Executive Summary

Anthropic has accused three Chinese AI labs – DeepSeek, Moonshot, and Minimax – of extracting Claude's advanced capabilities through over 16 million requests via 24,000 fake accounts. The "distillation" technique enables competitors to copy AI capacities at a fraction of development costs. OpenAI and Google raise similar allegations. Minimax has risen from the fourth-largest to the largest AI provider on OpenRouter – Anthropic lost its market share from 40% to 12% in one year. The matter affects not just Silicon Valley but jeopardizes US economic dominance.

People

Topics

  • AI Security and Data Theft
  • Geopolitical Competition
  • Artificial Intelligence Models
  • US-China Tech Rivalry

Clarus Lead

Three Chinese AI labs are illegally distilling capabilities from Claude at massive scale. Anthropic documents 16 million unauthorized API accesses through proxy networks – a coordinated campaign with geopolitical consequences. Why it matters: Minimax is now the market leader in developer APIs; Anthropic lost billions in revenue. This threatens not individual startups, but the entire US economy, which is built on proprietary AI models. The core conflict: Chip export controls don't work – distillation is faster, cheaper, and difficult to prevent.

Detailed Summary

Scope of Data Theft

Anthropic released evidence on February 23 of industrial theft campaigns. DeepSeek, Moonshot, and Minimax operated proxy networks that simultaneously managed over 20,000 fake accounts. The goal was to extract reasoning capabilities – not just answers, but the thinking steps behind them. The labs prompted Claude to reveal step-by-step solution paths to train new models with them. Distillation is legal when a company shrinks its own models. It becomes piracy when competitors copy others' outputs. The parallel: Like illegally recording Netflix movies, altering them for CGI, and selling them under your own label – except it happens in seconds across millions of prompts.

Economic Consequences for US Companies

OpenRouter, the largest neutral API platform, shows dramatic market shifts. A year ago, Anthropic controlled 40% of token usage. As of February 2026: only 12%. Minimax jumped to 20% – larger than Google and OpenAI combined. For Anthropic, this means (on OpenRouter alone) a loss of 750 million to one billion dollars annually. Add to this direct customer migration. Anthropic is preparing an IPO; growth has stalled – this data extraction hits the company at a critical moment.

Security and Geopolitics Dimension

Distilled models often lack security guardrails that US labs integrate. This enables potentially military and surveillance applications without controls. The FBI and US Trade Representative estimate annual IP theft by China at 225–600 billion dollars. Over two decades: four to twelve trillion dollars in asset outflow. Chip exports were restricted – but distillation is the faster hole in the fence. While the US imposes expensive GPU controls, Chinese labs train for free on US outputs.

Core Findings

  • Three Chinese labs generate 16 million illegal prompts through 24,000 fake accounts to distill Claude
  • Minimax becomes #1 provider on OpenRouter; Anthropic falls from 40% to 12% market share
  • Economic damage to US tech: 750 million–several billion dollars annually (verifiable portion only)
  • Geopolitical risk: Distillation bypasses chip export controls faster and cheaper
  • Security gap: Stolen models lack safeguards against military misuse

Critical Questions

  1. Data Quality: Anthropic documents the 16 million prompts with IP and metadata analysis – but how complete is this tracking? Could legitimate usage patterns (VPNs, regional access) be falsely classified as "distillation"?

  2. Own Responsibility: Anthropic paid 1.5 billion dollars for copyright violations (Libgen data in training). What guarantee is there that Anthropic didn't obtain its own models similarly – and thus holds the moral standing to accuse others?

  3. Causality in Minimax Success: Minimax demonstrably uses distilled capabilities – but can the market share jump be explained by this alone? Could better cost optimization, user experience, or regional demand be equally responsible?

  4. Feasibility of Countermeasures: Watermarking and rate-limiting don't work at scale. What technology could practically prevent distillation without destroying API usability?

  5. Geopolitical Consequences: If the US sanctions Chinese labs – couldn't they simply switch to open-source models (already available) and achieve AI progress at the same speed?

  6. Business Model Risk: Is offering frontier AI APIs over public internet infrastructure structurally vulnerable to distillation – and should US labs shift to local licensing and security audits?

Sources

Primary Source: [Everyday AI Show – Episode 720: China Stealing AI from the U.S.] – https://pscrb.fm/rss/p/www.buzzsprout.com/2175779/episodes/18739212-ep-720-china-stealing-ai-from-the-u-s-inside-anthropic-s-bombshell-allegations.mp3

Verification Status: ✓ 2026-02-24

This text was created with the support of an AI model.
Editorial Responsibility: clarus.news | Fact-Check: 2026-02-24