Preferred Networks (PFN) is a Tokyo-headquartered deep-learning research and AI infrastructure company that has positioned itself at the intersection of academic research and Japanese industrial application.
Founded in 2014 by Toru Nishikawa and Daisuke Okanohara as a spinout of Preferred Infrastructure, PFN built early credibility through its open-source deep-learning framework Chainer and through high-profile partnerships with Toyota Motor Corporation in autonomous driving and FANUC in industrial robotics.
The company has expanded into custom deep-learning hardware (the MN-Core line of accelerator chips), large-language-model research, and applied AI products in materials informatics, drug discovery, and industrial optimisation. Strategic equity investors have included Toyota, FANUC, and NTT, anchoring PFN to Japan's industrial backbone in a way that few pure-play AI startups elsewhere have achieved.
PFN is consistently described as Japan's leading independent AI research lab and is a recurrent reference in Japanese government policy discussions about national AI capability. The company has raised multiple late-stage private rounds and was last reported at unicorn-tier valuation.
It has not announced an imminent JPX listing, and the company's deep ties to large Japanese industrial conglomerates suggest a structure that prioritises long-term partnership over rapid public-market exit.
Preferred Networks is a useful reference case for foreign founders thinking about how to structure an AI company that wants to plug into the Japanese industrial ecosystem.
KK with corporate strategic investors. PFN is a Kabushiki Kaisha. The KK form is required to receive priced equity investment from large Japanese listed conglomerates like Toyota and FANUC, who run formal corporate-venture programs that map naturally onto KK preferred-share rounds.
A GK (Godo Kaisha) would not have been workable here because GKs do not have the share-class flexibility that priced strategic rounds require, and corporate investors are uncomfortable with GK governance.
Departure from old conglomerate norms. Despite having Toyota and FANUC as major investors, PFN is not a Toyota subsidiary or a FANUC subsidiary. It operates as an independent KK with its own board, its own research direction, and its own product roadmap.
This is a clear departure from the old keiretsu pattern where smaller technology firms would be effectively absorbed into the operational orbit of their large industrial partners. The Mercari-freee-Money Forward generation of Tokyo tech firms broke this mold by remaining independent through and beyond IPO; PFN is the equivalent example on the deep-tech side.
Founder voting and Japanese AI policy context. PFN cannot use dual-class voting shares under standard Japanese listing rules, so founder influence is preserved through pre-IPO ownership concentration, supportive long-term industrial investors, and board composition.
The company is also a frequent participant in Japan's AI-policy discussions, including the AI Strategy Council and the LLM Subcommittee, which gives it standing that pure venture-funded peers do not enjoy.
PFN raises priced equity from large Japanese listed conglomerates like Toyota and FANUC. Those investors require KK governance, share-class flexibility for preferred stock, and audit-committee or kansayaku oversight, all of which a GK cannot deliver cleanly. KK is also the only form that can list on the JPX, which is a strategic option PFN preserves.
The most likely path is the standard Tokyo growth-tech route. PFN would file on the JPX Growth tier (the post-2022 successor to the Mothers segment), then migrate up to Prime once it meets a tradable market cap of at least 10 billion yen, 800 shareholders, a 35 percent free float, and English-language disclosure of all material announcements. Mercari, freee, Money Forward, and Sansan all walked this path.
For most AI companies there is no general foreign-shareholder cap. The Foreign Exchange and Foreign Trade Act does require pre-notification for large stake acquisitions in sectors deemed national-security-relevant, which has been expanded to cover certain dual-use AI categories and IT systems handling national-security data.
Routine consumer or industrial AI is unrestricted, but transactions in defense-adjacent AI categories are reviewed.
GK incorporation is faster, cheaper, has no notarisation requirement for the articles, no kansayaku obligation, and lower ongoing compliance cost. The trade-off is that GKs cannot list on the JPX, do not have the same share-class flexibility for priced rounds, and look unusual to domestic Japanese venture investors.
The compromise is to start as a GK for an early operating subsidiary and convert to KK before the first priced round.
