The DeepSeek shock, one year on: efficiency made compute cheaper, so we used far more

One year ago today, a model from a Chinese lab almost nobody had heard of erased nearly $600 billion of NVIDIA's value in a single trading session. The thesis behind the selloff was clean and intuitive: if AI can be made this efficient, the world needs far fewer chips. Twelve months later, demand for compute is higher than ever. The market was right about the efficiency and exactly wrong about the conclusion. Why?

What actually happened

On 27 January 2025, DeepSeek's R1 model went viral, having reportedly been trained at a fraction of the cost of comparable Western frontier models. Investors did the obvious math: cheaper to train, cheaper to run, so less hardware needed. NVIDIA fell roughly 17% in one day, shedding close to $600 billion in market value, the largest single-day loss for any company in US history at the time (Reuters).

Then the opposite of the prediction happened. Cheaper inference didn't shrink demand. It detonated it. More products could afford to embed AI, more queries became economical, and total compute consumed went up, not down. Microsoft's Satya Nadella named it on the day of the crash, pointing to a 160-year-old idea (WWT).

The Jevons paradox, explained in one breath

In 1865, the economist William Stanley Jevons noticed something odd about coal. As steam engines got more efficient, Britain didn't burn less coal. It burned more. Efficiency lowered the cost of using coal, which expanded all the uses worth putting it to, and the expansion swamped the savings. Make a resource cheaper to use, and you tend to use much more of it.

Swap coal for compute and you have 2025. DeepSeek didn't reduce the appetite for GPUs; it lowered the price of intelligence, which invited a hundred new applications that were previously too expensive to bother with. The efficiency was real. The demand destruction was a mirage. The Stanford AI Index has tracked inference costs falling by orders of magnitude across recent years, and usage has climbed right alongside the price collapse, not against it (Stanford HAI, AI Index).

Where the demand actually lands

Follow the paradox to its end. If cheaper models drive an explosion in usage, that usage has to run somewhere, and most of it is inference, the steady, high-volume, around-the-clock work of answering queries rather than the one-off spike of training. Inference is exquisitely sensitive to running cost, because you pay it on every single request, forever.

So the Jevons wave doesn't just create more demand for compute in the abstract. It creates more demand for the cheapest place to run inference continuously. When the marginal cost of a query is what decides whether your product is viable, the price of the kilowatt-hour underneath it stops being a line item and becomes the business model.