Layer 2 · head to head
NVIDIA GB200 NVL72 vs Google TPU v7 (Ironwood)
Rack-scale coherent memory against a 9,216-chip inference pod. Different shapes of the same bet.
| NVIDIA GB200 NVL72 | Google TPU v7 (Ironwood) | |
|---|---|---|
| Can you get it | Rentable by the hour from multiple clouds; purchasable at scale. | Cannot be bought. Exists only inside one cloud, so choosing it is choosing that cloud. |
| Memory | 13.5 TB HBM3E across the rack | 192 GB HBM3E per chip |
| Bandwidth | 576 TB/s aggregate | 7.37 TB/s per chip |
| Compute | 1.44 exaFLOPS FP4 inference | 4,614 FP8 TFLOPS per chip |
| Power | ~120 kW nominal, 130-132 kW observed at full load | Not published |
| Interconnect | NVLink 5, 72 GPUs in one coherent domain | 9.6 Tb/s inter-chip (ICI) |
| Software | CUDA | JAX, XLA, PyTorch/XLA |
| Workload | both | inference |
NVIDIA GB200 NVL72
For: Frontier training and the largest inference deployments, where a single 72-GPU coherent memory domain is the point.
The catch: Liquid cooling is a mandatory architectural requirement, not a preference. At 120-132 kW a rack it renders most enterprise halls structurally and electrically unable to host it. This is a datacentre decision, not a hardware decision.
Economics: NVIDIA publishes two cents per million tokens and a 15x ROI claim for this configuration. Treat vendor TCO as a ceiling, not a forecast.
Google TPU v7 (Ironwood)
For: Inference at scale on Google Cloud. Ships in 256-chip and 9,216-chip configurations.
The catch: You cannot buy this. It exists only inside Google Cloud, so choosing it is choosing a cloud, permanently. It is also inference-optimised — do not benchmark it as a training part.
Economics: Vertically integrated: the price you see is a cloud price, not a chip price, and it is not comparable to a $/GPU-hour rental line.
Before either — can you power it?
~120 kW nominal, 130-132 kW observed at full load against Not published. In 2026 that comparison usually matters more than the FLOPS one: the US interconnection queue exceeds 2,600 GW with waits approaching five years, and roughly 80% of projects withdraw before energising.
Want this whole layer as one table?
The accelerators on these pages as a single CSV — memory, bandwidth, compute, power, interconnect, software stack, availability, who each part is for and the catch — ready to sort, filter or load into a notebook.
Get the silicon table — free for research and evaluation
Four short lines and the download starts at once (CC-BY-NC-4.0, built from these pages the moment you ask). What you type is used only to answer the request and is never sold. Prefer Hugging Face? Request it there.
Using it inside a product, a client report or anything you sell? That needs a commercial licence.