Osegi
Hardware·2 min read

Tensordyne targets Nvidia in AI inference, not in training

German-American startup Tensordyne wants to beat Nvidia on speed and cost when running AI models, while leaving the training market alone.

Close-up of an AI accelerator chip on a circuit board in a data center, symbolizing competition with Nvidia.
Image: t3n Magazin
SUMODialoediviewergiftdetect

Tensordyne, a German-American chip startup, wants to challenge Nvidia where AI models are run rather than where they are built. The company aims to execute models faster and more cheaply than Nvidia's hardware, and it hopes to be valued at several billion US dollars soon.

Inference instead of training

Co-founder Gilles Backhus makes clear that Tensordyne is not going after Nvidia's strength in training large models. His stance on that: he wishes luck to anyone who tries. The startup concentrates on model execution, often called inference, where the pitch is better speed and lower cost.

According to Backhus, a switch in the calculation method is supposedly enough to get there. The claim is bold, since it suggests that a different approach to the underlying arithmetic could deliver an advantage over the market leader without competing head-on across the whole AI workload.

A market Nvidia still dominates

The challenge is considerable. Even with rivals such as Cerebras, Sambanova and Etched pushing ahead, Nvidia remains the benchmark for AI chips in 2026. The financials show it: in the months of May, June and July, Nvidia generated almost 100 billion dollars in revenue, of which about 60 billion dollars remained as a net figure.

AMD, the main competitor outside the large hyperscalers, for which chips make up a larger or smaller part of overall business, reported roughly 13 billion dollars in revenue for the third quarter of 2026 and a net profit of 2.4 billion dollars.

Why it matters

For companies that deploy AI, running models is an ongoing cost, so any credible alternative that is faster and cheaper could ease pressure on budgets and on dependence on a single supplier. Whether Tensordyne can prove its claims against Nvidia's scale remains to be seen, and its multi-billion-dollar valuation goal depends on exactly that.

Apps worth a look

SUMO: the number puzzle

Tap the right tiles, add them up, hit the target. Easy to learn, hard to master. Free, for all ages.

Tags:tensordynenvidiaai chipsinferencestartupamd

More from this day

See all →

More in: Hardware

See all →