A little-known US artificial intelligence startup called Tensordyne is making a bold claim: its new chip, dubbed Napier, outperforms Nvidia's flagship Blackwell GPU by 13 times in tokens per second — the standard measure of how fast an AI chip can generate text output.
According to TweakTown, which reported on the announcement, the Napier chip is built on a 3-nanometer process node, placing it on par with the most advanced silicon manufacturing available today.
Tokens per second is a critical benchmark for large language model inference — the phase where a trained AI model actually responds to a user's prompt. A 13x advantage, if verified, would represent a generational leap rather than an incremental improvement.
The caveat here is significant: this is a startup's self-reported claim, and independent benchmarks have not been cited. The history of chip startups includes many ambitious announcements that faced scrutiny once put to real-world tests. Nvidia's Blackwell architecture currently powers the most advanced AI data centers on the planet and has been validated by major cloud providers.
Still, the claim is notable enough to watch. If even partially borne out, a chip offering dramatically higher tokens-per-second throughput at competitive cost could meaningfully disrupt the AI infrastructure market — and challenge Nvidia's near-total dominance in the space.