China’s DeepSeek Unravels America’s AI Industry

A Chinese AI company called DeepSeek has reached the capabilities of OpenAI, Meta, Google, and Anthropic for a fraction of the cost and computing power — putting those other companies’ call for nearly unlimited funding and resources to compete with China into question.

Why It Hurts: The success of DeepSeek’s models is sending shockwaves across the AI industry and greater financial markets, including a trillion-dollar stock route among top AI companies, a 17% drop in Nvidia’s stock ($589 billion in lost value, the most ever for a company), and a 3.2% drop in the Nasdaq.

Behind the Code: DeepSeek, created by the Liang Wenfeng-led hedge fund High-Flyer Capital Management to optimize its trading, has stunned industry analysts and investors this past week.

• DeepSeek’s reasoning models, R1 and V3, reportedly perform almost as well as Meta’s Llama, OpenAI’s ChatGPT and o1, Google’s Gemini, and Anthropic’s Claude.
• With V3, it accomplished the feat by spending $5.6 million and using just a bit more than 2,000 older-generation Nvidia chips — far fewer resources than the hundreds of millions of dollars and tens of thousands of chips used to train leading AI models.
• DeepSeek partly accomplished this by having its models train on the outputs of other AI models, skipping once-thought-to-be essential steps for developing systems (and raising interesting questions about the reliability and dangers of having AI learn from other AIs).
• All the hype has led DeepSeek’s free consumer app, which runs as an open-source system, to rocket to the top of the chatbot App Store chart, besting ChatGPT.
• And to add more chaos to the mix, DeepSeek also released two multimodal AI tools, Janus-Pro-1B and Janus-Pro-7B, that are allegedly as good as the latest versions of DALL-E and Stable Diffusion.

The Future: a16z’s Marc Andreessen called DeepSeek’s models “one of the most amazing and impressive breakthroughs I’ve ever seen,” while Foundry CEO Jared Quincy Davis likened it to “as if someone released a $20 iPhone.” Companies are already announcing that they’ve switched over to DeepSeek to cut down on costs. The adoption is happening despite DeepSeek’s outputs adhering to Beijing’s censorship rules — a lot to unpack there.

All of this is very bad news for Meta. Mark Zuckerberg has positioned Llama to be the world’s most dominant open-source model and is spending boocoo bucks to make that dream a reality. DeepSeek threatens to render that ambition to the same fate as Meta’s metaverse hopes (RIP). Unsurprisingly, The Information reports that Meta has already set up at least four “war rooms” to break down DeepSeek’s models and recreate Llama in its image — a copy of a copy.