Tenstorrent Blackhole: The RISC-V AI Chip You Can Actually Buy

The Product Lineup

Tenstorrent's Blackhole product family spans from a single PCIe card to a full 32-chip rack server. Every product uses the same open RISC-V ISA and the same open-source software stack — you write code once and scale from desktop to data center.

Breaking News: TT-Deploy JP (July 2026)

At the late-June 2026 TT-Deploy JPevent in Tokyo, Tenstorrent unveiled major performance milestones for the Blackhole and Galaxy architectures, further cementing RISC-V's viability for agentic AI:

• Kimi K2.6: Tenstorrent vendor claim — up to 900 tok/s/user on Galaxy superclusters (~3× vs GPUs in TT marketing). Not AGICY-measured; not the 300 tok/s fleet ASSUMPTION; not Kimi K3.
• LTX 2.3 Fast (Video Generation): Generated 144 frames of 1080p video with audio and lip-sync in approximately 6 seconds (4× faster than GPUs).
• TT-Ascalon S: Unveiled a new RISC-V CPU IP tailored specifically for the rigorous branching and logic demands of emerging agentic AI workflows.

Amidst these technical reveals, CEO Jim Keller formally shot down rumors of an $8–10 billion acquisition by Qualcomm. Tenstorrent is not for sale; the company is focused on its independent path, international deployments, and scaling towards an eventual IPO.

The Open Source Stack

Tenstorrent's entire software ecosystem is open-source and Apache 2.0 licensed — no license fees, no vendor lock-in, full auditability:

• TT-Metalium — Low-level kernel programming framework that gives you direct hardware control. Write custom operators, optimize memory layouts, and profile every cycle.
• TT-NN — High-performance neural network library with pre-optimized operators for transformers, attention, MoE routing, and common inference patterns.
• TT-Forge — Model compilation layer that ingests PyTorch, ONNX, and JAX models and maps them automatically to Blackhole hardware. Drop in your existing models and run.

All repositories are available on github.com/tenstorrent. The community includes contributors from enterprises, universities, and sovereign AI programs worldwide.

From Desktop to Datacenter

Tenstorrent's product line is designed as a continuous scaling path. You start developing on a single card and scale to sovereign infrastructure without changing your code:

• Blackhole p150 — Single PCIe card for development and inference on models up to ~30B parameters. Fits any standard workstation.
• TT-QuietBox 2 — 4 Blackhole chips in an acoustically optimized desktop enclosure. Run 120B parameter models from your office, under 50 dB.
• Galaxy Server — 32 Blackhole chips in a 4U air-cooled chassis. Run 671B+ parameter models at 400+ tokens/sec per user. No liquid cooling required.
• AGICY Fleet (planned) — 1,801 Galaxy servers (57,632 chips) planned as EU-sovereign compute at Vasilikos (pre-COD) — not live installed capacity.

The same TT-Metalium code compiles across all tiers. Scale your workload by adding chips, not rewriting software.

Why Open Hardware Matters

The shift from proprietary to open AI hardware is not ideological — it is strategic. Four structural advantages drive the transition:

• No vendor lock-in — RISC-V is an open ISA maintained by RISC-V International. No single company controls the instruction set, and no license can be revoked.
• No export controls — Unlike NVIDIA's high-end GPUs, RISC-V AI accelerators face no US export restrictions. Deploy anywhere in the world without regulatory risk.
• EU sovereignty compliance — Open silicon combined with open software meets the EU AI Act's requirements for transparency, auditability, and data sovereignty at the hardware level.
• Full auditability — The entire stack from ISA specification to compiler to runtime is inspectable. No black-box firmware, no hidden telemetry, no undocumented microcode.