Facebook
X
LinkdIn
Copy Link
In the years since ChatGPT’s debut, the world of artificial intelligence development has been defined by a single obsession: scale. Companies have raced to build ever-larger models, train on datasets of unimaginable size, and spend billions on the infrastructure required to sustain this rapid growth. The logic has been simple: bigger is better.
The pursuit of scale has inflated the industry, driving massive valuations. Nvidia — the shovel and picks provider of this new age — rose to a trillion-dollar valuation fueled by its GPUs being indispensable for AI development. Over the weekend, Meta announced plans for a data center spanning half the size of Manhattan, further reinforcing the industry’s commitment to infrastructure-heavy strategies.
The emergence of DeepSeek signals that the AI race is entering a new phase
Last week, this mindset was epitomized by the announcement of the Stargate Project, a $500 billion initiative to build the world’s largest AI infrastructure in Texas. It embodies the belief that sheer scale will secure dominance and unlock the elusive goal of Artificial General Intelligence (AGI) — the Holy Grail of modern AI research.
As US tech giants rely on sprawling data centers powered by Nvidia GPUs, US sanctions on advanced chip sales — first imposed under Biden and set to continue under Trump — have cut China off from critical AI hardware, forcing its developers to innovate with far fewer resources.
This is the environment that birthed DeepSeek. Founded just two years ago, the relatively unknown Chinese AI startup has emerged as a formidable challenger to the “bigger is better” narrative, achieving what many thought impossible: delivering performance comparable to the West’s cutting-edge models at a fraction of the cost.
DeepSeek’s flagship model, R1, excels in fields such as mathematics, coding and reasoning — domains typically dominated by resource-intensive models. Yet, remarkably, R1 was allegedly developed for under $6 million, a fraction of the billion-dollar investments of its Western counterparts. Now sitting at the top of Apple’s App Store, this achievement has upended assumptions about the economics of AI — and the market, understandably, took notice.
The impact was immediate. Shares in AI-related companies such as Nvidia, Microsoft and Meta dropped sharply as investors began questioning the sustainability of the scale-first strategies that have driven the industry. Nvidia, whose valuation was once thought untouchable, fell as much as 13 percent at the open — shedding $465 billion off its market value in what Bloomberg described as the largest single-day rout in stock market history.
Even European tech firms felt the tremors: ASML, the Dutch chip equipment maker, saw its stock tumble by over 10 percent, while Siemens Energy, a supplier of AI-related hardware, suffered a staggering 21 percent drop. One analyst told Bloomberg that DeepSeek could “potentially derail the investment case for the entire AI supply chain.” Altogether, the announcement appears to have triggered a trillion-dollar selloff in tech stocks.
Eye-watering sums aside, the shock is reminiscent of the Sputnik moment. Just as the Soviet Union’s launch of Sputnik 1 in 1957 forced the US to reevaluate its complacency in the space race, DeepSeek’s rise has laid bare the vulnerabilities in the West’s AI strategies. The reliance on trillion-dollar infrastructure and brute-force compute now appears precarious against a leaner, more resourceful competitor.
So, has the bubble finally burst? A little bit. But it’s not about which model you use — it’s about how you use it. The past three years have been defined by ever-larger models and shattered benchmarks, but the true breakthroughs now come from applying AI in smarter, more targeted ways.
To use a term coined by former AI researcher Leopold Aschenbrenner in his seminal AI essay series last year, the “unhobbling” is well underway. This shift marks a move away from sheer power scaling toward fully unlocking the potential of existing models. Techniques such as Retrieval Augmented Generation (RAG), chain-of-thought prompting and the application of smaller-language models (SLMs) are proving that AI doesn’t need to know the entirety of The Iliad — it needs to be precise, adaptable and fit for purpose.
The rise of DeepSeek exemplifies this shift, but it also highlights the growing bifurcation of AI across the world. While the unhobbling unlocks latent potential, it cannot escape the values, biases and constraints of its creators. Despite its technical brilliance, DeepSeek’s R1 bot is shaped by its environment: one heavily influenced by China’s political and social norms.
Want to ask about holiday ideas in Taiwan? Forget it. How about what happened in 1989? Computer says no. These omissions aren’t bugs — they’re features. They reflect the parameters set by a government that exerts tight control over technology, ensuring that sensitive topics are filtered out. It’s the same reason why Elon Musk’s Grok is chockful of cringeworthy humor. These models that will one day likely shape most of our lives are being shaped by those who create them.
That said, in the longer term, it would be wrong to suggest that compute is anything less than a critical factor in shaping the AI race. The UK’s AI Opportunities Plan, with its promise of a twentyfold increase in compute capacity and AI growth zones, marks a step in the right direction, for instance. Still, ambition is one thing; delivering infrastructure at the speed AI demands is quite another.
Fortunately for nations lacking in computing power, the emergence of DeepSeek signals that the AI race is entering a new phase. By innovating under constraint, DeepSeek has redefined what’s possible, proving that progress doesn’t always require sprawling infrastructure or trillion-parameter models. Instead, it calls for rethinking priorities: refining tools, embracing efficiency and adapting AI to serve practical, targeted purposes. Just make sure that purpose doesn’t involve anything to do with Winnie the Pooh and you’ll be fine.
