Seoul exits for Korean corpus and HyperCLOVA/EXAONE evaluation

Why Korean-language collection is harder than it looks

Korean content is materially under-represented in commodity training corpora relative to the market its model ecosystem targets. Naver publicly reports that HyperCLOVA X was trained on substantially more Korean-language data than typical Western frontier models — the company has cited a figure several orders of magnitude greater than GPT-4 on Korean corpus specifically. LG AI Research's EXAONE 4.0, released July 2025, is positioned around enterprise and reasoning workloads with Korean-domain depth.

For an AI team that is not Naver or LG and wants to build competitive Korean capability, the collection problem has three angles:

• Publisher geoblocking. Major Korean press (Chosun, Joongang, Hankyoreh) and portal-native content (Naver, Daum/Kakao) geoblock non-KR ASNs aggressively or rate-limit them into irrelevance.
• Academic/government corpora. KISS, DBpia, RISS, data.go.kr — strongly biased toward KR-resident access.
• Eval authenticity. Running Korean-language evaluation against global model APIs from a non-KR IP under-samples the Korean- context adjustments these models make (notably, content policy localization).

What the KR pool routes through

• Datacenter — AWS ap-northeast-2 (Seoul), GCP asia-northeast3, Naver Cloud Platform (Gasan, Chuncheon). KINX peering in Gasan.
• Residential — KT, SK Broadband, LG U+, with district-level routing in Seoul (Gangnam, Mapo, Jongno), Busan, and Daejeon.
• Mobile — SK Telecom, KT, LG U+ 4G/5G.

Good fit

• Korean-language training corpus collection under Article 35-5
• Evaluation against HyperCLOVA X, EXAONE, and multilingual models from a Seoul-resident vantage
• RAG ingestion of Korean government and academic open-data
• Competitive AI intelligence in the Korean market

Not a good fit

• PIPA-sensitive PII collection outside lawful basis
• Scraping of content that would trigger Article 35-5's "unreasonable prejudice" carve-out
• Targeting Korean minors-focused platforms or educational platforms without clear consent chain

ASN depth in the Korean residential pool

Korean broadband concentrates around three telecoms on near-universal Gigabit FTTH:

• KT (AS4766) — largest residential FTTH, legacy Korea Telecom footprint.
• SK Broadband (AS9318) — SK Telecom's broadband arm, strong urban presence.
• LG U+ (AS17858) — third of the three, particularly strong in specific metro areas.

Mobile: SK Telecom (AS9644), KT, LG U+ carry all 4G/5G. Korea has among the most advanced 5G deployments globally with ~20-30% 5G SA subscriber share as of 2025.

Why the Korean pool matters for AI workloads

• High-fidelity 5G mobile eval — Korean 5G networks are the most production-hardened globally. Mobile-anchored AI evaluation that wants realistic advanced-5G conditions anchors here.
• Korean-language training corpus — low-resource language in most global training corpora. Korean residentials from Seoul and Busan are essential for KR-language AI work that needs authentic regional content.
• Regional eval for APAC audience — Korea's cultural and linguistic distinctness from Japan and China makes it a required origin for pan-APAC model evaluation.

Top Korean metros

• Seoul — dominant. Government, business, media, AI research (Samsung, LG, Naver, Kakao headquarters). AWS ap-northeast-2 home region (Seoul).
• Busan — southern port city, distinctive regional character.
• Incheon — gateway metro, Incheon International Airport, dense logistics and tech industry.
• Daegu, Gwangju — regional anchors with useful content diversity.

Korean legal landscape

• PIPA (Personal Information Protection Act) — Korea's primary privacy framework. PIPC is the supervisory authority. One of the most stringent privacy regimes in APAC. Explicit consent is the default lawful basis for many types of data processing, including training-data collection.
• Copyright Act Article 35-5 (TDM exception) — permits data analysis (including ML training) subject to an "unreasonable prejudice" carve-out similar to the Japanese framework. Korean publishers are using this more aggressively than Japanese counterparts.
• AI Basic Act — Korea passed an AI framework law in 2024-2025 with tiered risk obligations. Foundation-model providers face disclosure and audit obligations particularly around training data.
• Network Act — telecom and online-service obligations, including cross-border data transfer controls that can apply to AI workloads.

Latency and edge routing

KR edge routes through AWS ap-northeast-2 (Seoul). Latency:

• OpenAI API (APAC): ~10-25ms p50 from Seoul
• Anthropic API: ~15-30ms p50
• Google Gemini (asia-northeast3): ~5-15ms p50

Residential adds 30-120ms; mobile 50-180ms. Korean fiber residentials have among the most consistent latency globally.

Related

• Korean residential proxies for AI
• Korean 5G mobile proxies for cellular AI eval
• Multilingual benchmark methodology