Tenstorrent TT-QuietBox & Galaxy — Liquid-Cooled AI Workstations, Servers and Superclusters

Specifications

Overview

Tenstorrent's system line goes from the desk to the datacenter. TT-QuietBox is a whisper-quiet, liquid-cooled workstation built to run models up to 120 B parameters locally at a starting point of $9,999. TT-QuietBox servers extend the same architecture into sovereign, scale-out production systems starting at $70,000. Above both sits Tenstorrent Galaxy - multi-server supercluster configurations for AI workloads spanning video, code and large language models.

The scaling claim matters for buyers: Tenstorrent positions Galaxy as delivering near-linear scaling that preserves quality and interactivity, rather than the degraded interactivity that batch-oriented multi-node inference often produces. Underneath, the systems are built on Blackhole Tensix processors in PCIe card form, with the p150a variant exposing four passive QSFP-DD 800G ports so cards can be linked directly to pool memory and improve performance.

For organisations that want an open, RISC-V-based alternative to proprietary accelerator ecosystems, the differentiator is the software stack: TT-Metalium and TT-NN are fully open source, giving developers full access to the hardware. Tenstorrent also licenses the underlying AI and RISC-V IP for organisations building their own silicon.

Key Benefits

Deskside capability. Run 120 B-parameter models from a quiet, liquid-cooled workstation starting at $9,999. Sovereign scale-out. TT-QuietBox servers from $70,000 for production AI with near-linear multi-node scaling. Open software stack. TT-Metalium and TT-NN are fully open source with access to the metal. Flexible deployment shapes. Deskside, rack-scale and multi-rack supercluster options on one architecture.

Applications

Local LLM development and inference, sovereign AI deployments, production scale-out inference, video and code generation workloads, RISC-V and open-source AI research, and private AI clusters.

Related Products