Tensordyne’s Napier Chip Uses 400-Year-Old Math to Smash AI Inference Costs

Picture this: you’re running an LLM and asking it to write a novel. Every word it generates costs electricity, cooling, and rack space. Now imagine that same novel costing 17 times less power. That’s the claim — and it comes from a startup that dug up some of the oldest math in human history to solve one of the newest problems in tech.

On June 15, 2026, Tensordyne announced Tensordyne Napier (TDN), a 3nm ASIC inference processor that the company claims delivers 17x more tokens per watt and 13x higher throughput than Nvidia’s GB300 NVL72 rack. The system was taped out and is in production at TSMC, with commercial shipping slated for the second half of 2027. Volume production is expected in mid-2027.

Let’s break down what’s actually new here — and what’s worth skepticism.

The Old Math in a New Chip

Napier’s secret ingredient isn’t some exotic quantum effect or a new neural architecture. It’s logarithmic mathematics, a concept discovered by 16th-century Scottish mathematician John Napier (yes, the namesake of the “log” function you probably hated in high school).

Here’s the trick: the logarithm of A times B equals the logarithm of A plus the logarithm of B. In practice, this means you can replace expensive, power-hungry multiplication operations with cheap, simple addition operations inside the chip’s logic circuits.

“We’ve turned multipliers into adders,” explained Tensordyne founder and VP of AI Gilles Backhus to IEEE Spectrum. “Adders are smaller and more energy-efficient logic circuits. So Napier can pack more compute into a smaller area and still save on power.”

The catch has always been conversion — getting data back and forth between the logarithmic domain and the floating-point format that neural networks actually use. Tensordyne says its engineers cracked this problem. If true, it’s a genuinely clever piece of silicon design.

Why Inference Is the Real Money Pit

Here’s the context most people miss: inference is now more expensive than training for most AI companies. The world’s largest Generative AI firms spend over 50% of their current revenues on inference infrastructure. Every time someone asks ChatGPT a question, every time an AI agent takes action, every time a recommendation engine fires — that’s inference, and it never stops.

Training is a one-time cost. Inference is a perpetual tax on every interaction.

Nvidia’s been trying to solve this with ever-bigger GPU clusters. But the industry is starting to realize that general-purpose GPUs might not be the optimal tool for inference at scale. Enter specialized ASICs — chips built for one job, doing it exceptionally well.

Tensordyne went further than just optimizing for inference. They re-architected the entire stack:

• TDN Math — logarithmic computation replacing multiplication
• TDN AIP — the processor, combining 256MB of SRAM with 144GB of HBM3E in a two-tier memory architecture
• TDN Link — a proprietary any-to-any scale-up interconnect with sub-microsecond latency between processors

The result is a 72-chip pod (the TDN72) that Tensordyne claims delivers 1,300 tokens per second at 120kW — versus 800 tokens per second at 1,500kW for a configuration of nine Nvidia Rubin + Groq LPX racks. The cost per million tokens? $11 for Napier, $150 for the Nvidia/Groq alternative.

The Skepticism Filter

Now, the important part: none of these benchmarks have been independently verified. Tensordyne is a startup making bold claims about hardware that doesn’t exist yet. The company has over a dozen Letters of Intent and more than $200 million in forecasted demand — impressive, but still pre-revenue.

A few things to watch for:

1. Real-world model compatibility. Log-math introduces approximations. Will it handle all model architectures, or only specific ones? Tensordyne says LLMs and transformers work well in log space, but the proof is in actual deployments.

2. The 2027 timeline. That’s a long way to go. Plenty of AI chip startups have made bold promises; execution is where the industry separates from the fantasy.

3. Ecosystem lock-in. Unlike Nvidia’s CUDA, Tensordyne doesn’t have a decades-old developer moat. Convincing teams to migrate inference workloads to a new silicon architecture takes more than good benchmarks.

That said, the partnership with Broadcom (providing fundamental IP, silicon, and packaging design) and HPE Juniper Networks (the interconnect) lends credibility. And Tensordyne board member Kevin Johnson — the former Starbucks CEO and Goldman Sachs board member — brings serious capital-market heft.

What This Means for the Industry

If Tensordyne delivers even half of what it promises, the implications are significant:

• Inference economics flip. $11 per million tokens versus $150 isn’t a marginal improvement — it’s a different business model. Companies that thought AI was too expensive to deploy at scale might suddenly find it isn’t.

• Air cooling wins. The Napier compute trays are air-cooled, not liquid-cooled. For data center operators who’ve been forced into expensive liquid-cooling infrastructure, that’s a real cost saver.

• Specialized silicon gets serious. The trend away from GPU-dominant inference is accelerating. Groq, Cerebras, and now Tensordyne are all proving that domain-specific chips can outperform general-purpose accelerators for specific workloads.

• Logarithmic math goes mainstream. If this works, expect to see more chips exploring alternative number formats. Nvidia’s own chief scientist Bill Dally presented on log-math research at HotChips in 2023 — the question was always whether someone would build a chip that actually used it.

Quick Quiz

1. What mathematical operation does Napier replace multiplication with to save power?

Addition, using logarithmic number representation where log(A×B) = log(A) + log(B).

2. Why is inference increasingly expensive for AI companies compared to training?

Training happens once; inference happens every time a model is used. With billions of daily interactions, inference becomes a perpetual, scaling cost that can exceed 50% of revenue.

3. What’s the main reason skepticism is warranted about Tensordyne’s claims?

The benchmarks are self-reported and the chip doesn’t exist yet — no independent verification exists until real systems ship in late 2027.

---

Sources

• Forbes — Tensordyne Revives Logarithmic Math In A Bid To Cut AI Power Use — Karl Freund, June 15, 2026
• IEEE Spectrum — Logarithmic Math Fuels Bold Tensordyne Inference Claim — IEEE, June 2026
• Tensordyne — Tensordyne Announces Breakthrough Inference System to End AI’s Speed vs. Cost Trade-Off — Press Release, June 15, 2026