Tenstorrent Galaxy Blackhole: RISC-V AI at Scale in 2026

Galaxy Blackhole: General Availability

In April 2026, Tenstorrent formally transitioned the Galaxy Blackhole from limited preview to general availability. The system is now orderable at volume by enterprises, cloud providers, and sovereign infrastructure programs worldwide. Early production units have already been deployed in Japan as part of a strategic partnership to build sovereign AI capacity in the Asia-Pacific region.

Each Galaxy server packs 32 Blackhole RISC-V compute chips into a standard air-cooled 4U form factor — no liquid cooling, no custom racks, no proprietary networking. This density, combined with standard Ethernet interconnects, allows data center operators to deploy at scale using existing facilities and cooling infrastructure.

Volume production marks a critical inflection point: the Galaxy is no longer a prototype or demo — it is a commercially available product with a verified price point of ~$110,000 per server, compared to roughly $3M for an NVIDIA DGX H100 rack.

Jim Keller's Philosophy: ‘No New Laws’

Jim Keller — who previously led silicon design at AMD (Zen), Apple (A-series), Tesla (FSD chip), and Intel — founded Tenstorrent on a contrarian conviction: the fundamental laws of computing have not changed. Rather than chasing exotic architectures, Keller insists that disciplined engineering of known principles yields superior results.

His design philosophy rests on three classical constraints. Rent's Rule governs the relationship between logic blocks and I/O pins — the Galaxy architecture maximizes external bandwidth relative to compute cores. Amdahl's Law dictates that serial bottlenecks dominate parallel scaling — Blackhole minimizes synchronization overhead across its 32-chip mesh. And the universal principle of balancing I/O, memory, and compute ensures no single resource becomes a bottleneck.

The result is a chip that does not chase peak FLOPS but instead optimizes for sustained throughput per watt and per dollar — the metrics that actually matter for production AI inference at scale.

The Open Hardware Stack

Tenstorrent's software ecosystem is fully open-source, standing in direct contrast to NVIDIA's proprietary CUDA lock-in:

• TT-Metalium — Low-level kernel programming framework that exposes full hardware control, enabling custom operator development and optimization.
• TT-NN — High-performance neural network library with pre-optimized operators for transformer architectures, attention mechanisms, and MoE routing.
• TT-Forge — Model compilation and compatibility layer that ingests PyTorch, ONNX, and JAX models, automatically mapping them to Tenstorrent hardware.
• TT-QuietBox — Acoustically optimized development enclosure for running a full Galaxy server in an office environment (under 50 dB).

The entire stack is Apache 2.0 licensed, allowing enterprises to audit, modify, and redistribute the toolchain without licensing fees or vendor dependencies.

Why AGICY Built on Tenstorrent

AGICY's decision to anchor its European sovereign AI infrastructure on Tenstorrent Galaxy rather than NVIDIA DGX is driven by four structural advantages:

• 27:1 CapEx ratio — A Galaxy server at ~$110K delivers comparable inference throughput to an NVIDIA DGX rack at ~$3M, enabling 1,801 servers for the cost of ~67 DGX racks.
• Air-cooled operation — Zero liquid cooling required. The Galaxy runs in standard data center environments, eliminating the capital and operational overhead of liquid cooling infrastructure.
• Open RISC-V ISA — The instruction set architecture is an open standard, immune to export restrictions, licensing revocations, or single-vendor control.
• EU sovereignty compliance — Combined with AGICY's EU-jurisdiction deployment in Cyprus, the open hardware stack ensures data and compute sovereignty at both silicon and software levels.

Investment Case

Tenstorrent has confirmed active IPO planning, with CEO Jim Keller publicly positioning the company as the leading RISC-V AI infrastructure provider. The timing aligns with explosive demand for non-NVIDIA AI compute, driven by export control concerns and the global push for sovereign AI capacity.

AGICY's planned deployment of 1,801 Galaxy servers — totaling 57,632 Blackhole chips at an estimated €184M fleet hardware CapEx (HW only; Phase 1 programme CapEx €462.2M LOCK) — ranks among the largest planned commercial RISC-V AI fleets in Europe (design target; not a verified market ranking). This scale is intended to validate the Galaxy architecture for institutional-grade workloads.

For investors, the combination of Tenstorrent's IPO trajectory and AGICY's committed deployment volume creates a dual exposure: hardware platform appreciation and sovereign compute revenue generation.