Updated daily · last refresh 2026-10-01 · 123 rates · 24 clouds
GPU pricing
The current cost of AI compute, in one table: on-demand rental rates per GPU-hour across 24 clouds, the cheapest and typical price for every accelerator that matters, and used purchase prices where a real secondary market exists. Marketplace rates refresh automatically every day (last refresh 2026-10-01); provider list prices get a full manual review monthly (last: 2026-09).
Three things make GPU pricing confusing, and this page is built to cut through all of them. First, the same physical GPU spans a 2-4x price range depending on who rents it to you: marketplaces set the floor, neoclouds compete in the middle, and hyperscalers charge for their ecosystem (the cloud index profiles all of them). Second, list price is a ceiling: reserved and spot commitments routinely cut 30-70%. Third, renting is not the only option: with used datacenter GPUs widely available, ownership beats rental at surprisingly modest utilization, which the ROI calculator will show you in about ten seconds. Click any GPU below for its full pricing breakdown across every provider we track.
Datacenter GPU pricing
| GPU | Cheapest $/hr | Median $/hr | Used unit price |
|---|---|---|---|
| B200 | $3.75 | $5.88 | - |
| MI325X | $2.40 | $2.80 | - |
| H200 | $2.30 | $3.44 | - |
| Gaudi 3 | $2.30 | $2.30 | - |
| H100 SXM | $1.90 | $2.49 | ~$23,000 |
| MI300X | $1.85 | $2.83 | - |
| H100 PCIe | $1.60 | $2.56 | ~$19,500 |
| GH200 | $1.35 | $1.42 | - |
| A100 80GB | $0.87 | $1.35 | ~$6,800 |
| L40S | $0.80 | $1.09 | ~$6,200 |
| A10 | $0.75 | $1.06 | ~$1,100 |
| RTX PRO 6000 | $0.50 | $0.66 | ~$8,500 |
| A100 40GB | $0.48 | $0.94 | ~$3,400 |
| L4 | $0.32 | $0.73 | - |
| V100 | $0.21 | $0.58 | ~$450 |
| T4 | $0.10 | $0.44 | ~$320 |
Consumer GPU pricing (marketplace rentals)
AI ASIC pricing
On-demand list prices per GPU-hour, normalized per GPU for multi-GPU instances; used prices are indicative for working pulls. Methodology.
How to use these numbers
Start with what fits
GPU pricing only matters for cards that can hold your model. The VRAM calculator narrows the table to the ones that can.
Convert $/hr into $/million tokens
Hourly price is the input; cost per token is the answer. The estimator does the conversion for any model and GPU.
Check the buy-vs-rent line
Used prices above plus your utilization decide whether renting makes sense at all.
See where prices are heading
GPU price trends: H100, A100, H200 and 4090 rates charted from launch to today, extended daily by this index.
See the value ranking
Bandwidth per dollar, the metric that predicts serving cost, ranked across this whole table.
Common questions
How much does it cost to rent a GPU?
As of 2026-09, on-demand GPU pricing runs from about $0.10/hr for a T4 to $8.65/hr for a B200 at hyperscaler list rates. The most-used tiers: A100 80GB from $0.85/hr, H100 from $1.65/hr, H200 from $2.10/hr. The same GPU commonly spans 2-4x between the cheapest marketplace and the most expensive hyperscaler.
Why does the same GPU have such different prices?
Provider class drives it: marketplaces (peer-to-peer hosts) set the floor, GPU-first neoclouds sit in the middle, and hyperscalers charge a premium for their ecosystem and enterprise features. Reserved and spot commitments then cut 30-70% off any list price.
Is it cheaper to buy or rent a GPU?
At sustained utilization above roughly 25-50%, buying wins: a used A100 80GB at about $6,800 works out near $0.35-0.75 per GPU-hour over three years including power, below most rental rates. At low or bursty utilization, renting wins because idle owned hardware still costs money.