a launch fails, the prepaid hour comes back.

Confidential A100 · not available yet

For data you can't expose

No A100 in Intel TDX confidential mode today. For sensitive data, H100, H200 and RTX PRO 6000 Blackwell run as confidential VMs with attestation you generate yourself.

See confidential VMs →

The A100 in short

  • ArchitectureAmpere
  • Memory80 GB HBM2e
  • Memory bandwidthup to 2,039 GB/s (SXM)
  • Performance19.5 TFLOPS FP32
  • Power300-400 W

Source: NVIDIA

What people run on an A100

  • Training and fine-tuning medium to large models
  • Large-batch inference and embedding generation
  • Multi-day research workloads on datacenter hardware

From zero to a running A100

  1. Pick a machine on the standard GPUs page, filtered by GPU, price and stock.
  2. Launch with a template (PyTorch, vLLM, SGLang…) or your own image, from the console, the API or the CLI.
  3. Check what you got: one read-only command compares the machine with its listing.
curl -fsSL https://voltagegpu.com/hwcheck.sh | bash -s -- --gpu "A100" --gpus 1

How the hardware check works → · Open source on GitHub → · A real mismatch, caught →

Real check on a rented standard A100, 17 October 2026: matches its listing →

Other GPUs you can rent

B300 · B200 · H200 · H100 · RTX 5090 · RTX 4090 · RTX 3090 · A6000 · L40S · RTX A4000 · RTX A5000

NVIDIA A100 facts: prices, billing and limits

Prices read from our live inventory on 17 November 2026 at 10:00 UTC, in US dollars.

Confidential VM (Intel TDX)

No A100 Confidential VM available at this reading.

Standard GPUs (no enclave)

No standard A100 in stock at this reading.

Billing

  • A Confidential VM price covers the whole VM, every GPU in it; a standard GPU price is per GPU-hour.
  • Prices are shown per hour and billed per second. The first hour is charged when the machine starts; after that the running time is prorated to the second, and unused prepaid time is refunded when you stop it.
  • Storage volumes on standard machines are billed separately, on the data stored.

Limits

  • Spot capacity can be reclaimed by the host, and it is not always cheaper than on demand.
  • We have run no attestation on B200 and B300 GPUs, so we claim none for them.
  • Bare-metal servers, standard GPUs and sandboxes have no enclave and are never attested: use them only for data that is not sensitive. The confidential container tier, which gave the tenant no attestation, was retired on 1 October 2026.

What a Confidential VM proof establishes

  • An Intel TDX trust domain, not in debug mode
  • A platform Intel rates up to date
  • Both proofs are bound to one manifest
  • NVIDIA GPU attestation, on the GPUs where we ran it
  • Fresh, when you drew the challenge

What it does not establish

  • Attacks by someone with physical access to the server
  • Which software runs in the VM
  • That the trust domain and the GPU are the same machine
  • That your SSH session ends inside the attested trust domain
  • The NVLink fabric of 8-GPU nodes

Each point, with its sources, on the proofs page →