Everything About AI
NextFin News - Preferred Networks, Japan's most valuable startup, is seeking a public listing in Tokyo, and the move exposes the uncomfortable arithmetic behind the global AI chip race: even a 300 billion yen champion runs out of private money when the bill is measured in the tens of billions. The company's bet is that public-market capital, plus a strategy built on efficiency rather than brute force, can keep it in a contest dominated by a rival whose quarterly research budget now exceeds Preferred Networks' entire lifetime funding.
The listing push arrives as Tokyo's IPO window is unusually welcoming. Between January and August 2026, the average market capitalization of companies newly listed on the Tokyo Stock Exchange reached a record 30.9 billion yen, roughly 60% above both 2025 and the 2013-2025 average. Japanese startups are waiting longer and arriving bigger. But Preferred Networks' target valuation is roughly ten times that average, and the capital it needs to turn chip designs into a real business dwarfs anything the domestic market has recently absorbed. This is not just a funding event. It is a stress test of whether a mid-sized Japanese firm can sustain a credible challenge in AI silicon against competitors with ten to a hundred times its resources.
The Company That Refused to Pick a Lane
Preferred Networks was founded on March 26, 2014, as a spin-off from Preferred Infrastructure by Toru Nishikawa and Daisuke Okanohara, and it has spent more than a decade refusing the usual startup choice between software and hardware. The Tokyo-based firm designs its own deep-learning accelerators, operates its own supercomputers, trains foundation models under the PLaMo brand, and sells industry-specific AI applications into manufacturing, retail, healthcare, robotics and entertainment. That vertical integration is now colliding with a capital-intensive reality.
The financing trail shows how quickly the bill has grown. In December 2024 the company raised 19 billion yen in a round led by SBI Group, with participation from Sekisui House Innovation & Communication, the Development Bank of Japan, Mitsubishi Corporation and Wacom, alongside debt from Shoko Chukin Bank, SMBC, MUFG Bank and Resona Bank. That took total known funding to approximately 36 billion yen. Extension rounds in April and June 2025 added at least another 5 billion yen, and the company has said it intends to keep raising through the cycle. Its valuation has climbed past 300 billion yen, making it Japan's highest-valued unicorn.
The IPO is the next step on a path the leadership has been describing in public for some time. Nishikawa, who co-founded the company and now serves as chairman, said in 2024 that the company aimed to raise tens of billions of yen over the following year or two and was courting overseas investors for the first time. The company's president, Daisuke Okanohara, framed the ambition plainly in a 2025 interview:
"For the past 10 years, we have been working on using AI to solve real-world problems."
Nishikawa added the geographic constraint that now drives the listing:
"There are limits to what we can do in Japan alone, and the global market is much bigger than the Japanese market."
The timing is deliberate on both sides of the balance sheet. On the demand side, the Tokyo market is paying up for scale: of the 15 companies that listed on the exchange's emerging markets from January to August 2026, only seven were valued below 10 billion yen at listing, while outliers such as GO and autonomous-driving developer Tier IV pulled the average to its record. On the supply side, Preferred Networks is approaching the point where private rounds cannot keep pace with chip-development costs. A listed equity currency also buys something private funding cannot: the ability to do follow-on offerings, to use shares for hiring and acquisitions, and to offer the disclosure discipline that overseas partners and enterprise customers increasingly demand.
Why a Software Company Started Making Chips
The first question any investor will ask is why an AI startup would take on one of the hardest problems in technology. The answer is a bottleneck that hit the industry around 2022: the accelerators everyone wanted were made by one company, and demand far outstripped supply. Preferred Networks' own research and its supercomputer operations gave it a front-row seat, and it drew a specific conclusion. The way to compete with Nvidia is not to out-brute-force it. It is to out-specialize it.
The company's MN-Core processors are built for a narrow class of workloads — the matrix operations at the heart of deep learning — and discard the general-purpose circuitry that makes GPUs flexible but power-hungry. The first-generation MN-Core, fabricated on TSMC's 12-nanometer process, drew 500 watts and delivered 524 TFLOPS at half precision. The MN-3 supercomputer built around 160 of those chips topped the Green500 list of the world's most energy-efficient supercomputers three times between 2020 and 2021, peaking at 39.38 gigaflops per watt with 64% execution efficiency.
The second generation sharpened the pitch. MN-Core 2 reaches 393 TFLOPS at TF16 precision with an efficiency of 1,192 gigaflops per watt. In the company's own comparison materials, that sits against Nvidia's A100 at 312 TFLOPS FP16 drawing 400 watts on a 7-nanometer process, the H100 at 989 TFLOPS at 700 watts on 4 nanometers, and AMD's MI300X at 1,300 TFLOPS at 750 watts — with MN-Core 2 drawing 330 watts. The claim is not raw performance leadership. It is performance per watt and performance per dollar, the two metrics that matter when a data center's electricity bill begins to exceed its hardware bill.
The newest piece is the MN-Core L1000, an inference processor under development for generative AI that uses a 3D-stacked architecture placing memory directly above logic. The company says this delivers far higher bandwidth than the high-bandwidth memory in current high-end GPUs, and because it uses large-capacity DRAM rather than costly SRAM, it can run models with more than 70 billion parameters on a single server at up to ten times the token throughput of existing processors. The target availability is 2027.
This is the crux of the strategy. Nvidia's stronghold is frontier training, where customers pay for maximum performance and ecosystem compatibility. Preferred Networks is aiming at inference — the cheaper, higher-volume end of the market where efficiency, memory bandwidth and total cost of ownership matter more than peak FLOPS. It is a narrower beachhead, but it is a beachhead that actually exists for a company of this size.
The Capital Problem Behind the Technology
Technology is not the binding constraint. Money is. Designing a chip is expensive; taping out at an advanced node, packaging it and building a supply chain is far more expensive. In February 2024, Preferred Networks announced that Samsung Electronics would manufacture its AI accelerators using Samsung's 2-nanometer gate-all-around process and 2.5D advanced packaging — the first publicly revealed order for Samsung's cutting-edge 2-nanometer contract process. That is a genuine technical coup. It is also a commitment that runs into the hundreds of millions of dollars once masks, engineering samples, packaging and inventory are counted.
Junichiro Makino, a chief technology officer at Preferred Networks, said at the time:
"We are excited to lead in AI accelerator technology with Samsung Electronics' 2nm GAA process. This solution will significantly support Preferred Networks' ongoing efforts to build highly energy-efficient, high-performance computing hardware that meets the ever-growing computing demands from generative AI technologies, especially large language models."
Private markets have been generous, but they have limits. Preferred Networks' cumulative disclosed funding is on the order of tens of billions of yen — roughly a quarter of a billion dollars at recent exchange rates. Nvidia, by contrast, reported research and development spending of $7.05 billion in the quarter ended July 2026, and about $23.6 billion over the trailing twelve months. The gap cannot be closed by venture rounds. An IPO converts a private story into a public currency, and it imposes the disclosure and governance discipline that makes a company legible to the pension funds, corporate partners and government programs it needs as partners.
There is also a national-policy dimension that investors should not ignore. Japan has made semiconductor sovereignty a strategic priority, backing the leading-edge foundry Rapidus and channeling public capital into AI. Preferred Networks' alliance with SBI, its government-linked chip efforts and its work with Samsung fit a pattern: Tokyo wants at least one credible domestic AI-silicon contender. A listed Preferred Networks is easier for public funds, corporate partners and domestic institutions to back than a private startup. The state has an interest in the company's success, which lowers some risks — but it also means the company's fortunes are tied to policy continuity as much as to product execution.
What the Market Is Pricing, and What It Is Not
The conventional read of this IPO is straightforward: Japan's AI champion lists, investors cheer the Nvidia-alternative narrative, and the stock rises on scarcity value. That is the first-order trade, and it is probably already embedded in the company's private valuation. The 300 billion yen mark was set in a funding environment where AI premiums were near their peak, and where "Japan's only AI chip contender" carried a monopoly-on-patriotism premium.
The second-order question is harsher, and it is the one public markets will eventually force. Investors do not pay for roadmaps; they pay for revenue, margins and proof that technology converts into cash. Preferred Networks' last disclosed revenue was 8.486 billion yen for the fiscal year ended January 2021. Its business today spans custom AI solutions, foundation models — PLaMo 2.0 Prime won a Nikkei product award in early 2026 — and chip development. None of that is yet a chip-revenue story. The MN-Core L1000 is not due until 2027 at the earliest, and until then the company will be burning cash on design, tape-outs, packaging and inventory.
That creates an expectation gap that will define the stock's second act. The listing narrative sells Preferred Networks as an AI chip company. The financials, for now, are those of an AI services and research company with a chip project attached. Investors who buy the former valuation will eventually demand the latter's economics, and the transition will be measured in quarterly disclosures that offer few places to hide.
There is also a competitive timing risk that the scarcity narrative papers over. By 2027, when the L1000 is meant to arrive, Nvidia's next-generation architectures will not be standing still, and a field of well-funded challengers — Cerebras, Groq, Tenstorrent, AMD, and the hyperscalers' in-house silicon — will have moved further down their own learning curves. Preferred Networks' window is narrow: it must prove its efficiency advantage in real customer deployments before incumbents compress the same metric through process-node advantages the company cannot match on its own. The 2-nanometer Samsung timeline is the fulcrum; any slippage there undermines the entire valuation.
Commercialization is beginning to show up beside the roadmap. In March 2026 the company formed a joint venture, GMO Preferred Security, with GMO Internet Group and GMO Cybersecurity by Ierae, taking a 49% stake in an entity designed to sell a Japan-built AI environment spanning hardware to software. It is a small step toward turning chip capability into billable infrastructure, and it signals that management knows the listing story needs revenue, not just silicon.
The Counter-Thesis: Vertical Integration Is a Trap
The strongest argument against Preferred Networks' approach is that vertical integration in semiconductors is a trap for a company of its size. Nvidia's moat is not just silicon; it is CUDA, a software ecosystem thirty years in the making that locks developers into its hardware. AMD has spent more than a decade and tens of billions of dollars trying to pry that lock open with ROCm, with gains that remain incremental. Preferred Networks' custom stack — its own chips, its own compilers, its own models — multiplies the integration burden. Every layer it builds in-house is a layer it cannot improve as fast as a specialist that focuses on that layer alone.
The counter-thesis also has a demand-side prong. An efficiency edge matters only if customers value lower total cost of ownership more than they value ecosystem compatibility. Enterprise buyers have repeatedly shown they will pay a premium for Nvidia because migrating models, tooling and staff is expensive and risky. A 30% efficiency gain does not overcome that inertia unless the price gap is overwhelming, and Preferred Networks has not yet demonstrated that gap in a production deployment at scale.
The falsifying signal is concrete and observable. If Preferred Networks cannot announce at least one named, paying enterprise customer for MN-Core 2 or the L1000 by the end of fiscal 2027, and if chip-related revenue remains below 10% of total revenue two years after listing, the vertical-integration thesis has failed and the company should be valued as a services firm, not a semiconductor contender.
Outlook: Three Time Horizons, Three Different Stocks
The IPO is the right move at the right time, but it converts a private-market story into a public-market test, and the judgment changes depending on the horizon.
In the short term, the listing should price well. Scarcity value is real: Japan has almost no listed pure-play AI champions, the domestic IPO window is open, and the AI-infrastructure narrative still commands a premium. Early trading will likely reward the story, particularly if the final offer size lands in the tens of billions of yen range the company has signaled.
Over the medium term, the stock will be judged on milestones, not narrative. The key markers are the L1000's 2027 delivery, named customer wins for MN-Core 2, and the trajectory of chip gross margins. Any slippage on the Samsung 2-nanometer timeline would hit hard, because the entire valuation rests on the chip roadmap being real and on schedule. Base case: the company lists at a rich but defensible multiple, delivers the L1000 on time, and chips reach low-double-digit revenue share by year three. Upside case: the L1000 demonstrates a clear cost-per-token advantage in production, enterprise customers sign, and the stock rerates as a genuine semiconductor name. Downside case: the 2-nanometer schedule slips, chip revenue stays immaterial, and the market reprices the company as an AI services firm at a fraction of its listing multiple.
In the long term, the structural question is whether a mid-sized Japanese firm can sustain a credible challenge in AI silicon against competitors with an order of magnitude more resources. The cyclical leg of the AI capital-expenditure boom will not last forever; when it turns, efficiency-focused challengers benefit, but only if they have survived to that point. Preferred Networks' bet is that inference economics will matter more than brand by the end of the decade. That is a structural claim, and it is defensible — but it is not yet proven.
What to watch is a short list: the final IPO size and valuation, the 2027 L1000 launch, named customer wins, quarterly research spending as a share of revenue, and the single most telling number — chip revenue as a percentage of total revenue two years post-listing.
Preferred Networks is not trying to beat Nvidia at Nvidia's game. It is trying to change the game to one where watts and memory bandwidth matter more than brand — and this IPO is the bill for that ambition. Whether the market pays it depends on whether efficiency, eventually, beats ecosystem.
Explore more exclusive insights at nextfin.ai.
