A single number made Wall Street panic: $6m. That is roughly what DeepSeek, a Chinese artificial-intelligence lab, spent training R1, a reasoning model that matched the performance of OpenAI's flagship o1. The American firm charges $15 per million input tokens and $60 per million output tokens for o1. DeepSeek charges $0.55 and $2.19 respectively — about 97% less. On January 27th 2025, investors did the maths. NVIDIA lost nearly $600bn in market capitalisation in a single day, the largest single-day loss in stock-market history. Total losses across American markets approached $1trn.

The reaction was not mere hysteria. It was the market repricing a set of assumptions that had quietly underpinned the AI boom — assumptions that turned out to be far less solid than they appeared.

The premise that cracked

The prevailing theory of AI progress was straightforward: more compute produces better models. Buy more chips, build bigger data centres, hire more engineers. The companies best positioned to do all three — and the firms supplying them, above all NVIDIA — were therefore worth extraordinary sums. This logic justified stratospheric valuations and a capital expenditure frenzy that saw American tech giants pledge hundreds of billions of dollars in infrastructure spending.

DeepSeek did not disprove the scaling hypothesis outright. But it demonstrated that the relationship between spending and capability is far less linear than the bulls had assumed. The lab achieved frontier-level performance through a combination of architectural ingenuity, efficient training techniques and, reportedly, a constraint that turned into an advantage: limited access to the most advanced chips, thanks to American export controls on high-end semiconductors to China.

When necessity mothers invention, the offspring can be inconvenient for incumbents. DeepSeek's engineers, unable to simply throw more compute at their problems, found cleverer ways to solve them. The result was a model that was not merely cheap to use but cheap to build — and, crucially, open-source, meaning anyone could download and run it.

The cost of being expensive

The pricing gap between R1 and o1 is worth dwelling on. A 97% cost reduction does not represent incremental improvement; it represents a different economic reality. At those prices, applications that were previously uneconomical become viable. Companies that could not afford to run inference at scale suddenly can. The addressable market for AI expands dramatically — but the revenue per query collapses.

This is a familiar dynamic in technology. When the cost of a thing falls by orders of magnitude, the companies that built their business models around its scarcity face an uncomfortable reckoning. OpenAI and its peers have raised tens of billions of dollars on the premise that building and running frontier models is extraordinarily expensive, and that this expense constitutes a moat. DeepSeek suggested the moat may be shallower than advertised.

The open-source dimension compounds this. R1 is not just cheap to access via API; it can be run locally, modified and improved by anyone. This hands a meaningful tool to researchers, startups and governments who would rather not route their queries through American servers. It also hands ammunition to the open-source AI movement, which had been arguing for years that closed, proprietary models were neither necessary nor obviously superior.

What the shock revealed

The market reaction was, in one sense, an overreaction. NVIDIA's chips are still essential. Data centres are still being built. The world's appetite for compute has not vanished because one Chinese lab found a more efficient training recipe. NVIDIA's share price recovered much of the lost ground in subsequent weeks.

But the shock was useful. It forced a public reckoning with questions that had been too uncomfortable to ask during the boom. How much of AI's stratospheric valuation rested on the assumption that expensive was synonymous with good? How defensible are the moats of companies whose chief advantage is access to capital and compute, rather than some irreproducible scientific insight? And what happens to the investment case for AI infrastructure if the efficiency of model training keeps improving?

DeepSeek also scrambled the geopolitical narrative. American export controls on advanced chips were intended to keep China behind in AI. They may instead have produced Chinese engineers who are, by necessity, better at doing more with less. The policy may still be worth pursuing — denying China access to the most powerful chips remains strategically significant — but the assumption that it would simply freeze Chinese progress looks naïve.

The new landscape

What comes next is a competition on multiple fronts simultaneously: capability, efficiency and cost. Western labs have responded with their own efficiency improvements; OpenAI, Google and others have cut prices and rushed out cheaper model tiers. The pace of change is accelerating.

For users, this is straightforwardly good news. AI is getting cheaper, faster. For investors, the picture is murkier. The AI boom was partly a bet on scarcity; scarcity is looking harder to sustain. That does not mean the technology is less transformative — if anything, cheaper AI is more likely to be widely adopted. But the distribution of the gains may look quite different from what the market priced in before a Chinese lab, working under constraints, decided to think more carefully about efficiency.

Six million dollars. In the context of an industry that burns billions, it is a rounding error. As a demonstration of what careful engineering can achieve, it was priceless.