Japan’s AI Third Pole Strategy: Reducing US LLM Dependency and Performance Realities

Japan’s domestic AI is pursuing a “third pole” strategy to reduce reliance on US LLMs. Based on METI policy and vendor efforts, this article examines current performance gaps and future outlook.

📑Table of Contents

1. Background and Objectives of Japan’s Domestic AI “Third Pole” Initiative
2. Performance Comparison: Gaps with Mythos and Domestic Model Strengths
3. METI Policy and Support Measures for Reducing US Dependency
4. Enterprise Cases: Fujitsu PHOTON and NEC Initiatives
5. Role of Hybrid Architectures and Open-Weight Models
6. Challenges and Future Outlook
7. Summary and Implications for Readers

What to understand first

---

Background and Objectives of Japan’s Domestic AI “Third Pole” Initiative

Japan finds itself positioned between US models from OpenAI and Anthropic and Chinese counterparts. METI’s 2026 AI strategy update positions domestic models as a “third pole,” aiming to decrease dependence on US LLMs for enterprise and government applications.

The push stems from concerns over data sovereignty and security, driving the need for AI infrastructure that can be managed domestically. ITmedia AI+ reporting confirms that multiple Japanese companies announced de-US dependency roadmaps in June 2026.

---

Performance Comparison: Gaps with Mythos and Domestic Model Strengths

Domestic models trail frontier systems by 15-25% on general benchmarks like MMLU and coding tasks. However, they excel in Japanese-language processing and domain-specific applications such as manufacturing and healthcare.

Metric	Mythos (frontier)	Domestic Models	Gap/Advantage
MMLU	85-90%	65-75%	15-25% behind
Coding	High	Medium-High	10-20% behind
Japanese	Standard	Strong	+5-10% advantage
Manufacturing Domain	Standard	Strong	Advantage

Source: ITmedia AI+ (June 2026), METI AI Strategy documents (June 2026)

---

METI Policy and Support Measures for Reducing US Dependency

METI targets 30% domestic AI infrastructure share in critical sectors by 2028. Subsidies and preferential procurement are being used to support vendors like Fujitsu and NEC.

Key measures include promoting open-weight models through partnerships with RIKEN and AIST, and developing guidelines for hybrid usage that combine local and selective external APIs.

---

Enterprise Cases: Fujitsu PHOTON and NEC Initiatives

Fujitsu’s PHOTON architecture claims 475x Transformer throughput via hierarchical chunk processing and multi-query integration. It targets enterprise use cases with practical performance.

NEC is expanding government and municipal deployments with its own models. Both vendors emphasize hybrid strategies that minimize US API calls while maintaining flexibility.

---

Role of Hybrid Architectures and Open-Weight Models

Full domestic self-sufficiency remains challenging. Hybrid approaches—local fine-tuning as the base with selective frontier API calls—are emerging as the practical path.

Open-weight models from research institutions enable domestic customization, helping control costs, latency, and security requirements simultaneously.

---

Challenges and Future Outlook

The primary challenge is closing the performance gap. Achieving the 2028 target will require increased compute investment and expanded datasets, plus securing skilled talent.

Expect more concrete benchmark results and enterprise adoption cases following the June 2026 announcements. The effectiveness of METI policies will be decisive.

---

Summary and Implications for Readers

Japan’s third pole AI strategy demonstrates a viable path toward reduced dependency, yet realistic performance limitations remain. Organizations should evaluate hybrid approaches tailored to their specific use cases.

Readers are encouraged to monitor METI official resources and ITmedia coverage for ongoing developments. Domestic model progress merits continued attention.

Frequently Asked Questions

Q1: What exactly is the third pole strategy?

It refers to Japan’s initiative to build an independent domestic AI ecosystem alongside US and Chinese powers. METI leads with a 30% share target in key sectors.

Q2: Are domestic models ready for practical use?

They lag on general tasks but perform adequately for Japanese and domain-specific work. Hybrid supplementation with frontier models is common.

---

Practical check

Q3: How does Fujitsu PHOTON perform?
A: It reports 475x throughput compared to standard Transformers, optimized for enterprise workloads.

Q4: Is the 2028 target achievable?
A: Progress depends on policy support and investment, but current announcements indicate steady advancement. Check METI documents for details.

---

Practical check

Q5: How should individuals or SMEs approach adoption?
A: Start with domestic model APIs or open-source versions, supplementing with frontier systems only when necessary.

Sources: ITmedia AI+ (https://www.itmedia.co.jp/aiplus/), METI AI Strategy (https://www.meti.go.jp/policy/ai/) (as of June 2026)