Market panic, muted impact, and the AI shock that wasn’t

• DeepSeek’s debut caused market panic but didn’t affect long-term AI chip demand.
• Nvidia, Broadcom, and ASML rebounded, posting strong gains after initial shock
• DeepSeek’s efficiency gains intensified the AI arms race instead of ending it.

Nearly a year ago, DeepSeek appeared to upend the global AI narrative overnight. A relatively unknown Chinese AI lab released models that claimed comparable performance to leading US systems, trained with fewer and less advanced chips.

Markets reacted brutally. Shares of Nvidia, Broadcom and ASML plunged in a single session as investors priced in the risk that demand for expensive AI infrastructure could weaken. Almost every tech CEO was asked about what the impact of Deepseek meant. There were cautious words, some worried words but it was clear that the AI world sat up and took note.

That moment now looks more like a shock of perception than a structural turning point.

Fast forward 12 months and the same companies that were punished by the market have gone on to post some of their strongest gains. Nvidia not only recovered but became the first company to cross a $5 trillion valuation. Broadcom and ASML also delivered solid returns through 2025. If DeepSeek was meant to dent the AI spending boom, it did not.

Why the initial shock faded faster than expected

Part of the explanation lies in what DeepSeek did next. Since its headline-grabbing releases, the company has pushed out a series of updates rather than a fundamentally new model. These releases improved efficiency and refined existing capabilities, but they did not change the underlying economics of AI in a way that would threaten chip demand. Investors gradually came to see DeepSeek’s progress as evolutionary rather than disruptive.

Another reality has also set in. Training frontier AI models still requires enormous computing power. While software optimisation matters, it has not replaced the need for vast amounts of hardware. US export controls have further limited access to the most advanced chips in China, making it harder for companies like DeepSeek to scale at the same pace as their Western counterparts. That constraint has tempered expectations about how quickly Chinese labs can leapfrog the leaders.

At the same time, the pace of innovation in the US has not slowed. New flagship models from OpenAI, Google and Anthropic have reinforced the idea that the frontier continues to move forward, absorbing efficiency gains rather than being undermined by them. Instead of reducing spending, companies have doubled down, betting that smarter models will unlock even more use cases and justify even larger infrastructure investments.

In hindsight, DeepSeek’s debut mattered less for what it changed technically and more for what it exposed psychologically. It shattered complacency and reminded markets that AI leadership is contested. But once the initial fear passed, fundamentals reasserted themselves.

That does not mean the story is over. DeepSeek is still active, still publishing research, and still searching for ways to do more with less. Another surprise is always possible. But a year on, the whole Deepseek ‘episode’ made one thing clear: efficiency alone has not broken the AI arms race. It has, if anything, intensified it.