Tenstorrent Reaches $6.2B Valuation to Scale Open-Weight AI Chips

The Toronto-based AI chip startup has reached a $6.2 billion valuation, TechCrunch reported, capital the company will deploy to scale its open-weight accelerator portfolio.

That figure puts Tenstorrent among the most valuable private challengers to Nvidia's data-center dominance. Open-weight silicon — chips paired with publicly published model weights and an open instruction-set architecture — has become a structural theme as buyers look past CUDA lock-in. The company's marquee customers include research labs, hyperscalers, and sovereign-AI programs that want hardware they can run without paying the Nvidia tax.

Who is Tenstorrent?

Founded in 2016 by veteran chip architect Jim Keller, Tenstorrent builds data-center accelerators around the RISC-V instruction set. Keller, whose résumé includes stints at AMD, Apple, and Tesla, has steered the company toward an open architecture that competes directly with Nvidia's proprietary GPU stack.

Tenstorrent sells hardware for both training and inference, positioning itself as a non-Nvidia option for hyperscalers, sovereign-AI programs, and enterprises running models from Meta, Mistral, and DeepSeek. The company has shipped earlier-generation parts and continues to tape out new devices aimed at higher-throughput training workloads.

That positioning has matured alongside the open-weight model trend. Each round of releases — Meta's Llama family, Mistral's model drops, DeepSeek's recent surge — has widened the addressable market for non-Nvidia silicon. Open-weight models port more cleanly to RISC-V and other open architectures than to closed GPU stacks, and Tenstorrent has positioned itself as one of the better-funded beneficiaries.

What does $6.2B buy?

The funding arrives as Tenstorrent moves from early-customer deployments to broader production. Earlier-generation accelerators shipped in lower volumes; newer tape-outs target higher throughput for training workloads, with partner foundries handling the silicon.

Scaling means wafer allocations at foundry partners, more aggressive packaging capacity, and continued work on the software stack — compilers, kernels, runtime libraries that make non-Nvidia silicon usable for production AI. Software maturity has been the harder problem for every Nvidia challenger, and Tenstorrent is no exception.

Each funding round has lifted Tenstorrent's valuation mark. Reaching $6.2 billion signals that public and private investors now price the company on production capability rather than roadmap promise. The number also ranks Tenstorrent among the better-capitalized AI silicon startups globally, a notable shift for a company that began as a small Toronto engineering shop.

Where does Tenstorrent fit?

Tenstorrent operates in a market where Nvidia still controls the lion's share of data-center accelerator revenue, but a credible second tier has formed. Groq, Cerebras, and SambaNova compete for the slice of training and inference workloads that doesn't run on H100s or Blackwell chips. Tenstorrent's RISC-V base and open-weight positioning differentiate it within that group, especially for buyers who specifically want instruction-set independence.

Open-weight silicon also carries geopolitical weight. As export controls tighten around advanced AI chips, RISC-V-based accelerators offer a path for sovereign-AI programs in regions that cannot reliably procure Nvidia hardware. Tenstorrent has courted those buyers, particularly in markets where the U.S. export-control regime has throttled access to top-end GPUs. The pitch is simple: silicon customers can deploy without running into Washington's licensing wall.

What's next?

Production scaling takes time. Wafer allocations, packaging capacity, and software-stack hardening don't move at funding-round pace. But the $6.2 billion valuation gives Tenstorrent runway to compete for design wins that previously went to better-capitalized rivals — and to keep pace with a market where Nvidia's product cadence sets the tempo.

Tenstorrent's open-weight bet will be tested against Nvidia's next platform refresh. If RISC-V silicon can absorb a meaningful share of training and inference workloads over the next 18 months, the $6.2 billion mark looks cheap.