Anthropic Opens Seoul Office — What the Korean Enterprise Wave Reveals About Regional Routing Architecture

Anthropic's announcement of its Seoul office is, on the surface, a standard regional expansion story. Under the surface, it contains one of the clearest case studies available of a routing architecture constraint that affects AI teams globally: when regulated-industry enterprises adopt Claude at scale, they cannot simply call the Anthropic API directly. They route through cloud providers — AWS Bedrock, Google Vertex AI — to satisfy in-region data-residency and security requirements.

That constraint is not unique to Korea. It applies to the EU, Japan, healthcare in the US, financial services in Singapore, and most enterprise procurement contexts worldwide. The Korean deployments make it visible because Anthropic named the companies and described the path explicitly.

What happened

On June 17, Anthropic opened a Seoul office and announced a cluster of Korean enterprise deployments:

• Samsung SDS is deploying Claude to employees across Samsung Electronics for knowledge work, agentic workflows, and software development.
• Hanwha Solutions is bringing Claude to global employees through AWS Bedrock, explicitly citing in-region data-residency and security requirements.
• LG CNS (the IT-services arm of LG Group) is rolling out Claude to thousands of employees.
• NAVER — Korea's dominant cloud and AI platform — has deployed Claude Code across its entire engineering organization.
• Nexon is using Claude Code for live-service game development.

Anthropic also signed a Memorandum of Understanding with Korea's Ministry of Science and ICT to collaborate on AI safety and Korean-language safety evaluation with the Korea AI Safety Institute.

Why it matters for AI engineering teams

The Bedrock route is the only compliant route for Korea. Hanwha's explicit call-out — "meeting strict in-region data-residency and security requirements" via AWS Bedrock — is a reminder that the Claude API at api.anthropic.com does not offer Korea-region data residency. Teams operating under Korean data-protection rules, or serving Korean enterprise clients who require domestic processing, must route Claude requests through AWS Bedrock's ap-northeast-2 (Seoul) region.

The same pattern holds across regulated industries globally:

• EU GDPR teams route through Bedrock eu-central-1 (Frankfurt) or Vertex AI EU regions.
• Healthcare teams under HIPAA or regional equivalents route through their contracted cloud provider, not the direct Anthropic API.
• Financial services in Asia-Pacific route through in-region cloud endpoints for both regulatory compliance and audit-trail requirements.

The direct API and the cloud API are not interchangeable. They differ in authentication models, rate limits, pricing, available model versions, and which features are in beta. The Claude Platform on AWS (Anthropic's managed agent runtime on Bedrock) adds another layer: it supports Managed Agents, self-hosted sandboxes, and IAM-based authentication — none of which are available on the direct Anthropic API. Teams choosing their routing architecture must decide which path their compliance requirements actually demand, not just which path is simpler. See the TheRouter docs for how to configure multi-endpoint routing policies.

Model availability is not uniform across paths. Not every Claude model is available in every Bedrock region at launch. When Fable 5 shipped, AWS availability varied by region and lagged the direct API. Teams routing through Bedrock for compliance must monitor Bedrock model availability separately from Anthropic's direct model catalog, and may need a fallback routing policy that handles gaps between model availability on the Anthropic API and the AWS regional endpoint.

Language-specific safety evaluation is becoming a routing signal. Anthropic's MOU with the Korea AI Safety Institute includes evaluating model safety in Korean. As regulators in more countries develop language-specific AI safety benchmarks, routing decisions may eventually need to incorporate which model versions have passed safety evaluations in a given language or jurisdiction. That is not yet an API parameter, but the regulatory direction is clear.

The router/operator angle

For teams running a multi-provider AI gateway, the Korea enterprise wave surfaces three architectural decisions that apply globally:

1. Routing tier by compliance requirement. A team may need the Anthropic direct API for low-latency, non-regulated workloads, and Bedrock for workloads that touch regulated data. Routing policies should encode this split — not just by model name, but by data classification. A gateway that routes all Claude requests to api.anthropic.com without inspecting data-classification headers will fail regulatory requirements without warning.

2. Model availability monitoring per endpoint. If you route to Bedrock for compliance and a newer Claude model is not yet in your required Bedrock region, your fallback chain needs to handle that gap. That means either pinning to an available model version on Bedrock, routing non-regulated traffic to the direct API's newer model, or holding on the older model until regional availability catches up.

3. Credential isolation. Hanwha's Bedrock deployment implies separate credential management: an AWS IAM role with Bedrock permissions, not an Anthropic API key. If your gateway handles both routing paths — direct Anthropic API and Bedrock — you need credential isolation so that API keys do not cross the compliance boundary. Teams using the same gateway configuration for regulated and non-regulated traffic should apply separate credential scopes to each routing tier.

What to watch

The Seoul deployments indicate that Asia-Pacific is the next major growth region for enterprise Claude adoption, following Europe. Operators building routing infrastructure for global enterprise customers should:

• Verify which Claude models are available in your required Bedrock regions today, and build monitoring for when regional availability changes.
• Audit whether your current routing policies enforce data-residency boundaries, or silently route all traffic to the direct API.
• Review the Anthropic-Korea AI Safety Institute partnership for any future language-specific model certification signals that could affect model selection policy for Korean-language workloads.

The direct Claude API is not going away for most teams. But the Korean enterprise rollout is the clearest public statement yet that for regulated industries worldwide, the path to Claude runs through a cloud provider, not a direct API key — and routing infrastructure must treat that boundary as a first-class configuration constraint, not an afterthought.

For teams already running a multi-provider gateway, TheRouter's provider routing supports separate credential scopes per endpoint and per-request routing policies that encode data-classification rules.