Building your own AI chip is supposed to be the ultimate act of technological independence. So it says something interesting that one of Japan’s most valuable AI companies has decided the fastest route to a homegrown chip runs through foreign wallets.
Preferred Networks, the Tokyo-based AI unicorn, is courting overseas investors for the first time, according to Nikkei Asia. The money is meant to speed up chip development and fund global expansion, with an IPO penciled in somewhere in the next three to five years. The company also recently pulled in an additional 5 billion yen in an extension round, bringing its latest funding round to 24 billion yen total.
The mainstream read on this is “Japan wants a piece of the AI chip race.” I think the more useful read is the opposite direction: even a company with a clear national-champion story, real customers, and government tailwinds cannot fund silicon on domestic capital alone. That is not a knock on Japan. It is a statement about how expensive the hardware layer of AI has become.
Why chips are the hardest thing to build alone
If you follow AI agents for a living, or just try to keep up, most of what you see is software. A model gets released, someone wraps it in a tool-calling loop, and suddenly it can book flights or triage support tickets. Software iterates in weeks. Chips do not.
Designing a processor means years of engineering before a single unit ships, then fabrication contracts, then a software ecosystem so developers can actually use the thing. Every one of those steps burns capital with no revenue attached. This is why almost nobody attempts it, and why the companies that do tend to be either enormous already or backed by very patient money.
Preferred Networks is neither enormous nor, apparently, sitting on enough patient domestic money. Hence the outreach abroad.
The timing is the story
Here is what makes this more than a routine fundraising note. Japan passed major reform of its foreign investment screening regime under FEFTA, with the legislation clearing on May 29, 2026. In plain terms, Japan has been tightening how carefully it reviews foreign money coming into sensitive sectors.
AI chips are about as sensitive as a sector gets. So Preferred Networks is walking toward international capital at the same moment the rules governing that capital just got rewritten. That is not necessarily a contradiction. Screening regimes are designed to sort investment, not block it. But it does mean the company is threading a needle: enough foreign capital to fund silicon, structured in a way that survives review.
I find that tension genuinely interesting, because it is going to show up again and again. Every country wants domestic AI capability. Almost no country can pay for it without outside help. The screening rules and the funding needs are pulling in opposite directions, and companies like Preferred Networks are where that pull gets resolved in practice.
What the Mitsubishi deal tells us
On June 2, 2026, Mitsubishi Heavy Industries and Preferred Networks announced a business alliance to jointly develop Japan-made AI technologies for mission-critical applications, aimed at making social infrastructure more intelligent and autonomous.
That phrase, mission-critical, matters for anyone trying to understand where AI agents are actually heading. Consumer chatbots can be wrong occasionally and the cost is mild annoyance. Infrastructure cannot. An agent monitoring industrial systems or managing power distribution operates under a completely different standard of reliability.
And that standard is exactly why chips enter the conversation. When you need predictable performance, low latency, and control over your own stack, borrowing someone else’s hardware roadmap becomes a liability. Custom silicon is one answer to the reliability problem, not a vanity project.
The non-technical takeaway
If you are reading this site to understand AI agents rather than semiconductor finance, here is what I would take away.
- The agent software everyone talks about sits on top of a hardware layer that very few companies can afford to build. That constraint shapes which agents exist and who controls them.
- National AI strategies are colliding with capital reality. Wanting sovereign AI and being able to self-fund it are separate questions.
- Mission-critical agents, the kind running infrastructure rather than answering emails, are pushing companies toward custom hardware. Expect more of these chip stories, not fewer.
- An IPO three to five years out means this is a long game. Chip bets are measured in years, not quarterly releases.
Preferred Networks may well pull this off. It has funding momentum, a major industrial partner, and a clear thesis about where reliable AI needs to run. But the more revealing detail is not the ambition. It is that the ambition required a passport.
đź•’ Published: