GPU comparison
TPU v6e vs H100: specs and rental prices
Google Cloud TPU v6e (Trillium) against NVIDIA H100: 32 GB against 80 GB of memory, lowest listed on-demand price $2.70 against $2.50 per GPU-hour or chip-hour. Specs from Google and NVIDIA, prices as of Oct 29, 2026.
In short
- The H100 lists 80 GB of GPU memory against 32 GB for the TPU v6e: 2.5× higher.
- Memory type is HBM on the TPU v6e and HBM3 on the H100.
- The H100 lists 3,350 GB/s of memory bandwidth against 1,638 GB/s for the TPU v6e: 2.05× higher.
- Not every figure can be compared: at least one vendor page does not publish: FP8 tensor throughput; TF32 tensor throughput; FP64 throughput; Maximum power.
- The lowest listed on-demand price is $2.50 per GPU-hour for the H100 (Hyperstack) and $2.70 per chip-hour for the TPU v6e (Google Cloud).
- Per GB of GPU memory that is $0.084 per hour for the TPU v6e and $0.031 for the H100.
This summary is generated from the tables below. 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.
Specifications side by side
| Specification | TPU v6e (Trillium) | H100 SXM | Difference |
|---|---|---|---|
| GPU memory | 32 GB | 80 GB | H100 2.5× higher |
| Memory type | HBM | HBM3 | Different |
| Memory bandwidth | 1,638 GB/s | 3,350 GB/s | H100 2.05× higher |
| FP8 tensor throughput | Not published | 3,958 TFLOPS (with sparsity) | — |
| FP16 / BF16 tensor throughput | 918 TFLOPS (basis not stated) | 1,979 TFLOPS (with sparsity) | Not compared: the vendors state the figures on different bases |
| TF32 tensor throughput | Not published | 989 TFLOPS (with sparsity) | — |
| FP64 throughput | Not published | 34 TFLOPS | — |
| Maximum power | Not published | 700 W | — |
| GPU-to-GPU interconnect | 800 GB/s | 900 GB/s | H100 1.13× higher |
Compared versions: TPU v6e (Trillium) and H100 SXM. The figure in the stronger cell is shaded; power is shown without a winner. Two throughput figures are compared only when both vendors state them on the same basis (both with sparsity, or both dense), and “derived” means a total the vendor publishes divided by the GPU count it publishes. Full tables with sources: TPU v6e specifications · H100 specifications.
Rental prices
| Rental price | TPU v6e | H100 | Difference |
|---|---|---|---|
| Lowest listed on-demand price | $2.70 per chip-hour · Google Cloud · checked Oct 29, 2026 | $2.50 per GPU-hour · Hyperstack · checked Oct 29, 2026 | TPU v6e 1.08× higher |
| Lowest listed spot price | No fresh listing | $0.79 per GPU-hour · Nebius · checked Oct 29, 2026 | — |
| Price per GB of memory | $0.084 per GB of GPU memory per hour | $0.031 per GB of GPU memory per hour | TPU v6e 2.7× higher |
| Fresh listings | 6 from 1 providers | 26 from 12 providers | — |
Prices cover every version and instance size the providers list, so they are not tied to the versions compared above; the price per GB of memory divides the lowest fresh on-demand price by the memory of the compared version. Listed prices are not like for like across providers. See every listing: TPU v6e prices · H100 prices.
More on these GPUs
- Google Cloud TPU v6e (Trillium): specs, VRAM and rental prices
- NVIDIA H100: specs, VRAM and rental prices
- A100 vs H100
- H100 vs H200
- H100 vs B200
- L40S vs H100
- H100 vs MI300X
- RTX PRO 6000 vs H100
- Trainium2 vs H100
- TPU v5e vs TPU v6e
- TPU v5p vs TPU v6e
- Ironwood vs TPU v6e
- All GPU comparisons · Rental cost calculator
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 v6e vs H100. Retrieved from https://datacentercensus.com/gpu-prices/compare/tpu-v6e-vs-h100 |
|---|---|
| MLA | “TPU v6e vs H100.” Data Center Census, Oct 29, 2026, https://datacentercensus.com/gpu-prices/compare/tpu-v6e-vs-h100. |
| Link | https://datacentercensus.com/gpu-prices/compare/tpu-v6e-vs-h100 |
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.