Marc Bolitho

CEO @ Tensordyne

[Tensordyne @TensordyneInc]

Nobody gets more fast tokens out of a megawatt than us. Nobody.

· Bit-accurate log math is locked in.
· Chip is taped out with Broadcom and in wafer fab on TSMC 3nm.
· System bring-up of compute tray with HPE Juniper scale-up fabric is ahead of schedule.

Today we're publishing our whitepaper on how we solved for what really matters in AI inference.

[Marc Bolitho @MarcBolithoAI]

The trade-off is over.

Right now, large models and inefficient hardware force AI builders and companies into a painful choice: do you want ultra-fast inference for your users, or one that's profitable to run? The industry's standard response is to pick one and try to make it work. And the consequences are already playing out: price hikes, lower quantization, companies restricting access to AI, and increasingly creative strategies to finance the endless cost of inefficient inference.

We believe AI needs to be both fast and low cost.

Today, I am incredibly proud to unveil Tensordyne Napier, the first inference system built from the math up through the scale-up networking to eliminate the trade-off between AI speed and cost.

We named the platform after John Napier, the legendary mathematician who invented logarithms 400 years ago. Why? Because by embedding logarithmic math directly onto our silicon, our team has shrunk the physical footprint and power consumption of AI compute to unprecedented levels.

As the only company to have cracked log math for AI compute, we are uniquely positioned to right-size every specification of the chip — delivering exceptional speed at remarkable efficiency:

» Running massive 2T GPT MoE models at 1,000 tokens/sec typically requires significant infrastructure investment, such as 9 racks of Nvidia Rubin + Groq, or 14 racks of AWS Trainium3 + Cerebras. Tensordyne Napier achieves this exact performance on a single rack drawing just 120 kW instead of 1.5 MW — comparable to the output of a standard wind turbine.

» When executing next-gen reasoning workloads like DeepSeek-R1, a single Tensordyne rack delivers 13x more tokens per second and 17x more tokens per megawatt than an Nvidia NVL72 GB300 at 210 tokens per second. That efficiency translates to up to $33M in additional annual revenue per rack compared to legacy alternatives.

» Despite packing 608 PFLOPS of FP8 dense compute into a single rack - more than any other system on the market today - it is 100% air-cooled. Our pod and rack integrate seamlessly into standard industry racks, so you can deploy instantly without expensive liquid-cooling retrofits or data center renovations.

Fast and low cost AI has arrived.

Get in touch if you'd like to secure allocation.