I spent a week researching the Chinese “transfer station” economy reselling Claude at 10% of retail. The supply chain is wilder than I expected.

Exploring the Intricate World of China’s “Transfer Station” Market for AI API Reselling

Over the past week, I undertook an in-depth investigation into a niche yet increasingly influential market segment in China—referred to locally as “转运站” or “transfer stations.” These operations are centered around reselling access to advanced AI APIs, notably Claude, at a fraction—around 10%—of the official retail price from providers like Anthropic. What I uncovered reveals a supply chain far more sophisticated and layered than surface-level observations suggest. This article aims to demystify the complex mechanics underpinning this ecosystem.

The Multi-Layered Supply Chain Architecture

At its core, the transfer station economy operates through an elaborate eight-layer modular supply chain:

Account Generation and Management:
“Account farmers” leverage sophisticated automation tools to produce thousands of Anthropic accounts. They utilize antidetect browsers, including Multilogin, AdsPower, and GoLogin, paired with residential proxies to simulate genuine user devices.
Network traffic is meticulously crafted using curl_cffi to mimic Chrome’s TLS fingerprints at the network layer, bypassing basic detections.
Phone Number Verification Bypasses:
SMS verification requirements are circumvented via SMS-Activate-class APIs, which interface with physical SIM bank hardware—namely Hybertone GoIP servers housing hundreds of real SIM cards—allowing rapid, bulk number provisioning.
KYC (Know Your Customer) Circumvention:
Since regulations now mandate government ID verification and live selfies (effective from April 2026), these operations employ multiple strategies:
- AI-generated, fake IDs produced by services like OnlyFake-class.
- Real-time deepfake injection techniques, utilizing virtual cameras through software such as OBS Virtual Camera with DeepFaceLive or Deep-Live-Cam.
- Human-in-the-loop farms recruiting individuals from low-income regions to authenticate identities effectively.
Relay Infrastructure and API Access Management:
The relays are built on a foundation of open-source repositories, including one-api, new-api, claude-relay-service, claude2api, clewdr, and clove.
These relay networks pool OAuth tokens—identified by patterns like sk-ant-oat01-... and sk-ant-ort01-...—and rotate them across requests, allowing thousands of users to access the APIs via shared farmed accounts.

Technical Evasion and Quality Control Measures

An eye-opening CISPA Helmholtz audit examined 17 such relay networks and uncovered noteworthy performance issues:

Performance Degradation:
Up to 47.21% reduction compared to direct access to the official API.
Many relays route traffic through alternative models such as Haiku, GLM, or Qwen, relabeling responses to mimic genuine outputs.
Model Fingerprint Spoofing:
Nearly 46% of audited endpoints failed model-fingerprint verification, indicating attempts to mask the true origin of responses.
Data Logging & Harvesting:
All interactions—prompts and responses—are logged meticulously.
As disclosed by Anthropic in February 2026, a network comprising over 20,000 accounts has accumulated approximately 16 million exchanges, fueling secondary datasets.
These datasets, like DeepSeek and Moonshot, are openly shared on platforms such as HuggingFace, further enriching the publicly available AI training material.

Underlying Tools and Detection Evasion Techniques

The ecosystem leverages an array of sophisticated tools:

Reverse Engineering & Traffic Analysis:
Reverse-engineered traffic data, notably via eunomia.dev’s eBPF tools, decrypts how Claude’s traffic flows and detects anomalies.
Understanding the internal detection systems, including Anthropic’s Clio, reveals vulnerabilities—especially in cross-account monitoring—that these relays exploit.
OAuth & Traffic Modulation:
By reverse engineering OAuth flows and employing techniques like JA3/JA4 fingerprint evasion, relays effectively hide their identity and avoid detection.

Monetization Strategies and Market Dynamics

The “one fish, three meals” approach exemplifies how these relay networks maximize profitability. They generate revenue by:

Offering API access at a fraction of official prices.
Selling harvested user interactions and datasets for training or malicious purposes.
Exploiting weaknesses in detection systems to operate undetected long-term.

Broader Implications and Ethical Considerations

This ecosystem’s complexity exposes significant challenges around API security, user verification, and data privacy. It underscores the necessity for providers like Anthropic to bolster detection mechanisms and tighten access controls. Furthermore, the proliferation of such reselling markets accentuates the importance of understanding supply chain vulnerabilities in the AI ecosystem.

Conclusion

The Chinese “transfer station” market for reselling AI API access demonstrates a high degree of technical ingenuity and operational sophistication. From account creation and verification to traffic routing and data harvesting, these operations form an underground economy that challenges conventional security paradigms. As AI models become more embedded in society, awareness of these dynamics becomes crucial for developers, security professionals, and policymakers alike.

Sources Consulted:

ChinaTalk article by Zilan Qian (May 2026)
CISPA Helmholtz paper Real Money, Fake Models (arXiv 2603.01919)
Anthropic’s February 2026 dataset disclosure
eunomia.dev’s eBPF reverse-engineering analyses
Public repositories of relay software on GitHub

Stay tuned for further insights into AI supply chains and security mechanisms.