Anthropic just caught China copying its homework (16 million times)

On Monday, Anthropic published a blog post accusing DeepSeek, Moonshot AI, and MiniMax of running 24,000 fake accounts against Claude. Over 16 million exchanges. Targeted extraction of Claude's best capabilities: agentic reasoning, tool use, and coding.

The same day, a 33-year-old analyst published a fictional memo on Substack imagining an AI-driven economic collapse. The Dow dropped 800 points.

Then, on Tuesday, Meta signed a $60 billion chip deal with AMD. Nvidia reports earnings today. It was that kind of week.

🔓 The Big Story: Anthropic Catches Three Chinese Labs Copying Claude at Industrial Scale

Anthropic published a blog post accusing DeepSeek, Moonshot AI, and MiniMax of running 24,000 fake accounts against Claude - over 16 million exchanges. Targeted extraction of Claude's best capabilities: agentic reasoning, tool use, and coding.

The technique is called distillation: bombard a capable model with carefully designed prompts, collect the outputs, use those outputs to train your own system.

How the three campaigns broke down: - DeepSeek: 150,000+ exchanges targeting reasoning, asking Claude to produce "censorship-safe alternatives to politically sensitive queries" - Moonshot AI: 3.4 million exchanges focused on agentic reasoning, coding, computer-use agent development, and computer vision - MiniMax: 13 million exchanges targeting agentic coding and tool orchestration

Anthropic identified the actors through IP address correlation, request metadata, and corroboration from other AI companies that spotted the same behavior. The labs used proxy networks to mask their locations.

The timing is deliberate. This announcement lands as Washington debates tighter AI chip export controls. Anthropic's blog explicitly ties distillation to national security.

The uncomfortable reality: The entire foundation model industry is built on extracting value from other people's work. Anthropic settled a copyright lawsuit last September for $1.5 billion over allegations that it trained Claude on books downloaded from shadow libraries. The line between "legitimate training" and "illicit distillation" is blurrier than anyone in San Francisco wants to admit.

💰 Meta Bets $60 Billion on AMD. Nvidia Reports Today.

Meta and AMD announced a five-year deal worth up to $60 billion for AI chips. AMD will deliver six gigawatts of compute power starting H2 2026, built around a custom GPU based on its MI450 architecture. Meta also gets the option to buy up to 10% of AMD through performance-based warrants at $0.01 per share. AMD stock jumped 9%.

The broader picture: The five largest US cloud and AI infrastructure companies have committed to over $650 billion in capital expenditure for 2026 - nearly double 2025. Meta alone plans $135 billion this year.

Nvidia reports Q4 FY2026 earnings. Wall Street expects ~$66 billion in revenue (up 68% YoY). Jensen Huang said Nvidia has $500B in combined Blackwell and Rubin chip visibility through end of calendar 2026.

📉 A Fictional Substack Post Crashed the Stock Market

Citrini Research published "The 2028 Global Intelligence Crisis" - a 5,000-word thought experiment written as a dispatch from June 2028, looking back on how AI agents destroyed the economy. Unemployment above 10%. S&P 500 down 38%.

The Dow dropped 800 points on Monday. The iShares Expanded Tech-Software Sector ETF hit a 52-week low, down nearly 30% YTD.

Why it resonated: The SaaSpocalypse from three weeks ago was still fresh. Citrini's scenario didn't invent a new fear. It organized the fear that was already there into a coherent narrative.

🎯 Quick Hits

• Google shipped Gemini 3.1 Pro (Feb 19) - More than doubled reasoning performance, scoring 77.1% on ARC-AGI-2. Available in Gemini app, NotebookLM, Vertex AI, and GitHub Copilot.
• ChatGPT ads are live - Started Feb 9 for US users on Free and Go tiers. $60 CPM, $200K minimum advertiser commitment. Target, Adobe, Expedia, Qualcomm among first brands.
• Waymo now operating in 10 US cities
• DeepSeek reportedly trained on Nvidia's Blackwell chips - Apparent violation of US export controls (Reuters)
• AI energy crunch is real - Power constraints reshaping where AI companies can build and scale. Meta's 6-gigawatt AMD deal is roughly the output of six nuclear reactors.

💭 One Thing I'm Thinking About

The underlying tension around training data isn't going away. The value in these models comes from aggregating human intelligence at massive scale. The question of who gets to benefit from that aggregation, and who doesn't, is going to define the next decade of AI policy.

Right now, the answer seems to be: whoever has the best lawyers and the most sympathetic geopolitical framing.