Preferred Networks

PFN / 株式会社Preferred Networks

Preferred Networks is a privately held Japanese AI company that develops its own AI accelerators, computing infrastructure, Japanese-language foundation models, and industrial applications.

Private company
Profile as of September 14, 2026
Not ranked: no published revenue or market capitalisation
How the ranking works →

Preferred Networks' vertical integration creates opportunities for hardware-software optimization and domestic AI sovereignty, while requiring it to compete against global hyperscalers and frontier-model developers with far larger capital budgets. The absence of audited financial and operating data leaves the economic durability of that strategy unproven.

Key figures

Revenue: Not disclosed
Revenue growth: Not disclosed
Operating margin: Not disclosed
Net income: Not disclosed
R&D spend: Not disclosed
Capital expenditure: Not disclosed
Latest private valuation December 2024: More than 300 billion yen
Total capital raised in latest round 2024-12-23 to 2025-06-30: At least 24 billion yen
MN-Core 2 peak performance: 393 TFLOPS FP16
PLaMo-100B parameters: 102 billion

Overview

Preferred Networks is a privately held Japanese AI company founded in 2014 that attempts vertical integration across the AI stack. It develops its own AI accelerators, computing infrastructure, Japanese-language foundation models, and industrial applications. Its semiconductor product is the MN-Core series, which includes the 7-nanometre MN-Core 2 and the upcoming MN-Core L series designed for generative-AI inference. Beyond hardware, the company provides the Preferred Computing Platform cloud service and the PLaMo family of foundation models. This integrated approach allows it to optimize systems for targeted workloads, particularly in industrial automation, materials discovery, visual inspection, mobility, and robotics. While its strategic relevance comes from this broad technological footprint, its financial disclosure, commercial scale, customer concentration, and chip-production economics remain largely private.

The AI angle

Preferred Networks earns revenue through direct product development and deployment across multiple layers of the AI stack. Demand for Japanese-language generative AI supports its PLaMo models and related APIs, while demand for local AI compute supports its Preferred Computing Platform and MN-Core hardware. Industrial physical-AI work drives its inference chips, robotics, autonomous machinery, and automotive projects. The company matters most in the silicon and model developer layers because it combines proprietary AI-accelerator design with Japanese-language models and deploys both through its own compute and industrial application stack. Its model layer includes PLaMo Prime and PLaMo 2.1 Prime, with the PLaMo-100B model featuring 102 billion parameters trained on 2 trillion tokens. The company also targets edge devices such as automobiles and manufacturing equipment with PLaMo Lite. Despite this broad product portfolio, the company does not disclose the percentage of revenue attributable to AI processors, cloud services, models, or applications, making its exact commercial scale in each segment difficult to assess.

Technology and moat

The company's primary differentiator is its attempt to control a larger portion of the AI stack than most Japanese start-ups. It designs MN-Core accelerators, builds systems around them, operates an associated cloud platform, develops Japanese-language models, and adapts these technologies to physical and industrial applications. This gives it more control over workload optimization than a model-only developer or a systems integrator using third-party GPUs. The MN-Core approach emphasizes performance per watt and system-level efficiency, with MN-Core 2 consuming 330 watts while providing 393 TFLOPS FP16 peak performance. Its language models are differentiated by their Japanese-language training and deployment positioning. However, the company must fund costly chip design, wafer procurement, systems deployment, model training, and application delivery without the disclosed revenue base, capital budget, or external customer volumes of global hardware rivals like NVIDIA or AMD. It does not lead these competitors on broad ecosystem depth, developer tooling, installed base, or training scale.

Five-pillar assessment

Scale and market position: A prominent Japanese AI start-up with a valuation exceeding 300 billion yen, though its commercial scale, revenue, and accelerator shipment volumes remain undisclosed.
Technology and R&D: Develops the 7-nanometre MN-Core 2 accelerator, liquid-cooled AI servers, and the 102-billion-parameter PLaMo foundation model optimized for Japanese.
Supply-chain centrality: Backed by major Japanese industrial and financial groups, with key partnerships including Toyota, Mitsubishi Corporation, and TSMC for chip manufacturing.
Financial momentum: Raised at least 24 billion yen in a 2024-2025 financing round and launched its PFCP cloud service and PLaMo Prime models, though financial trajectory is private.
Governance and quality: Independent private company led by its founders, with strategic corporate shareholders but limited financial and operational transparency.

Supply chain and relationships

Preferred Networks depends on external semiconductor foundry capacity to manufacture its proprietary hardware. Credible press reporting states that Taiwan Semiconductor Manufacturing Company manufactures the 7-nanometre MN-Core 2 accelerator. The company also relies on external suppliers for advanced packaging, high-bandwidth memory, EDA software, and server components. Historically, its supercomputer clusters have used NVIDIA GPUs, Intel processors, and Mellanox networking. On the customer side, named relationships are limited but include Japan's Digital Agency, which adopted PLaMo Translate for its government AI environment, and the National Bank of Cambodia for an AI liquidity-forecasting proof of concept. Toyota conducts joint research with the company on physical-AI inference using the MN-Core L Series. The company's visible relationships are heavily concentrated among Japanese strategic investors, government institutions, and industrial groups, and it does not publicly identify its largest revenue customers or disclose revenue concentration.

Customers

Japan Digital AgencyAdopted PLaMo Translate for Gennai government AI environment
National Bank of CambodiaAI liquidity-forecasting proof of concept

Suppliers

TSMCReported manufacturer of MN-Core 2
NVIDIAHistoricalHistorical infrastructure relationship
IntelHistoricalHistorical infrastructure relationship
Mellanox TechnologiesHistoricalHistorical infrastructure relationship

Partners

Toyota Motor CorporationJoint research on MN-Core L Series inference
Mitsubishi CorporationCo-established PFCI joint venture
Internet Initiative Japan Inc.Co-established PFCI joint venture
Mitsubishi Heavy Industries, Ltd.AI for mission-critical machinery
GMO Internet Group, Inc.GMO Preferred Security joint venture

Competitors

NVIDIAInferredCompetes in AI acceleration and cloud computing
Advanced Micro DevicesInferredCompetes in AI accelerators
Fujitsu LimitedInferredCompetes in Japanese HPC and AI infrastructure

Geopolitics and risk

The company's reliance on external foundries creates material geographic and geopolitical exposure. Because it designs AI processors but does not own a fabrication plant, its reported use of TSMC for manufacturing exposes it to foundry allocation limits, advanced-node availability, and packaging capacity constraints. This concentration in Taiwan means that any cross-strait escalation, shipping disruption, or regional instability could delay MN-Core production and limit the company's ability to supply servers and cloud capacity. Furthermore, the company depends on advanced semiconductor manufacturing tools, EDA software, high-bandwidth memory, and server components that sit within US-led export-control regimes. While its strategic shareholder base links it to major Japanese industrial groups, which supports domestic commercial access and aligns with Japanese AI sovereignty goals, this may also concentrate its business-development focus in Japan and limit its global diversification.

Risk matrix
Risk	Severity	Why it matters
External foundry dependence	High	Relies on TSMC for manufacturing, creating exposure to foundry allocation and packaging capacity.
Taiwan Strait exposure	High	Disruption could delay MN-Core production and limit server and cloud capacity.
US export-control exposure	High	Depends on US-regulated EDA software, memory, and semiconductor manufacturing tools.
Nvidia ecosystem competition	High	Must compete against Nvidia's massive installed base, developer ecosystem, and cloud availability.
Capital-intensity risk	High	Scaling custom chips, cloud capacity, and model development requires continuing capital.
Commercialization risk for MN-Core L	High	Commercial availability, yields, and customer orders are not disclosed.
Private-company disclosure risk	High	Lack of audited financials constrains external assessment of commercial progress.
Revenue-concentration risk	Medium	Visible relationships are concentrated among Japanese strategic investors and government institutions.

Governance and ownership

Preferred Networks is an independent, privately held company led by its co-founders, Toru Nishikawa and Daisuke Okanohara. Its strategic shareholder list includes Toyota Motor, Fanuc, NTT, Hitachi, Mitsubishi Corporation, Chugai Pharmaceutical, and the Development Bank of Japan. While these relationships support commercial access, the company does not disclose ownership percentages, voting-right distribution, or executive compensation. As a private entity, it does not publish audited consolidated financials, segment revenue, operating cash flow, capital expenditure, or customer concentration. This lack of financial transparency constrains external assessment of its commercial progress and the economic durability of its capital-intensive vertical integration strategy.

What to watch

Commercial availability, performance benchmarks, and customer shipments for the MN-Core L1000 and MN-Core L Series.
Whether PFN secures disclosed production capacity from TSMC or another foundry for inference accelerators.
Whether PFCP obtains named commercial customers beyond PFN-group projects, and PFCI begins operations in early 2026.
Whether the Digital Agency expands PLaMo Translate deployment to additional ministries and agencies.
Whether Toyota's physical-AI research produces a commercial MN-Core L design win for robotics or vehicles.
Whether PFN raises additional capital, publishes an IPO timetable, or provides audited financial disclosure.

Recent developments

June 2026Began joint research with Toyota's Frontier Research Center on MN-Core L Series inference.
December 2025Japan's Digital Agency adopted PLaMo Translate for its Gennai government AI environment.
October 2025Released PLaMo 2.1 Prime with automated tool calling.
June 2025Raised an undisclosed equity amount in a further extension of its financing round.
December 2024Launched PLaMo Prime and announced a 19 billion yen first close in a financing round.
October 2024Launched Preferred Computing Platform, a cloud service based on MN-Core 2.

Coverage on AsiaAI.FYI

Why the 35% Surge in Global Chip Sales Is a Warning for Western Tech Investors? August 7, 2026
Zhipu AI Surges Past Trillion Yuan Market Cap in China’s AI Boom June 22, 2026

Guides that cover Preferred Networks

About this profile

Compiled with AI-assisted research from company filings, market data, and published reporting as of September 14, 2026, then reviewed by AsiaAI.FYI. Figures marked Estimate are not company-reported. Check primary filings before relying on any number.

Confidence: C. Identity, products, announced financing, and named partnerships have direct company support, but audited financial performance, revenue history, and shipment volumes are undisclosed.

Main sources: PFN corporate company disclosures; PFN press releases; PFN technical product pages; Credible business press reporting; Technical industry reporting.

All 54 sources

Position in the AI stack

A vertically integrated AI provider spanning custom accelerator design, cloud infrastructure, Japanese-language foundation models, and industrial applications.

Legal name: Preferred Networks, Inc.
Headquarters: Tokyo, Japan
Founded: 2014
Founders: Toru Nishikawa
Daisuke Okanohara
Leadership: Daisuke Okanohara, Co-Founder and Chief Executive Officer
Toru Nishikawa, Co-Founder and Chairman
Ownership: Private company
Business: AI accelerators, cloud infrastructure, foundation models, and industrial applications
Sector: AI Software & Models
Chip Design
Servers, PCs & ODMs
Website: preferred.jp/en