Preferred Networks, the Tokyo AI startup behind Japan's homegrown answer to Nvidia, is opening its books to foreign investors for the first time and eyeing an IPO to pay for mass-producing its own chips.
CEO Daisuke Okanohara told Bloomberg News this week that Preferred Networks wants overseas money in the company for the first time in its history. That's a shift. A decade of fundraising went almost exclusively to Japanese backers: Toyota, Fanuc, NTT, SBI Group. All Japanese money. Nikkei Asia reported the same courtship of foreign capital, tying it directly to the cost of scaling PFN's chip line. The company is one of Japan's most valuable AI startups. It carries a valuation of roughly $2 billion, or about 300 billion yen, according to Nikkei Asia and PitchBook.
Here's the thing that separates Preferred Networks from nearly every AI startup you've heard of: it isn't chasing a chatbot. PFN builds robots. It makes Kachaka, a small autonomous transport robot, and works with Toyota and Fanuc on industrial and manufacturing AI, the unglamorous work of getting machines to sort, carry and assemble things correctly on a factory floor. That's a different business than training a large language model, and it comes with a different chip problem.
Betting on its own silicon #
Training and running AI models on Nvidia's GPUs means competing for the same scarce hardware everyone else wants, and paying whatever premium that scarcity demands. PFN decided years ago it would rather design its own silicon. The result is the MN-Core line, a chip project the company has run since 2016, built around 3D-stacked memory rather than the architecture Nvidia uses. Different memory, different math. PFN says the approach lets its chips move data faster for the math AI training actually requires, without carrying the general-purpose flexibility a GPU pays for and PFN doesn't need.
PFN manufactures its current MN-Core chips on TSMC's 12-nanometer and 7-nanometer processes. But for its next-generation MN-Core L Series, it turned to Samsung's 2-nanometer node, a move TrendForce described in February 2024 as a snub to TSMC. Samsung offered PFN a turnkey deal - design, production, advanced packaging - in one relationship, rather than PFN piecing services together from separate vendors. It's a small order set against Samsung's overall foundry business. Small order, big stakes. Samsung's 2nm capacity is what PFN needs to make its chips fast enough to matter against Nvidia's newest GPUs.
Japan's Sakana AI and China's Qwen are rewriting the AI pecking order in ways US labs did not expect
Tokyo's Sakana AI released Fugu on June 22, a 7-billion-parameter orchestration model that outscores GPT-5.5 and Claude Opus 4.8 on software engineering benchmarks by routing tasks across frontier models rather than competing on raw scale. Meanwhile Anthropic has accused Alibaba of running 28.8 million fraudulent queries against Claude between... - Japanese AI startup challenging US dominance - Chinese AI model catching up to
PFN has already put real money behind that plan. It raised 19 billion yen in December 2024 in a round led by SBI Group, with the Development Bank of Japan, Mitsubishi Corporation, Sekisui House and Wacom also participating, then added another 5 billion yen in April 2025. That brought the total to 24 billion yen, roughly $126 million. PFN has said the funding goes toward developing and producing the MN-Core L1000, its most advanced chip yet.
The real test #
An IPO is not imminent. Okanohara has said the listing sits three to five years out, timed to when chip production actually scales rather than to any near-term funding gap. That's a longer runway than most AI startups admit to publicly. But PFN's real problem is capital intensity. Building and mass-producing custom silicon costs far more than fine-tuning someone else's model, and courting foreign investors now is how PFN buys itself time to get there without diluting the Japanese backers who got it this far.
Frankly, the more interesting story here isn't the IPO timeline. It's that a country largely absent from the Nvidia-versus-China chip narrative just put its most valuable AI company forward as a third option. That's new territory. Japan doesn't have an obvious national champion in this race. PFN's pitch, custom silicon aimed at robotics rather than chatbots, gives it a genuinely different story to tell investors in New York or London who've already heard every version of the LLM pitch. Whether that lands with foreign capital the way it has with Toyota and Fanuc at home is the real test PFN just signed up for.
Also read: Mark Spitznagel Says Stocks Will Melt Up Past 8,000 Before a 1929-Style Crash • Zhang Yiming Is Personally Building ByteDance's Real-Time World Model • Google's $3.2 Billion TeraWulf Backstop Shows How AI Data Centers Get Built
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.