# DeepSeek and the Repricing of AI: Does Building Bigger Models Still Make Sense?

Author: Norbert Biedrzycki
Published: 2025-02-15
Language: en
URL: https://norbertbiedrzycki.pl/en/deepseek-and-the-repricing-of-ai-does-building-bigger-models-still/
Summary: On a single Monday in January, Nvidia lost close to $600 billion of market value. Not because demand for AI had collapsed, but because a Chinese lab showed that intelligence might be much cheaper…

**On a single Monday in January, Nvidia lost close to $600 billion of market value. Not because demand for AI had collapsed, but because a Chinese lab showed that intelligence might be much cheaper than the market had assumed.**

On 20 January, DeepSeek released R1, a reasoning model whose performance on several benchmarks was comparable to the best American models. What shocked investors was not only the quality but the reported cost. DeepSeek said the final training run of its underlying V3 model used around $5.6 million of GPU time, a tiny fraction of what Western labs are believed to spend. The model was released openly, and its app briefly topped download charts in the United States.

On 27 January, Nvidia shares fell around 17 percent, the largest one day loss of market value for any company in history. Power and data centre stocks fell with it.

The irony is that only days earlier, the Stargate project had been announced in Washington, with plans to invest up to $500 billion in AI infrastructure in the United States over four years.

**What the market got right, and what it may have got wrong**

It is important to be precise about what DeepSeek did and did not show. The $5.6 million figure covers only the final training run. It excludes research, experiments, failed runs, salaries and the hardware the company already owned, which by most estimates was substantial. Some of its efficiency may also have come from learning from existing frontier models.

Even so, the direction is undeniable. DeepSeek demonstrated that clever engineering, under hardware constraints imposed by US export controls, can deliver near frontier performance at much lower cost. Constraint, as so often in the history of technology, turned out to be a source of innovation.

**The Jevons question**

Within hours of the sell off, several technology leaders pointed to the Jevons paradox. In 1865 the economist William Stanley Jevons observed that more efficient steam engines did not reduce coal consumption in Britain. They increased it, because cheaper energy opened up new uses.

If the same happens with AI, falling costs per unit of intelligence will drive an explosion in usage, and total demand for compute will keep growing. That is the bull case for infrastructure. The bear case is that efficiency gains will outpace demand growth in the short term, leaving parts of the infrastructure overbuilt and putting pressure on pricing.

My honest view is that both can be true at the same time, but for different players. Total demand for compute will probably keep rising. Yet the premium paid for frontier scale alone may shrink, and the advantage of simply having the biggest budget will erode.

**What changes for investors**

For investors, DeepSeek shifts the central question. For two years, the dominant logic was: whoever builds the biggest model wins. Now the question is: who can deliver the best performance at the lowest cost per useful outcome?

This favours several groups. Application companies, whose input costs fall. Enterprises, which will get better models for less. Open ecosystems, which become more credible. And the efficiency layer: companies optimising inference, compression and specialised hardware.

It is less favourable to business models that depend on charging a premium for access to a single frontier model, and to infrastructure investments that assume today's prices per unit of compute will persist.

For boards, the practical lesson is simpler. Do not lock your AI strategy into a single model or a single supplier. The cost curve is moving too fast.

*Investment question: Does AI investment success depend on owning the largest computing infrastructure, or on making intelligence dramatically cheaper?*
