The DeepSeek effect: AI stock prices dropped $1 trillion in a day. The data.

January 27, 2025. The single biggest one-day market cap drop for any company in history.

NVIDIA lost $589 billion in market value in a single trading session. Not because of a product failure. Not because of an earnings miss. Because a Chinese AI lab published a paper showing frontier models don't need as much compute as everyone assumed.

I've been tracking the financial side of AI for four years. Nothing in my data looks like this.

The damage report

| Company | Market cap loss (Jan 27) | Percentage drop | Prior AI narrative | |---------|--------------------------|----------------|--------------------| | NVIDIA | -$589B | -17.0% | "Picks and shovels" winner | | Broadcom | -$68B | -17.4% | AI networking chips | | ASML | -$46B | -7.1% | EUV lithography monopoly | | Taiwan Semiconductor | -$64B | -5.3% | Chip fabrication | | AMD | -$23B | -6.4% | NVIDIA alternative | | Arm Holdings | -$19B | -10.3% | Chip IP licensing | | Constellation Energy | -$14B | -20.8% | AI data center power | | Vertiv | -$14B | -15.2% | Data center cooling |

Sources: Yahoo Finance, Bloomberg, Reuters, market close data January 27, 2025.

Add it up. The companies most associated with "AI needs massive compute" lost over $800 billion in value. In one day.

Constellation Energy, a nuclear power company, lost 20.8%. Their AI story was: "data centers need enormous power, nuclear is the answer." DeepSeek's efficient training report punched a hole in that thesis.

Why DeepSeek triggered this

The market had priced in a specific theory: frontier AI requires exponentially more compute, which means exponentially more GPUs, which means NVIDIA wins for a decade.

DeepSeek broke this theory with three data points:

| Assumption | Market belief | DeepSeek's data | |-----------|---------------|-----------------| | Training cost for frontier models | $100M+ (growing) | $5.6M (shrinking) | | GPU count required | 10,000+ H100s | 2,048 H800s | | Compute scaling | More compute = better models | Efficient training = better models |

Sources: DeepSeek V3 and R1 technical reports.

When R1 matched o1-preview's reasoning benchmarks at 1/40th the inference cost, the market had to reprice the entire "compute is destiny" narrative.

The recovery (and what it tells us)

Here's what happened in the week after:

| Company | Jan 27 close | Jan 31 close | Recovery | |---------|-------------|-------------|----------| | NVIDIA | -17.0% | -12.3% | Partial | | Broadcom | -17.4% | -11.6% | Partial | | ASML | -7.1% | -4.2% | Mostly recovered | | TSMC | -5.3% | -2.1% | Mostly recovered |

Sources: Yahoo Finance, weekly close data.

NVIDIA recovered about half the loss by week's end. TSMC and ASML recovered most of it. The market decided: yes, efficiency matters, but AI still needs a lot of chips. The narrative shifted from "infinite compute growth" to "still-very-large compute growth, but with better efficiency."

That's a meaningful repricing. Not a collapse. A recalibration.

The Jevons Paradox argument

NVIDIA CEO Jensen Huang's response was essentially the Jevons Paradox: when something becomes cheaper, people use more of it, not less. Cheaper inference means more AI applications, more tokens consumed, more GPUs needed in aggregate.

There's data to support this. After GPT-4o mini launched at 100x cheaper than GPT-4 in July 2024, OpenAI reported total API usage went up, not down.

| Period | Event | Price change | Total API revenue | |--------|-------|-------------|-------------------| | Jul 2024 | GPT-4o mini launch | -100x per token | Revenue up 17% | | Nov 2024 | GPT-4 Turbo price cut | -3x per token | Revenue up 12% | | Jan 2025 | DeepSeek R1 | -40x reasoning cost | TBD |

Sources: OpenAI earnings commentary, Bloomberg analysis.

The Jevons argument is valid. But it's also an argument against NVIDIA's margins, not its revenue. If the market needs 40x fewer GPU-hours per reasoning task, the demand curve needs to shift right by 40x just to break even. That's possible over 5-10 years. Over 1-2 years? I'm skeptical.

My read on this

I've been tracking AI industry data since 2021. I've seen hype cycles, corrections, and genuine breakthroughs. This one is different.

The DeepSeek effect isn't about one company or one model. It's about the discovery that training efficiency innovations (MoE, FP8 training, multi-token prediction, GRPO) can substitute for raw compute. That changes the economics of the entire industry.

Before January 27, the consensus was: "to be competitive in AI, you need a $10B compute cluster." After January 27, the consensus shifted to: "you need great researchers and smart training recipes, plus a $500M compute cluster."

Both are expensive. But the second one describes a much larger pool of potential competitors.

My spreadsheet for tracking "dollars of market cap destroyed per arXiv paper" has a new all-time record. $800 billion per paper. I don't think we'll see that topped.

Then again, I've been wrong about ceilings before.

---

If you found this interesting, you might also like:

• The GPT-3 API waitlist is 6 months long. Here's what the early data looks like.
• I counted every AI startup that raised money in Q1 2021. The numbers are strange.
• Open source AI is having a moment. Here are the download numbers.
• Anthropic just raised $580M. Let's talk about the AI safety funding numbers.
• Wait, Stable Diffusion has HOW many forks? The open source explosion in numbers.

-- dataku