The Blackhole hardware and the software hurdle
The Tenstorrent Blackhole SoC consists of 120 Tensix cores dedicated to machine learning tasks and four L2CPU blocks. Each L2CPU block contains four SiFive X280 cores. The cores use a high-speed Network on Chip. The Blackhole P100a is $999, while the P150a is $1399. For professional development, the TT-LoudBox is $12,000, and the liquid-cooled TT-QuietBox is $15,000. The TT-QuietBox includes four n300 cards, an AMD EPYC 8124P 16-core CPU, 512GB of RAM, and 4TB of NVME storage. I find the software stack to be a significant problem for developers trying to harness this compute. The "Linux on Blackhole" project uses minimal device tree patches to enable mainline kernel boots. This architecture uses PCIe BARs of 2MB and 4GB to allow software to access any location on the Network on Chip. Jim Keller, who designed architectures for AMD, Apple, and Tesla, leads the company in its mission to address software 2.0 demands. The company also released Wormhole processors, which use Tensix cores to provide native scalability across multiple chips. These Wormhole n150 and n300 accelerators are 3/4 size PCIe cards that operate at up to 160W and 300W respectively. The Wormhole processors can combine to act as one single processor in a workstation, which looks like a single massive network of Tensix cores to software.
LG and Samsung strategic investment
LG Electronics plans to integrate Tensix AI accelerator core IP and Ascalon CPU core IP into smart TVs and automotive chips. This licensing follows a strategic relationship between the two firms. Samsung Securities and AFW Partners led the $693 million Series D round in December 2024. Tenstorrent also closed a $1.29 billion Series E round in August 2026, which sets the post-money valuation at $5.76 billion. LG Technology Ventures and Hyundai Motor Group also contributed to the funding. Other investors in the Series D included XTX Markets, Mesh Ventures, Export Development Canada, Bezos Expeditions, and Samsung Venture Investment Corporation. I view the massive capital from these Korean giants as a sign that Tenstorrent’s programmable silicon can replace traditional proprietary models. The company’s scale reaches from individual developers to 32-chip systems. LG and Tenstorrent also plan to collaborate on future generations of RISC-V CPU, AI accelerator, and video codec IP.
| Funding Round | Date | Amount | Post-Money Valuation |
|---|---|---|---|
| Series E | August 2026 | $1.29 billion | $5.76 billion |
| Series D | December 2024 | $693 million | $2.61 billion |
| Series C | May 2021 | $213.2 million | $1 billion |
Economics of edge AI deployment
The cost of running AI models on edge devices changes when companies move away from expensive Nvidia hardware. An 8-GPU NVIDIA H100 server costs between $250,000 and $320,000, and an 8-GPU NVIDIA H200 server costs between $320,000 and $420,000. Tenstorrent aims to serve edge use cases and fringe needs that Nvidia ignores. The Blackhole architecture allows for flexible memory configurations because each L2CPU has access to dedicated and shared DRAM pools. I suspect that the ability to scale AI workloads natively without virtualization gives a cost advantage to companies deploying smart appliances. You already know that high-performance RISC-V cores are expensive, but the Blackhole P150a is $1399. A single NVIDIA H100 GPU costs between $25,000 and $30,000, while an H200 costs between $32,000 and $40,000. The Blackhole architecture uses a heterogeneous grid of cores linked by a high-speed Network on Chip; this setup allows for flexible memory configurations and native scaling from a single card to 32-chip systems. Will the current software immaturity prevent widespread adoption in consumer appliances?
