Tokyo exits for Japanese corpora and Article 30-4 collection

Why Japan gets disproportionate weight in AI collection

Two reasons Japanese coverage matters beyond its share of the open web:

• Legal clarity for TDM. Japan's Article 30-4 is, as of 2026, the broadest text-and-data-mining carve-out in any major economy. For AI teams based outside Japan but running collection from a Japanese datacenter with a legal basis in Japan, this materially reduces copyright exposure on training-corpus workloads that would be more contested under EU or UK law.
• SquadProxy scarcity. High-quality Japanese long-form text is under- represented on the commodity corpus side (Common Crawl, CC-News). Active collection from Japanese-resident IPs returns substantially more JP content than collection via US exits, because many Japanese publishers geo-route or quietly geoblock.

What the JP pool routes through

• Datacenter — AWS ap-northeast-1 (Tokyo), GCP asia-northeast1, IIJ-hosted infrastructure. Tokyo-to-JPNIC peering is one hop.
• Residential — NTT fibre (the dominant home-broadband ASN), KDDI au Hikari, SoftBank Hikari. City-level routing in Tokyo 23 wards, Osaka, Kyoto, Fukuoka.
• Mobile — NTT Docomo, KDDI au, SoftBank, Rakuten Mobile 4G/5G.

Tsukuba and the AIST corridor

AIST, RIKEN AIP, and the University of Tsukuba form a research corridor with unusually strong NLP output for the country's size. Collection that targets pre-print servers, institutional repositories, and academic mirrors in Tsukuba benefits from a local exit — some of these sources rate-limit foreign ASNs aggressively.

Good fit

• Training-corpus collection for Japanese under Article 30-4
• Evaluation of multilingual models on Japanese knowledge (J-MMLU, JGLUE-style benchmarks)
• RAG ingestion of Japanese government (e-Gov, Kokkai), academic, and technical literature
• Competitive intelligence against Japanese-market model APIs

Not a good fit

• APPI-sensitive personal data without consent or cross-border safeguards
• Circumvention of publisher opt-outs declared under the Article 30-4 "unreasonable prejudice" exception
• Anything contacting Japanese minors-focused platforms

ASN depth in the Japanese residential pool

Japanese broadband concentrates around the major telecoms plus specific regional carriers:

• NTT East (AS4713) and NTT West (AS2914) — dominant residential FTTH via Flet's. The NTT pool covers the largest share of Japanese residential traffic.
• KDDI (AS2516) — au Hikari fiber + cable, strong urban share.
• SoftBank (AS17676) — BB fiber + cable, second-largest broadband share after NTT.
• J:COM (AS9605) — cable TV-based broadband, specific metros.

Mobile: NTT DoCoMo, KDDI au, SoftBank, Rakuten Mobile carry essentially all JP 4G/5G. Rakuten Mobile's CG-NAT footprint is distinct from the big three, useful for mobile evaluation diversity.

Top Japanese metros

• Tokyo — dominant. Government, business, media (NHK, Asahi, Yomiuri, Nikkei), AWS ap-northeast-1 home region. Default origin for JP-anchored AI workloads.
• Osaka — Kansai region anchor, business + manufacturing. AWS ap-northeast-3 home region (alternative APAC edge).
• Nagoya — Chubu region, industrial (Toyota), distinct regional press (Chunichi Shimbun).
• Fukuoka — Kyushu anchor, growing IT / startup cluster, distinct regional content.
• Sapporo — Hokkaido anchor, northern Japan regional content.

Japanese legal landscape for AI data collection

• APPI (Act on the Protection of Personal Information) — Japan's primary privacy framework. Extraterritorial reach for processing concerning Japanese data subjects. Cross-border transfers to jurisdictions without adequacy decision require specific safeguards.
• Copyright Act Article 30-4 — broad exception permitting computational analysis (explicitly covering ML training) of copyrighted works, UNLESS the use "unreasonably prejudices" the rightsholder's interests. This is one of the most permissive TDM regimes globally, but the "unreasonable prejudice" carve-out has been invoked by some publishers — the legal boundary is evolving.
• Telecom Business Act + MIC guidance — telecommunications operations involving Japanese subscribers have MIC oversight. SquadProxy operates with a JP-registered intermediary for the residential pool.
• Japan AI Promotion Basic Act (2025) — the recent statute framing AI development and deployment. Voluntary guidelines at the moment but the compliance framework is tightening through 2026-2027.

Latency and edge routing

JP edge routes through AWS ap-northeast-1 (Tokyo). Latency:

• OpenAI API (APAC routing): ~15-30ms p50 from Tokyo
• Anthropic API: ~20-40ms p50
• Google Gemini (asia-northeast1): ~10-20ms p50

Residential adds 30-120ms; mobile 50-180ms. NTT fiber footprints have the most consistent latency; SoftBank is slightly more variable under peak.

Related

• Japanese residential proxies for AI
• Japanese 4G mobile proxies
• Japanese 5G mobile proxies
• Multilingual LLM benchmark methodology