Explainer
TPU vs GPU: how do they compare?
A TPU is a machine-learning chip designed by Google; a GPU is a graphics processor from NVIDIA or AMD that is also used for AI. Here are their published specifications and listed rental prices side by side, as of Oct 29, 2026.
What is the difference between a TPU and a GPU?
A TPU (tensor processing unit) is Google’s own accelerator, and Google Cloud is where TPU time is rented. In the Census, 1 provider lists TPU prices, against 12 providers for NVIDIA and AMD GPUs. Google describes each TPU v6e chip as one TensorCore containing two matrix-multiply units, a vector unit and a scalar unit (Google Cloud TPU v6e documentation). GPU makers publish their own architecture pages, linked from each GPU hub.
Google prices TPUs per chip-hour; GPU providers price per GPU-hour. The Census keeps each unit as the provider states it. Prices were last checked Oct 29, 2026. A listing not re-checked for 7 days is greyed and left out of every lowest-price figure; after 14 days it is removed from the page.
How do TPU and GPU specs and prices compare?
| Chip | Memory | Memory bandwidth | Lowest on-demand price |
|---|---|---|---|
| TPU v5e | 16 GB | 800 GiBps | $1.20 per chip-hour · Google Cloud |
| TPU v5p | 95 GiB | 2765 GBps | $4.20 per chip-hour · Google Cloud |
| TPU v6e | 32 GB | 1638 GBps | $2.70 per chip-hour · Google Cloud |
| Ironwood | 192 GiB | 7380 GBps | $12.00 per chip-hour · Google Cloud |
| H100 | 80 GB | 3.35 TB/s | $2.50 per GPU-hour · Hyperstack |
| B200 | 180 GB HBM3E | 7.7 TB/s | $6.00 per GPU-hour · Hyperstack |
Memory, bandwidth and price are the figures shown on each chip’s page, with its source. A figure a vendor does not publish is marked as such, never estimated. Google prints TPU memory in GiB for some chips, so those figures are not converted.
Head-to-head pages
- TPU v6e vs H100: 32 GB against 80 GB; lowest listed on-demand $2.70 per chip-hour against $2.50 per GPU-hour.
- Ironwood vs B200: 192 GiB against 180 GB HBM3E; lowest listed on-demand $12.00 per chip-hour against $6.00 per GPU-hour.
Where can you rent a TPU?
Google Cloud lists public per-chip-hour prices for TPU v5e, v5p, v6e and Ironwood, in several regions and with one- and three-year commitment prices. Every row is on the TPU v6e, Ironwood, TPU v5p and TPU v5e pages. AWS’s own chips are covered on the Trainium2 and Inferentia2 pages.
TPU vs GPU: common questions
What is the difference between a TPU and a GPU?
A TPU is a chip Google designed for machine-learning work, and Google Cloud is the provider that lists TPU prices in the Census. GPUs come from NVIDIA and AMD and are rented from 12 providers in the Census. Google’s TPU documentation says each TPU v6e chip has one TensorCore with two matrix-multiply units, a vector unit and a scalar unit.
Where can I rent a TPU?
Google Cloud lists public per-chip-hour prices for TPU v5e, v5p, v6e and Ironwood. The Census tracks no other provider that sells TPU time by the hour.
Is a TPU cheaper than a GPU?
Listed prices are not like for like, so the Census does not rank them. The lowest listed on-demand price is $2.70 per chip-hour for the TPU v6e and $2.50 per GPU-hour for the H100, but the chips differ in memory, speed and the software they run. Use the comparison pages for the specifications.
Why is a TPU price per chip-hour and not per GPU-hour?
Google Cloud states Cloud TPU prices per chip-hour. The Census keeps that unit and shows it as chip-hour wherever a TPU or other non-GPU chip is priced.
Related: GPU rental prices · GPU comparisons · Rental cost calculator · Neoclouds
Cite this page
Last updated Oct 29, 2026. Figures may change as records are updated — include the date you retrieved them.
| APA | Data Center Census. (2026). TPU vs GPU. Retrieved from https://datacentercensus.com/gpu-prices/tpu-vs-gpu |
|---|---|
| MLA | “TPU vs GPU.” Data Center Census, Oct 29, 2026, https://datacentercensus.com/gpu-prices/tpu-vs-gpu. |
| Link | https://datacentercensus.com/gpu-prices/tpu-vs-gpu |
Republishing a figure or table? Attribute it to Data Center Census with a link to this page. The underlying data is downloadable as JSON and CSV; the press kit has standing figures and contact details.