GPU rental prices, early October 2026: list prices barely moved, but where you rent changes the price up to 4.7×

Since October 7 we've collected on-demand GPU list prices three times a day: 42 snapshots so far, about 370 offers from 16 providers. Sources: the Lambda, DigitalOcean and Vultr APIs, and NVIDIA Brev's public catalog, which covers AWS, Google Cloud, Oracle Cloud, Nebius, Crusoe, Hyperstack, Massed Compute and others. The live table is on our GPU prices page. This is the first look at what twelve days of it show; the full October index follows on November 1.

List prices barely moved

Of 407 offers we tracked across the period, 14 changed price at all:

• One real on-demand cut: an 8×A100 offer from Excess Supply dropped from $3.60 to $1.62 per GPU-hour on October 10 (−55%).
• Back and forth: Google Cloud's 8×H100 alternated between $12.41 and $13.98 per GPU-hour, and its T4 offers moved by about 1%. This looks like the catalog reporting different regions on different days rather than price changes.
• Spot prices moved more: DigitalOcean's spot price for AMD MI355X rose from $2.97 to $5.50 per GPU-hour (+85%), MI350X from $2.46 to $3.50 (+42%), and NVIDIA B300 fell from $8.00 to $6.17 (−23%).

On-demand list prices behave like price lists, not markets: they change in steps, rarely. Whatever is moving in GPU pricing week to week (availability, spot and marketplace prices, negotiated discounts) doesn't show up in the list.

Where you rent matters far more

The same GPU model costs very different amounts depending on the provider. Taking each provider's cheapest on-demand offer per GPU-hour on October 19:

The big clouds sit at the top of each range. Most of the market clusters well below them: eight of eleven H100 providers list between $2.99 and $4.41.

The listing often doesn't say which GPU you get

Of 22 on-demand H100 offers, 13 don't say whether the card is SXM, PCIe or NVL. Eight say SXM, one says PCIe.

That gap matters more than most of the price differences above. In our H100 study, an H100 SXM served 2.25× as many tokens as an H100 NVL under the same latency target, and every NVL card we rented was set below NVIDIA's power range. The SXM was the cheapest per token although it cost the most per hour: $0.31 per million output tokens, against $0.43 on PCIe and $0.57 on NVL.

So a listing that is 25% cheaper per hour can be twice as expensive per token, and the listing alone won't tell you which case you're in.

What this means if you rent GPUs

• Shop across providers, not across weeks. List prices hardly move; waiting rarely pays. The spread between providers for the same card is 1.4–4.7×.
• Ask which variant before you compare. For H100s, ask SXM, PCIe or NVL, and the power limit. Then check it on the node: nvidia-smi -q -d POWER, or pip install tokenwatt-check (how to check a GPU's power limit).
• Compare cost per token, not price per hour. Divide the hourly price by what the node actually serves at your latency target. Our cost per token page has measured numbers for common cards.

Limits

These are list prices at collection time for on-demand capacity. They exclude storage, egress, reserved and committed-use discounts, and they don't tell you whether capacity is available: most offers in the catalog carry no availability flag. Brev's catalog reports one region per offer, which can change between snapshots. Twelve days is a short window. RunPod and peer-to-peer marketplaces where hosts set their own prices aren't included.

Prices are each provider's published list prices, collected through the sources named above. A provider that wants its listings corrected or removed can email hello@tokenwatt.io.

Data: latest snapshot (JSON) · GPU prices page.