Nvidia has the worst wipeout in history as China’s DeepSeek upends the AI hierarchy

Deep sink… Chinese AI startup DeepSeek blew a $1T+ hole in the stock market yesterday. AI’s “Sputnik moment” saw Nvidia wipe out ~$600B in market cap — the largest one-day loss is US history. DeepSeek’s free-to-use model, launched last week, is said to match the performance of OpenAI’s most advanced GPT offering. Salting the wound: DeepSeek said its model was trained for less than $6M (OpenAI spent roughly 20x that to train its GPT-4). Yesterday, DeepSeek dethroned ChatGPT as the top free app on Apple’s App Store.

• Reality check: DeepSeek built a serious AI contender despite using fewer (and less-advanced) chips. In the face of US chip-export restrictions, the Chinese co said it used ~2K reduced-capability Nvidia chips to train its model — a fraction of the tens of thousands of chips used by rival models.

• Skeptics: The CEO of ScaleAI suggested DeepSeek has 50K advanced Nvidia chips that it can’t admit to because of US export controls. Still, it’s a sobering wakeup call about China’s AI capabilities, which have advanced quicker than previously thought. 

• Reviews: Users have lauded DeepSeek’s transparency (the app shows its work with every answer) but criticized its censorship (think: refusing to answer q’s about the Tiananmen Square protests and massacre). It also sends US data to China.

“Did we just waste billions?”… DeepSeek’s success-for-less strategy is making investors shudder at the gargantuan amounts US tech companies have poured into AI. Last quarter alone, Amazon, Alphabet, Microsoft, and Meta spent $60B in capital expenditures, largely on AI investments like data centers. Meta recently said it’ll make $65B in capital expenditures this year as it executes plans for a city-sized AI data center. Microsoft committed $80B to AI data centers this year, and the AI infrastructure investment by OpenAI’s Stargate joint venture could grow from $100B to $500B+.