Preferred Networks is pushing deeper into custom semiconductors tuned to run AI models, with the MN-Core L series positioned as its first chips intended for broad commercial sale. The lineup focuses on generative AI inference - running models rather than training them - and the company says it delivers quick-response AI computation for devices and robots.
The company has spent about a decade building AI hardware, but up to now its chips have mostly been used by a small circle of research institutions and university labs. This next step is about turning that R&D into a product business.
The goal is to be profitable within three years, after which the company would be positioned to go public.
Production is the chokepoint. Surging AI investment is tightening supplies of key components, while capacity at foundries remains scarce. Preferred Networks is already encountering "significant challenges" securing manufacturing slots as chipmakers worldwide negotiate for priority. The MN-Core L series is planned to be fabricated on an older process node at Taiwan Semiconductor Manufacturing Co. As Okanohara put it, "The risk isn't whether we can sell the chips, but whether we can make enough of them."
This marks a shift from a long-held preference for engineering independence. The company previously leaned on backing from investors including Toyota Motor Corp. and Fanuc Corp., but the economics of AI hardware have changed, with Nvidia Corp. defining the pace and scale. Okanohara says Preferred Networks now needs far more scale than it once expected to meet rising demand from customers who want to reduce procurement and operating costs and rely less on a single US supplier.
"We will need to consider an initial public offering over the next few years to finance the huge capital needs for our semiconductor business," he said.
Markets move every day, and most of it never reaches you in time. Get the free Market Briefs daily newsletter and stay ahead of it.
Preferred Networks also builds large-language models tailored to industrial uses and is working with SoftBank Corp., Sony Group Corp., and Honda Motor Co. on a homegrown multimodal foundational model via their joint venture, Noetra Corp. Okanohara argues that countries need homegrown options across hardware, networking gear, and software to avoid sudden loss of access. As a cautionary example, he cited the US government's temporary ban earlier this year on foreign nationals using Anthropic PBC's Claude Fable and Mythos models.
The company's roots go back to 2006, when Okanohara and Toru Nishikawa co-founded the precursor to Preferred Networks while they were pursuing graduate studies at the University of Tokyo. He says they focused on engineering instead of profile-building - a choice he now regrets - and adds they should have promoted the company more, especially overseas. He insists their AI models are built entirely in-house and can compete globally, with more projects coming. In his words, "We're going to bring AI services originating in Japan to the world."
For your wallet, here's the takeaway: building a non-Nvidia path for AI requires big capital, scarce fab time, and patience. Preferred Networks is aiming for profitability in three years and lining up for an eventual listing, a timeline worth watching if you care where the next wave of AI infrastructure might originate.
