Preferred Networks, Inc.

What Preferred Networks is

Preferred Networks, Inc. (PFN) is a Japanese technology company. Its head office is in the Otemachi Building, 1-6-1 Otemachi, Chiyoda-ku, Tokyo ^[1]. The company was incorporated on 26 March 2014, and the history page on its own site dates its founding to March 2014 ^[1]. It is led by two co-founders: Toru Nishikawa is representative director and chairman, and Daisuke Okanohara is representative director and president ^[1]. In Japanese corporate practice the chairman is the senior board role while the president is normally the executive who runs the business day to day, so the two founders sit at the head of both the board and the operating side.

What the company works on

PFN describes itself as vertically integrating semiconductor and computing infrastructure through to solutions and applications ^[1]. Vertical integration means that one organisation covers several layers of a technology stack itself instead of assembling a product from parts bought from other suppliers. Here the layers named run from silicon and computing hardware up to the solutions and applications a customer finally uses.

The hardware side is the most visible. The company develops the MN-Core series of AI processors jointly with Kobe University ^[1]. An AI processor is a chip designed around the kind of arithmetic that neural networks repeat in very large volume — long runs of multiplications and additions on big blocks of numbers — rather than around the general-purpose instruction handling of an ordinary CPU. Building such chips together with a university means the design work is shared between company engineers and academic researchers.

PFN also runs its own supercomputer, MN-3, which took first place on the Green500 energy-efficiency ranking three times ^[1]. The Green500 lists supercomputers by how much computing they deliver per unit of electrical power, so a top position there is a statement about efficiency rather than about raw size. An in-house machine of this kind gives a company a place to test its own chips and to train models on hardware it controls.

On the model side, the company developed PLaMo, a Japanese large language model that it describes as built from scratch ^[1]. A large language model is a neural network trained on very large amounts of text so that it can continue, transform or answer questions about text; the company presents PLaMo as its own Japanese model developed in-house.

Shareholders and joint research

Shareholders listed by the company include Toyota Motor, Hitachi, Mitsubishi Corporation, Mitsubishi Heavy Industries, FANUC, NTT and SBI Group ^[1]. The list spans car manufacturing, heavy industry, factory automation, general trading, electronics and IT services, and financial services — large established Japanese firms rather than venture funds alone. That mix of owners is unusual for a company of this age and says something about where its technology is expected to be used.

The company also runs joint research with leading firms in industrial robotics, pharmaceuticals and energy ^[1]. Joint research of this kind means researchers from the company and engineers from the partner work on the same problems, usually with the partner supplying domain knowledge and data from its own field and the company supplying machine-learning methods.

Where it sits

Taken together, the company's own description places it across several layers of computing at once: chip design, a supercomputer, and large language models, with industrial partners and shareholders drawn from manufacturing, pharmaceuticals and energy ^[1]. Most AI companies occupy one of those layers. PFN's stated position is that it works on all of them, which is what the phrase about vertical integration is meant to convey ^[1]. For a reader trying to follow the field, it is a useful example of a company whose bet is that the hardware and the models have to be developed together, and whose backers come from the industries that would apply them.