Anthropic Says Chinese AI Firms Used 16 Million Claude Queries to Copy Model

Anthropic disclosed that three China-based AI firms—DeepSeek, Moonshot AI and MiniMax—ran industrial-scale distillation campaigns that produced more than 16 million exchanges with Claude via roughly 24,000 fraudulent accounts to harvest its capabilities. DeepSeek focused on reasoning and censorship-safe paraphrases (~150,000 exchanges), Moonshot on agentic reasoning, tool use, coding and vision (~3.4m exchanges), and MiniMax on agentic coding and tool use (~13m exchanges); Anthropic attributed the traffic via metadata, IP correlation and infrastructure signals and says the attacks used commercial proxy “hydra” clusters. The company warned illicitly distilled models strip safety safeguards and pose national-security risks, has implemented classifiers and tightened verifications, and the disclosure heightens regulatory, competitive and geopolitical risks for frontier AI developers.

Analysis

Market structure: The disclosures strengthen moats for well-capitalized frontier labs and cloud providers (GOOGL, AMZN, MSFT) that can invest in detection, verification, and IP defenses; expect 3–7% incremental pricing power for premium model/API tiers over 6–18 months as providers tighten access. Smaller AI startups and proxy/reseller businesses are losers—their access arbitrage and low-cost model cloning economics are permanently impaired, increasing concentration risk in top providers. Cross-asset: modest upward pressure on tech credit spreads if litigation/regulatory costs rise (+10–30bp tail), slight rise in implied vol for GOOGL/AI names near-term, and potential USD safe-haven flows if US-China tech decoupling escalates.

Risk assessment: Tail risks include US export-controls or criminal referrals to China-based firms, which could trigger >20% repricing for China AI equities (days–weeks). Immediate (0–14 days): headline-driven volatility; short-term (1–3 months): regulatory inquiries and vendor delistings; long-term (6–24 months): market consolidation and higher CAPEX for safety. Hidden dependencies: model security relies on cloud providers, KYC vendors, and proxy detection; failure in any link can amplify breaches. Catalysts: additional disclosures from Google/DOJ actions or EU AI Act enforcement within 30–90 days could accelerate repricing.

Key Decisions for Investors

• Establish a 0.75–1.5% long position in GOOGL (GOOGL/GOOG) within 2 weeks; set an 8% stop-loss and a 6–12 month target of +15% if regulatory/competitive moat strengthens.
• Initiate a pair trade: long GOOGL 1% / short BIDU 1% to capture US incumbents' defensible IP premium vs Chinese regulatory/geopolitical risk; reassess after 90 days or on any major policy announcement.
• Allocate 1% each to CrowdStrike (CRWD) and Palo Alto Networks (PANW) as 6–12 month beneficiaries of higher security spend; trim if either rallies >20% or if detection effectiveness materially declines.
• Buy a 3-month GOOGL 10% OTM call spread (size = 0.25–0.5% portfolio risk) to capture asymmetric upside from further enforcement/market consolidation; close at 50% of max profit or if implied vol spikes >40%.
• Reduce small-cap/public AI model vendor exposure by 20–30% vs benchmark immediately; re-evaluate after 60–90 days contingent on regulatory clarity and any DOJ/Commerce actions announced within that window.