TPU up to 50% better per dollar: will an independent rerun confirm it? — Contested · resolves 2026-11-07

The debate: on September 7, SemiAnalysis's Dylan Patel said Google's TPU delivers up to 50% better performance per dollar, based on their InferenceX results for TPUv7 Ironwood against B200 and B300.

Why it matters: lower cost per unit of compute and less dependence on Nvidia are the main case for in-house chips, and if the gap survives an independent rerun, buyers of inference hardware gain an independent number to compare against.

Who is on each side: our 2026-09-13 column rated the headline contested because the first-hand reading comes from a single firm's single round of outreach, and it waits for a third party with no Google ties to rerun it on public software.

This card is a yes-or-no question, not a bet on either side: after the public release of TorchTPU (the original piece said mid-October), will a named third party with no Google partnership, and not SemiAnalysis, rerun the same workload on the public software by 2026-11-07 and reach a same-direction conclusion on cost per token? If so, the headline's generality gains independent evidence and our column's contested rating moves to supported; if not, or if the result points the other way, the fact that the first-hand reading comes from a single firm's single round of outreach remains the main weakness and the rating stays contested.

Sources still being linked up — the node(s) behind this call don't have a clickable original link resolved yet. Logged for follow-up.

The test setup must match: TPUv7 Ironwood against B200 and B300, Qwen3.5 397B, FP8, 8k tokens in and 1k out. Only the third party's own signed measurement with a parameter table counts; secondhand accounts do not. Such a measurement with a same-direction result before the due date answers yes; none, or an opposite result, answers no. This is a yes/no question, not a bet by us; the result is recorded on this card as yes or no and is not counted as our call being right or wrong.

What would overturn it

If TorchTPU is not open-sourced in mid-October as the original piece said, or the open-source version cannot be compared with the branch used for the measurement, this card is void: the measurement's premise fails, so it scores neither a hit nor a miss.