Qualcomm discusses an $8 billion to $10 billion acquisition of Tenstorrent. Reports from June 2026 indicate these talks. Jim Keller leads the design team. He previously built the AMD Zen architecture and Tesla’s autonomous driving chip. Tenstorrent uses RISC-V to bypass the licensing fees that Arm and x86 require. This architecture allows designers to modify processors without single-vendor control. The company raised $693 million in December 2024. This Series D round, led by Samsung Securities and AFW Partners, valued the firm at $2.7 billion. Total funding exceeds $1.18 billion. Keller’s history at Apple with the A4 and A5 chips adds to the company’s weight in semiconductor circles. The company employs approximately 1,200 people across offices in Toronto, Austin, Santa Clara, Belgrade, Seoul, Tokyo, and Bangalore.
The Tensix architecture and hardware constraints
The Tensix core defines the Tenstorrent hardware stack. Each core contains five baby RISC-V processors and 1.5 megabytes of local SRAM. It lacks a cache hierarchy and uses software-managed memory instead. This design relies on explicit data movement via DMA operations rather than a hardware multithreading model to hide latency. The firmware update v19.5.0, released in February 2026, reduced the p150 card Tensix core count from 140 to 120 and its SRAM from 210 MB to 180 MB without providing a specific reason. This reduction in hardware resources directly impacts the peak performance of the p150 card.
| Product | Type | Compute (FP8) | Memory |
|---|---|---|---|
| Wormhole n300 | PCIe Card | 466 TFLOPS | 24 GB GDDR6 |
| Blackhole p150a | PCIe Card | 745 TFLOPS | 32 GB GDDR6 |
| Galaxy Blackhole | Rack System | 23 PFLOPS | 1 TB DRAM |
The Wormhole n300 pairs two processors to deliver 128 Tensix cores with 192 megabytes of SRAM and 24 gigabytes of GDDR6 memory. It provides 466 FP8 TFLOPS at 300 watts. The Blackhole p150a costs $1,399 and includes 32 gigabytes of GDDR6 memory. The Galaxy Blackhole delivers 23 petaflops of block FP8 compute and 56 ports of 800-gigabit Ethernet. It uses 1 terabyte of DRAM and 6.2 gigabytes of on-chip SRAM. The company sells these systems to customers wanting to scale out using standard Ethernet. The Tensix architecture follows a dataflow execution model where each core operates on data as it becomes available. The five baby RISC-V cores handle distinct pipeline stages: reading data from DRAM, unpacking data from SRAM buffers, managing matrix computations, managing vector computations, and writing results back.
Sovereign AI and neocloud deployment
David Bennett, a former Tenstorrent executive, leads the Japanese neocloud ai&. The company started in 2026. It manages an eight megawatt deployment using NVIDIA, AMD, and Tenstorrent hardware. ai& plans to deploy $500 million of its $2 billion infrastructure raise by the end of 2026. The firm uses Tenstorrent’s Blackhole Galaxy servers to run DeepSeek R1 671B at 350 tokens per second per user. This performance comes from 16 Galaxy units containing 512 Blackhole chips. Does the industry prefer open hardware enough to abandon CUDA? You probably already know the difficulty of replacing NVIDIA’s software stack.
Morgenrot partners with Tenstorrent to integrate the Arthur job management tool with Tenstorrent hardware and software. This integration targets AI data centers and high-performance computing environments. The partnership combines Morgenrot’s management capabilities with Tenstorrent’s RISC-V and AI technologies. ai& focuses on sovereign AI. This approach ensures countries and large companies maintain control over their AI stacks. Bennett targets the Japanese market, where he aims to bridge the gap in AI deployment. The company also builds an AI lab to post-train open source models with local data. Tenstorrent counts ai& as its largest customer.
