Tenstorrent TT-QuietBox 2: $9,999 RISC-V Desktop AI Workstation Released

The landscape of AI development is shifting, with a new emphasis on local, accessible compute power. Tenstorrent, the AI computing company founded by veteran chip architect Jim Keller, today announced the TT-QuietBox™ 2 (Blackhole™), a desktop AI workstation designed to deliver teraflop-class inference performance with a fully open-source software stack. Starting at $9,999, the system aims to address the growing costs and limitations of cloud-based AI infrastructure, offering developers and businesses a path toward “silicon sovereignty” and complete control over their AI pipelines.

The Rise of Inference and the Demand for Ownership

While AI model training often grabs headlines, the demand for inference – the process of using trained models to make predictions or generate outputs – has quietly surpassed it. According to Tenstorrent, inference now accounts for over 55% of cloud AI infrastructure spending, totaling $37.5 billion and continuing to accelerate. This surge in demand is driving a need for more efficient and cost-effective inference solutions. Currently, developers face a difficult choice: rely on per-token cloud fees that can quickly escalate with usage, or invest in proprietary hardware with limited transparency and customization options.

The TT-QuietBox™ 2 is positioned as an alternative, built on the premise that developers should have complete visibility and control over their compute infrastructure, from the underlying silicon architecture to the compiler. This approach is particularly appealing for small to medium-sized businesses seeking on-premise AI deployment without the complexities and costs associated with traditional server racks.

Inside the Blackhole: RISC-V and a Dataflow Architecture

At the heart of the TT-QuietBox™ 2 lies Tenstorrent’s Blackhole ASIC, built on the open-source RISC-V instruction set architecture. RISC-V has been gaining traction as an alternative to proprietary architectures like ARM, offering greater flexibility and customization. As noted in an interview with EE Times, Jim Keller sees RISC-V as crucial to the future of AI and custom silicon development. The system features four Blackhole ASICs working in a unified mesh, boasting 2,654 TFLOPS of compute power at BlockFP8 precision, supported by 128 GB of GDDR6 memory and 256 GB of DDR5 system memory.

A key innovation is the integration of compute and high-density SRAM on a single die. This “dataflow” architecture minimizes data movement between the processor and memory, a common bottleneck in conventional systems. By keeping data closer to the processing units, the TT-QuietBox™ 2 aims to deliver sustained throughput and avoid the performance limitations imposed by DRAM bottlenecks. Notably, the system avoids the use of High-Bandwidth Memory (HBM), sidestepping current supply shortages and price increases impacting the broader AI hardware market.

Real-World Workloads and Open-Source Software

Tenstorrent emphasizes the TT-QuietBox™ 2’s ability to run real-world AI workloads out of the box. The system is pre-configured to handle large language models (LLMs) like GPT-OSS 120B, capable of running a 120-billion-parameter model entirely on-device. Performance benchmarks include 476.5 tokens per second with Llama 3.1 70B and the ability to deploy Qwen3-32B as a private coding agent. Beyond LLMs, the workstation supports creative applications like image generation with Flux and video synthesis with Wan 2.2, ensuring that intellectual property remains secure and off third-party servers.

For scientific research, the TT-QuietBox™ 2 demonstrates impressive performance with Boltz-2, a biomolecular ML model. The system can predict the structure of a 686-amino-acid protein in just 49 seconds – a task that takes a modern CPU 45 minutes – matching the performance of high-end workstation GPUs at a fraction of the cost. The parallel processing capabilities allow for four protein structure predictions simultaneously, further increasing throughput.

Crucially, the TT-QuietBox™ 2 is underpinned by a fully open-source software stack. Tenstorrent’s TT-Forge AI compiler supports models from popular frameworks like PyTorch, ONNX, TensorFlow, JAX and PaddlePaddle. This open approach extends to lower-level components like TT-Metalium, the AI SDK, and TT-LLK, the low-level kernel software, providing developers with complete visibility and control over every stage of the AI pipeline. This level of transparency is particularly valuable for sovereign AI deployments and regulated industries requiring strict data handling guarantees.

Developer Experience and Practical Considerations

Tenstorrent has focused on creating a user-friendly developer experience. The TT-QuietBox™ 2 ships with Ubuntu 24.04 and the complete open-source software stack pre-installed, enabling quick deployment. Engineering improvements have reduced idle power consumption and heat output by approximately 50% compared to previous generations. The liquid-cooled chassis is designed for quiet, sustained operation in a desktop environment, eliminating the need for dedicated server rooms or specialized electrical work.

The system plugs into a standard 120V wall outlet, simplifying deployment for businesses and individuals alike. Availability is slated for Q2 2026, with a starting price of $9,999. Interested customers can join the waitlist at www.tenstorrent.com/waitlist/tt-quietbox.

Looking Ahead: Open Source and the Future of AI Compute

The launch of the TT-QuietBox™ 2 represents a significant step toward democratizing access to high-performance AI compute. By embracing RISC-V and an open-source philosophy, Tenstorrent is challenging the dominance of established players like ARM and offering a compelling alternative for developers and businesses seeking greater control and customization. The company’s success will likely depend on its ability to foster a vibrant ecosystem around its open-source tools and attract a community of developers willing to contribute to the platform. As Financial Content reports, Tenstorrent’s move to license its IP has already secured over $150 million in contracts from major companies like LG Electronics, Hyundai Motor Group, and Samsung Electronics, signaling strong industry interest in its approach.

The system will be showcased at the Game Developers Conference (GDC) 2026, March 11-13, at booth #1354. The long-term impact of the TT-QuietBox™ 2 will depend on continued software development, community adoption, and the evolution of the RISC-V ecosystem. The company’s commitment to open source and developer empowerment positions it as a key player in shaping the future of AI compute.