China’s DeepSeek miracle is not everything it seems

If we are to believe that China’s DeepSeek has matched Silicon Valley’s AI giants in breakneck time, on a shoestring and with a tiny fraction of the data and energy use, the world is indeed in convulsion.

This is a geopolitical earthquake of the first order, upstaging Donald Trump’s $US500 billion ($800 billion) Stargate bid for supremacy in artificial intelligence (AI) before he has even started.

Those who thought US trade restrictions limiting the export of Nvidia’s GPU chips would keep the US ahead of China were wrong.Credit: AP

China’s Communist Party could hardly have wished for a more obliterating propaganda stunt to cut Trump’s strident new America down to size. A president who judges success by Wall Street faces a sharemarket crash within days of taking office. The politburo has a collective sense of humour.

Immense power and prestige passes instantly from the US to China. We are looking at a future where Chinese firms run away with AI, undercutting the West by “commoditising” it, just as China has done with solar panels, electric vehicles and lithium batteries.

But there are strong suspicions that we have not been told the full story. “Did DeepSeek really build an OpenAI for $US5 million? Of course not,” said Stacy Rasgon, Bernstein’s star tech analyst.

It is too soon to judge whether the US embargo on cutting-edge chips has backfired by forcing Chinese firms to get off the “bigger is better” hyperscaler bandwagon of ever more computing power and think outside the box. “Have they failed, as the industry argues, or have they not been stringent enough, as I hear people with a national security background arguing?” asked Bill Bishop, from Sinocism.

On the face of it, the DeepSeek shock demolishes the commercial assumptions that underpin America’s hyperinflated tech bubble and has laid waste to digital darlings across the world, with collateral casualties to match. Uranium stocks fell 11 per cent on Monday since the vast expansion of data centres may not be needed, and nor will the nuclear plants to power them.

Meta has set up a “war room” to assess the DeepSeek damage. The inherent future value of Nvidia is torn to shreds – down $US600 billion in value in one day – and that company alone is (or was) 7 per cent of the S&P 500 index. You don’t need to buy its elite H100 graphic processing units (GPU) for deep learning at $US30,000 a shot. You can do it with fewer and less advanced H800 chips.

However, note that Nvidia praised DeepSeek as a boon for the industry since it should accelerate the global adoption of AI and, therefore, drive a surge in demand for Nvidia’s other GPUs.

Cracks are starting to appear in the miraculous DeepSeek story. The company is the research lab and pet project of Liang Wenfeng, the reclusive founder of the hedge fund High-Flyer and the new hero of the Communist Party.

On January 20, DeepSeek released the open source R1, claiming that it had trained the model in under two months with lower-grade chips at a trivial cost of $US5.6 million. “We’re done following. It’s time to lead,” he said in a rare public outing unearthed by China Academy.

He said the US embargo on high-end chips had forced his team to rely on “inference”, a form of AI training that depends on recognising patterns in data. “We never intended to be a disruptor; it just happened by accident,” he said. Mirabile dictu.

But what really went into the secret recipe at DeepSeek headquarters in Hangzhou? Alexandr Wang, head of Scale AI in San Francisco, said the gossip from China is that the lab actually used 50,000 of the best H100 GPUs, either stockpiled in advance or obtained on the global black market since the US imposed its chip blockade.

“They can’t talk about it, obviously, because it is against the export controls that the US has put in place,” he told CNBC. DeepSeek’s Liang Wenfeng himself told China’s premier at a symposium last week that the shortage of advanced chips was a “bottleneck” holding back his lab.

The inference is that DeepSeek is, in reality, acting as an agent of the regime, drawing on the full resources and strategic GPU reserves of the Chinese state and also lending itself to a subtle campaign to try to undermine support for the US embargo in Washington.

If so, China has not, in fact, come up with such an ultra-cheap model that smashes the competition or one that suddenly opens the way to the mass utilisation of AI.

DeepSeek’s claims on cost stretch credulity. Mark Zuckerberg’s Meta is spending more than $US60 billion this year on AI. Its Llama 4 is expected to cost $US600 million and will require 160,000 advanced GPUs to train the model. Open AI is burning through $US5 billion a year on its closed-source ChatGPT architecture. OpenAI’s Sam Altman has been soliciting $US5 trillion. Yes, trillion.

Few dispute the capabilities of DeepSeek, so long as you don’t ask about Tibet or the Uyghurs, but in that respect it is much like Wikipedia: unusable for politics.

DeepSeek founder Liang Wenfeng (right), pictured at a symposium presided by Chinese Premier Li Qiang earlier this month, says AI “should be affordable and accessible to everyone”.

Marc Andreessen, the éminence grise of Trump’s tech oligarchy, praised the company, calling the new R1 model the Sputnik moment for Silicon Valley. “One of the most amazing and impressive breakthroughs I’ve ever seen,” he said.

DeepSeek matches US rivals on maths, coding and reasoning, which is remarkable in itself since China lacks the full AI ecosystem and ranks 17th in the AI index of capital economics.

The US is first, followed by Singapore, the UK, and Switzerland. Europe’s big states lag behind – Germany in 12th place, France 20th and Italy 25th – if they can keep anything after the EU’s suicidal AI Act.

Nathan Benaich, founder of Air Street Capital, told Sifted that the “real losers” of the DeepSeek saga are European firms building language learning models. The dominoes will fall hard and fast. “These companies have already been struggling to differentiate themselves on price or performance, so DeepSeek R1 presents a perfect storm for them.”

The US intelligence agencies probably have a good idea whether DeepSeek is telling the full truth about its use of Nvidia GPUs, and that, in turn, will inform whether Washington opts for tougher sanctions. But if Trump goes that route, he will further split the world into opposed technology camps. The Global South will gravitate ever more towards China.

The proper conclusion to draw at this stage is that DeepSeek has genuinely shaken the tech world to its foundations and may have opened a path to “democratised” mass AI that does not gobble up the world’s electricity, even if the company is throwing sand in our eyes about the true cost and methods.

That alone is a seismic geopolitical event, as Microsoft’s Satya Nadella warned in Davos. “We should take the developments out of China very, very seriously,” he said.

Telegraph, London

The Business Briefing newsletter delivers major stories, exclusive coverage and expert opinion. Sign up to get it every weekday morning.