New Times,
New Thinking.

  1. Technology
29 January 2025

DeepSeek has exposed America’s AI fiction

The story which has pushed US tech companies to dizzying market valuations over the past two years is no longer convincing.

By Will Dunn

Everyone knows the old story about the space race: in the 1960s, Nasa needed to solve the problem of astronauts being able to write notes in space, so they spent millions developing a pen that could write in microgravity. The Soviets, the story goes, took pencils.

Most people also know this isn’t true. Pencil shavings, in the high-oxygen atmosphere of a spacecraft, are a fire hazard. The space pen was developed by a private company and sold to Nasa in a brilliant piece of marketing by the American businessman Paul Fisher.

The space-pencil story speaks to an old American anxiety, however, that the over-engineered products of capitalism could be superseded by cheaply made socialist alternatives. It is perhaps not a coincidence that the story emerged around the time of the Vietnam War, when American soldiers found their new M-16 rifles malfunctioned while Soviet Kalashnikovs kept firing.

This anxiety resurfaced this week after the debut of a new large language model, an AI chatbot called DeepSeek developed by a Chinese company; it was referred to by the American investor Marc Andreessen as a “Sputnik moment”. Microsoft and OpenAI are now said to be investigating claims that DeepSeek may have used data gathered from using their ChatGPT model. Mustafa Suleyman, CEO of Microsoft AI, recently claimed that any information accessed through the internet was “freeware, if you like”, but apparently Microsoft’s definition of “fair use” changes when it is Microsoft’s intellectual property being stolen.  

What has unsettled AI investors is not the power of DeepSeek but the fact that it can offer a fairly similar experience to OpenAI’s model, ChatGPT, at a fraction of the cost. The model of AI development that so many businesses and governments had been led to expect – a model built on the idea of relentless scaling-up of models and historic investment in energy-hungry data centres – has been upended.

It is hard to say exactly how much cheaper it has been to build DeepSeek than it was to build ChatGPT. Reports that the model cost around $5m to build are based on the cost of training the final version, not of creating the company itself or all of the research that went into it. But the difference in scale between the relatively small Chinese hedge fund behind DeepSeek and Microsoft, OpenAI’s main backer, is enormous, and DeepSeek’s model has been shown running on computers that would never be able to run the latest ChatGPT.

For end users, DeepSeek appears to offer a similar product that is orders of magnitude cheaper. For software developers who might use its models to, for example, build a chatbot, the “per-token” cost of using its V3 model is more than 100 times cheaper than the latest ChatGPT.

Subscribe to The New Statesman today for only £1 per week

The main implication of this is that the story which has pushed American tech companies to dizzying market valuations over the past two years is no longer convincing. The results of this change in confidence are spectacular: the market value of Nvidia, which sells the chips used for advanced AI models, declined by more than half a trillion dollars on 27 January – the biggest one-day fall for any company in US stock market history.

Particularly galling for US investors is that this may be the result of America’s own protectionist policy. In China, Nvidia is only allowed (under the 2022 Chips and Science Act) to sell a hobbled version of the processors it sells in the US, re-engineered to limit their performance. (Nvidia hasn’t said exactly how much the chips it sells to China are hindered, but they are thought to transfer data between processors at about half the rate of the full-speed US models.) DeepSeek seems to have accepted this, and written a much more efficient model around these limitations. What this means is that America’s attempt to hold back Chinese AI development appears to have driven it to become more efficient. The US government has unwittingly revealed that the story its own companies were telling its own investors was not strictly true.

That story went like this: AI would profoundly change societies and economies around the world. America had a first-mover advantage because it controlled the supply of the best processors, and it had companies that were investing the most in the technology. If it kept these companies at the forefront of the boom, it would continue to enjoy unrivalled economic dominance. This story was so convincing that all other concerns were swept aside. The intellectual property of America’s other industries was claimed as training data. Coal-fired power stations were kept burning to supply the data centres of Google and Meta with power for their AI models.

Companies also rapidly reoriented their politics to support the narrative that America would lead a world reshaped by AI. Nine years ago, as the US prepared to elect Donald Trump for the first time, the CEO of OpenAI Sam Altman penned an impassioned condemnation of Trump’s policies. On his personal blog, Altman decried the soon-to-be president’s “casual racism, misogyny, and conspiracy theories” and stated that he was “irresponsible in the way dictators are”. He found “chilling” parallels between America in 2016 and “Germany in the 1930s”, he warned that “demagogic hate-mongers lead down terrible paths”, he compared the lies of Donald Trump to the lies of Adolf Hitler. This month, Altman tweeted – as he announced a $500bn new AI initiative with the Trump administration – that Trump was going to be “incredible for the country”. 

This was all done in pursuit of scale, which Altman, Elon Musk and others believed would win the AI race. The strongest processors working with the largest amounts of data (and therefore the largest amounts of energy, water and other people’s intellectual property) would give America the edge. What appears to be happening instead is what followed the original iPhone: not another world-changing innovation, but an army of progressively cheaper imitations.

What this represents is not only a challenge to America’s dominance in AI but also to the efficiency of its capital markets, which suck up investment from around the world, including from your pension, for the benefit of a handful of tech companies that are increasingly blatant in their desire for political power beyond America’s borders. There are plenty of questions to ask about DeepSeek, not least its coy responses when asked about Uyghur people in Xinjiang, or what happened in Tiananmen Square. Yet there is room for cautious optimism about anything that challenges the dominance of America’s financial markets.

[See also: Donald Trump’s gladiator politics]

Content from our partners
The future of exams
Skills are the key to economic growth
Skills Transition is investing in UK skills and jobs