Tenstorrent Launches Galaxy Blackhole AI System for General Availability

Tenstorrent announced the general availability of its Galaxy Blackhole AI compute platform on April 28, 2026. The system is designed to provide general-purpose AI performance by unifying compute, memory, and networking into a single architecture, which the company describes as Networked AI.

The Galaxy Blackhole system aims to address the inefficiencies of fragmented infrastructure, where separate accelerators are often connected via external networking. Tenstorrent’s approach integrates these components natively to optimize the system for real-world AI workloads, specifically targeting large-language-model (LLM) inference and high-quality video generation.

Technical Specifications and Architecture

The hardware is delivered in a 6U chassis that houses 32 Blackhole accelerators. These accelerators are interconnected using a dense Ethernet mesh, providing 100 Tbps of aggregate bandwidth.

The platform is built on RISC-V architecture, an open standard instruction set. By utilizing this architecture, Tenstorrent seeks to provide a scalable alternative to the proprietary ecosystems currently dominating the AI hardware market.

According to reporting from The Register, the 6U chassis is priced at $110,000.

Performance Claims and Use Cases

Tenstorrent claims the Galaxy Blackhole system is 10x faster than leading GPU systems. The company asserts that this performance gain is particularly evident in AI video generation and LLM operations, specifically during the prefill and decode phases of inference.

The company focuses on three primary pillars of efficiency to compete with existing AI infrastructure:

• Sustained inference performance.
• High-speed memory access.
• Scalable networking.

These technical priorities are intended to support real-time AI applications, such as the generation of high-quality video in less than a second.

Industry Context

Led by CEO Jim Keller, Tenstorrent is positioning the Galaxy Blackhole as a challenge to the traditional GPU-centric model of AI scaling. While GPUs have historically been the primary tool for training and deploying AI, the Galaxy platform’s use of a networked AI architecture is intended to reduce the overhead associated with moving data between separate compute and networking layers.

The move toward general availability marks a transition from the prototype and early-access phases the company underwent in 2025 to a commercially available product targeting data centers and enterprises requiring massive scale for AI inference.