A startup called Tensordyne has reached a milestone in hardware development, completing the "tapeout" of a new AI chip — the step where a chip's design is finalized and sent off to be manufactured. What makes it unusual is the underlying math.
Instead of the standard floating-point arithmetic that powers virtually every GPU on the market, Tensordyne built its chip around a Logarithmic Number System, or LNS. The approach represents numbers as logarithms rather than in conventional form, which can dramatically simplify the kinds of multiply-heavy calculations that dominate AI workloads.
According to EE Times, the company claims its chip can consume an order of magnitude less power per token compared to GPU alternatives — meaning roughly ten times more efficient per unit of AI output. The publication did not independently verify those figures, and the chip has not yet shipped to customers.
The claim lands at a moment when the energy appetite of AI is drawing intense scrutiny. Data centers running large language models are straining power grids, and the cost of electricity has become one of the defining constraints on how quickly AI can scale. If a chip could genuinely deliver that level of efficiency improvement, it would be significant — not just for operating costs, but for the feasibility of running AI at the edge, in devices with tight battery budgets.
Tensordyne is one of a wave of startups betting that rethinking chip arithmetic from scratch is a more promising path than incremental improvements to GPU design. Whether the real-world numbers hold up once the silicon is in hand remains to be seen — but the tapeout means those answers are now a step closer.
If the power claims prove out, it could pressure the dominant GPU makers and open new markets for AI in power-constrained environments.