Tenstorrent (Wormhole, Blackhole)

Tenstorrent designs AI accelerators that combine RISC-V CPU cores with their Tensix tensor compute units. The current product line is Wormhole (shipping) and Blackhole (recently launched). Led by CEO Jim Keller, a well-known chip architect whose prior work included AMD's Zen and Apple's A4 / A5. Tenstorrent is the rare datapoint in 2026 trying to make both the ISA and the accelerator open at the AI-accelerator level. The tt-metal low-level kernel framework and tt-forge model compiler are open-sourced under Apache 2.0. Compare to NVIDIA's H100 line, where the ISA, the accelerator, and the entire CUDA software stack are proprietary; or AMD's MI300X, where the hardware is proprietary but ROCm is mostly open. Tenstorrent sits further along the open spectrum than either. Production-readiness: shipping for real workloads, but the software ecosystem is years behind CUDA. Customers willing to do their own kernel work (national labs, sovereign-compute programs, research groups) can use it; mainstream production AI inference is still on NVIDIA. The strategic question for Tenstorrent is whether the open-ISA + open-accelerator argument attracts enough early customers and contributors to close the software gap before NVIDIA's lock-in compounds further.