PFN and Tenstorrent Co-Develop Processor Specialized for Generative AI

Preferred Networks (PFN) and US-based Tenstorrent announced on October 20, 2026 that they have begun jointly developing systems including processors for generative AI. Tenstorrent will handle CPU development, while PFN will develop AI processors, an area where NVIDIA GPUs have become the de facto standard. The two companies will develop AI processor systemization technology by leveraging optical-electronic fusion technology and cutting-edge packaging technology.

The project was selected through a public call by the New Energy and Industrial Technology Development Organization (NEDO). Research and development will continue through fiscal 2031, after which the companies aim to commercialize the results.

Tenstorrent is a semiconductor development company led by Jim Keller, who has designed chips at AMD, Apple, and Tesla. It develops processors based on the open-source RISC-V instruction set architecture and ships products including the "Blackhole" AI processor, positioned as an alternative to GPUs. In Japan, it is developing edge AI accelerators on 2nm logic technology together with semiconductor company Rapidus.

PFN, in addition to developing the domestic LLM "PLaMo," is working with Kobe University on the "MN-Core," a dedicated processor optimized for matrix operations, which are central to AI computation. Its features include boosting computing performance while suppressing power consumption, and it has won first place in the power-efficiency category of the global supercomputer rankings three times. The next-generation model "L1000," announced in November 2024, improves LLM inference speed through wide memory bandwidth. A workstation equipped with the L1000 is said to achieve inference speeds of roughly 500 to 1,000 tokens per second for a 70-billion-parameter (70B) LLM. For future generations, the company believes a single node could run inference on a 1-trillion-parameter model with a 128K context length at about 500 tokens per second. The initial planned release of the L1000 was 2026, but it has now been revised to 2027.

In high-memory-bandwidth AI-specialized processors, US companies Cerebras and Groq are ahead. Cerebras has partnered with OpenAI and is part of the inference infrastructure for ChatGPT's "Ultrafast" mode. Regarding Groq, it has been reported that NVIDIA acquired a non-exclusive license to its technology for $20 billion and that it will be provided as part of the next-generation AI supercomputer "Vera Rubin."

Optical-electronic fusion technology transfers data using light instead of electricity. It is considered important for efficient AI computation spanning multiple nodes, as it enables high-capacity communication while consuming less power than electrical approaches.

Preferred Networks (PFN) and US-based Tenstorrent announced on October 20, 2026 that they have begun jointly developing systems including processors for generative AI. Tenstorrent will handle CPU development, while PFN will develop AI processors, an area where NVIDIA GPUs have become the de facto standard. The two companies will develop AI processor systemization technology by leveraging optical-electronic fusion technology and cutting-edge packaging technology.

The project was selected through a public call by the New Energy and Industrial Technology Development Organization (NEDO). Research and development will continue through fiscal 2031, after which the companies aim to commercialize the results.

Tenstorrent is a semiconductor development company led by Jim Keller, who has designed chips at AMD, Apple, and Tesla. It develops processors based on the open-source RISC-V instruction set architecture and ships products including the "Blackhole" AI processor, positioned as an alternative to GPUs. In Japan, it is developing edge AI accelerators on 2nm logic technology together with semiconductor company Rapidus.

PFN, in addition to developing the domestic LLM "PLaMo," is working with Kobe University on the "MN-Core," a dedicated processor optimized for matrix operations, which are central to AI computation. Its features include boosting computing performance while suppressing power consumption, and it has won first place in the power-efficiency category of the global supercomputer rankings three times. The next-generation model "L1000," announced in November 2024, improves LLM inference speed through wide memory bandwidth. A workstation equipped with the L1000 is said to achieve inference speeds of roughly 500 to 1,000 tokens per second for a 70-billion-parameter (70B) LLM. For future generations, the company believes a single node could run inference on a 1-trillion-parameter model with a 128K context length at about 500 tokens per second. The initial planned release of the L1000 was 2026, but it has now been revised to 2027.

In high-memory-bandwidth AI-specialized processors, US companies Cerebras and Groq are ahead. Cerebras has partnered with OpenAI and is part of the inference infrastructure for ChatGPT's "Ultrafast" mode. Regarding Groq, it has been reported that NVIDIA acquired a non-exclusive license to its technology for $20 billion and that it will be provided as part of the next-generation AI supercomputer "Vera Rubin."

Optical-electronic fusion technology transfers data using light instead of electricity. It is considered important for efficient AI computation spanning multiple nodes, as it enables high-capacity communication while consuming less power than electrical approaches.