DeepSeek

On January 20, 2025, a free app from an obscure Chinese AI lab topped the US App Store, surpassing ChatGPT in downloads. Two days later, NVIDIA's stock fell 17% in a single session — wiping out nearly $600 billion in market cap. The app was DeepSeek. The lab was a 122-person team in Hangzhou that nobody outside China had heard of two weeks earlier. And the model that caused the panic — R1 — had reportedly been trained for $6 million, compared to the hundreds of millions OpenAI spent on GPT-4. Wall Street had been pricing NVIDIA at a premium on the assumption that frontier AI required industrial-scale compute. DeepSeek, quietly, had proved that assumption wrong.

But DeepSeek is not primarily a chatbot company. It is a research lab — a subsidiary of High-Flyer, the $8 billion quantitative hedge fund founded by Liang Wenfeng — that treats AI model development the way physicists treat theoretical research: as a pursuit driven by curiosity about what is fundamentally possible, not by what is immediately monetizable. The chat interface is free. The models are MIT-licensed and open for anyone to use, fine-tune, or deploy commercially without restriction. The API is priced at up to 95% below GPT-5. DeepSeek has taken zero venture capital. Liang Wenfeng retains 84% ownership. The company's operating philosophy is perhaps best captured in its own words: "We want to explore the essence and limit of intelligence." In September 2025, DeepSeek V3.2 shipped with gold-medal performance on the 2025 International Mathematical Olympiad and International Olympiad in Informatics — and dropped API prices by 50% on launch day. In February 2026, Anthropic accused DeepSeek of using fraudulent accounts to scrape Claude conversations for training data, a claim that underscored how seriously the Western AI establishment now takes the threat. The V4 model, expected around Chinese New Year 2026, is rumored to focus on code generation and will reportedly train on NVIDIA's most advanced chips.