On Monday, January 27, 2025, a free app most Americans had never heard of a week earlier sat at #1 on the US App Store — and by the closing bell, chipmaker Nvidia had shed roughly $600 billion in market value in a single day, the largest one-day loss for any company in US stock-market history. The app was DeepSeek, a Chinese AI chatbot, and the story of how it briefly “broke” the most valuable chip company on Earth is one of the most revealing episodes in the whole AI era.
It’s also widely misunderstood. So let’s walk through the day DeepSeek broke Nvidia carefully: what actually happened, why a chatbot could crater a chip stock, and what the panic did — and didn’t — prove.
A Three-Week Earthquake
The shock didn’t come from nowhere. It built over about a month:
- December 26, 2024: DeepSeek quietly released V3, a powerful open-weights AI model, reporting a final-training-run cost of around $5.6 million — a figure that would soon go viral (and which, as we’ll see, needs a big asterisk).
- January 20, 2025: DeepSeek released R1, a reasoning model that benchmarked competitively with OpenAI’s best — and released it under an open MIT license, free for anyone to download and use.
- January 27, 2025: the DeepSeek app overtook ChatGPT as the #1 free app on the US iOS App Store. The same trading day, Nvidia’s stock fell sharply, and the AI-infrastructure trade unwound across the market.
How big was the drop? Here it’s worth being precise, because you’ll see different numbers. The most-cited figure is an 18% fall and about $600 billion in lost market value. Other outlets reported roughly 17% and $589 billion. That’s not a contradiction — it’s the same trading day measured slightly differently (intraday low versus close). Either way, it was a record decline for a single company — not a broad-market crash, but a staggering hit to one firm.
Why Would a Chatbot Crash a Chip Stock?
This is the part that confuses people. DeepSeek makes AI software; Nvidia makes the chips AI runs on. Why would one hurt the other?
The answer is a chain of assumptions the market had been leaning on. For two years, the story powering Nvidia’s historic rise was simple: cutting-edge AI requires enormous quantities of Nvidia’s most expensive chips. Every big lab was racing to buy more GPUs, and investors priced Nvidia as the indispensable supplier to that gold rush.
DeepSeek appeared to threaten that story on two fronts at once:
- It looked cheap to train. If a frontier-class model could be built for a tiny fraction of the assumed cost, maybe the world didn’t need to keep buying ever-more chips at any price.
- It ran on weaker chips. DeepSeek trained on Nvidia’s H800 — a deliberately downgraded model built to comply with US export restrictions to China, not the top-tier hardware US labs used. If world-class results were possible on the cut-down chip, the premium on the best chips looked shakier.
Put those together and nervous investors reached a scary conclusion: maybe AI won’t need as many top-end Nvidia GPUs as we thought. That fear — not any actual drop in chip demand — is what sold the stock. Marc Andreessen captured the mood by calling R1 AI’s “Sputnik moment”: the jolt of realizing a rival had leapt ahead where you assumed you were safely in front.
What the Panic Got Right — and Wrong
Here’s the crucial nuance: the crash was a reaction to an assumption, and the assumption was only half true.
The “$5.6 million” number was real but misleading — it covered only DeepSeek’s final training run, not the years of research, the failed experiments, or the billion-dollar-plus GPU cluster behind it. (We break the whole figure down in “Did DeepSeek Really Cost $6 Million?”.) DeepSeek’s parent company had also stockpiled an estimated 10,000 Nvidia A100 chips before export controls tightened — so “trained on weak chips alone” was never the full picture either.
And the market itself second-guessed the panic: Nvidia’s stock, though rattled, substantially recovered in the following weeks as investors digested that efficient models can actually increase total AI demand, not shrink it (cheaper AI means more people use more of it). The one-day record stood, but the “AI is over-built” thesis didn’t hold. (None of this is investment advice — it’s a look at what moved and why.)
What It Actually Proved
Strip away the market drama and a genuine, lasting lesson remains. DeepSeek showed that a team outside the small circle of US frontier labs could ship a reasoning-class, open-weights model that competed with the best closed systems — built efficiently, and under a real hardware handicap. That combination punctured a comfortable assumption: that frontier AI required both an enormous closed budget and unrestricted access to the very best chips.
It also reopened a real debate with a concrete data point instead of a hypothesis. Did US chip-export controls fail — pushing Chinese labs to innovate around the limits with clever engineering like Mixture of Experts and 8-bit training? Or did they still constrain China’s ceiling, given that pre-restriction chip stockpile? Reasonable people still argue both sides. What’s not in dispute is that one open-weights release from a former hedge fund rearranged how the world thinks about the cost, openness, and geography of cutting-edge AI — in a single day.
That’s why “the day DeepSeek broke Nvidia” is remembered less as a stock story than as the moment the AI race stopped looking like a closed, two-or-three-company contest — and started looking like anyone’s game.
❓ Frequently Asked Questions
How much did Nvidia lose the day DeepSeek launched? On January 27, 2025, Nvidia’s stock fell about 18%, erasing roughly $600 billion in market value in one day — the largest single-day loss for any company in US stock-market history. Some outlets reported slightly different figures (about 17% and $589 billion) based on intraday-versus-close measurement, but it was a record decline for a single company either way.
Why did DeepSeek cause Nvidia’s stock to crash? DeepSeek appeared to show that a top-tier AI model could be trained relatively cheaply and on downgraded export-restricted chips. That threatened the market’s core assumption that cutting-edge AI required buying huge amounts of Nvidia’s most expensive GPUs — so investors sold on the fear that future chip demand might be lower than expected.
Did Nvidia recover from the DeepSeek crash? Yes. Although the one-day drop set a record, Nvidia’s stock substantially recovered in the following weeks as investors reconsidered — recognizing that cheaper, more efficient AI can actually increase overall demand for AI (and chips) rather than reduce it.
Was the “$5.6 million” DeepSeek cost real? It was DeepSeek’s own figure — but only for the final training run. It excluded years of research, failed experiments, staff, and the billion-dollar-plus GPU cluster behind the model. The viral “trained for $6 million” framing left all of that out. Full breakdown here.
What was the “Sputnik moment” comparison? Venture capitalist Marc Andreessen and others compared DeepSeek-R1’s release to the 1957 Sputnik launch — the shock of realizing a rival had suddenly leapt ahead in a race you assumed you were leading. It captured how the US tech industry felt when an open-weights Chinese model matched its best closed systems.
📚 Keep Learning
- Did DeepSeek Really Cost $6 Million? — the viral number, fully unpacked.
- Learn AI and ML Concepts Through the Lens of DeepSeek — what made the model efficient enough to spook the market.
- Is DeepSeek Safe to Use? — the privacy and security questions that followed the hype.
📚 Recommended Resources
Disclosure: some links below are Amazon affiliate links. If you buy through them, I earn a small commission — at no extra cost to you — which helps keep this site running.
Artificial Intelligence: A Guide for Thinking Humans
Melanie Mitchell · Book · ~$28 on Amazon
The best antidote to AI hype cycles like this one. Mitchell teaches you to tell a genuine breakthrough from a market panic — exactly the judgment you needed to read the DeepSeek story clearly while the headlines were losing their minds. No math, just clear thinking.
Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow
Aurélien Géron · Book · ~$70 on Amazon
If the DeepSeek shock made you want to actually understand the technology under all the drama — what “training a model” really costs and involves — this is the clearest, most practical place to build that understanding, from the fundamentals up.