Tenstorrent: open RISC-V and open-source software as the NVIDIA antidote

Tenstorrent designs and sells computers for AI built around its Tensix tensor cores and RISC-V CPUs, combined with a fully open-source software stack (TT-NN, TT-Lang, TT-Forge) so customers can run almost any model without a proprietary compiler. Beyond hardware, it licenses its AI and RISC-V intellectual property to customers who want to design and own their own chips, aiming at sovereign-AI buyers who cannot or will not depend on NVIDIA.

That RISC-V plus a fully open-source AI stack can beat NVIDIA's lock-in: win buyers who want to own their silicon, back it with $693M, and ship Galaxy in volume.

Founded in Toronto in 2016 and led since 2021 by microprocessor legend Jim Keller (AMD Zen, Apple, Tesla), Tenstorrent set out to chip away at NVIDIA's dominant share of AI accelerators. Its thesis was that AI demand would outgrow proprietary GPU supply, and that buyers — national clouds, enterprises, and developers — would want hardware and software they could own, not a closed ecosystem they could only rent.

On December 2, 2024 Tenstorrent closed over $693M in a Series D at a $2B pre-money valuation, led by Samsung Securities and AFW Partners and joined by Bezos Expeditions, Hyundai Motor Group, LG Electronics, Fidelity, Baillie Gifford, XTX Markets and Export Development Canada, among others. The company said the round was oversubscribed and that it had signed customer contracts worth nearly $150M; the capital went to open-source software, hiring, global design centers, and building AI systems and clouds.

Still scaling: by May 2026 Tenstorrent had launched Galaxy Blackhole in volume production, linking 36 servers into a single supercomputer, and cited deployments with Equinix, Cirrascale, BetterBrain, ai& in Tokyo and Virtu Financial — a long way from NVIDIA's grip, but with shipped hardware, partners and record inference demos as proof of the open approach.

The open stack is the moat: a 100% open-source software layer and RISC-V cores let sovereign and enterprise buyers own their AI infrastructure instead of renting CUDA.

Cost is designed in, not bolted on: no silicon interposers or Ethernet switches keeps hardware cheap enough to win price-sensitive buyers NVIDIA cannot serve.

Licensing turns competitors into customers: selling AI and RISC-V IP means chipmakers who would never buy boxes still pay for the design.

When the incumbent owns the ecosystem, compete on ownership: open the software, license the silicon, and let buyers afraid of lock-in pay to own their stack.

As of May 2026 Tenstorrent shipped Galaxy Blackhole systems in volume: a 6nm chip with GDDR6 memory and direct-attach Ethernet, scaling 32 chips into a Galaxy and 36 Galaxies into a supercluster, with a 90% pass rate running Hugging Face models on a fully open-source stack. Equinix, Cirrascale, BetterBrain, ai& in Tokyo and Virtu Financial were named deployments. The December 2024 Series D and reported ~$150M in contracts underpinned the claim that the RISC-V open-source bet was commercial, not theoretical.