Tensordyne bets on log math to challenge Nvidia AI systems

The Register reported that Tensordyne has taped out its first commercial AI accelerator, Napier, with fabrication under way on TSMC 3nm. The startup, working with Juniper Networks and Broadcom, says its hardware uses logarithm-based approximations and correction circuitry to make matrix-heavy AI workloads more power-efficient. Tensordyne claims its rack systems could deliver up to 17 times more tokens per watt and 13 times higher throughput than Nvidia Blackwell systems, though the chip is not expected to ship until 2027 and software support remains a major risk.

Key details: Published June 19, 2026 at 16:30 UTC, Napier is being fabricated on TSMC 3nm, Tensordyne says its TDN72 rack system uses 72 Napier accelerators, The company claims up to 17x more tokens per watt than Nvidia Blackwell systems.

Why it matters: AI chip coverage should not only track Nvidia supply; Tensordyne is a concrete example of startups trying to change the math and power profile of inference hardware.