But there are strong suspicions that we have not been told the full story. “Did DeepSeek really build an OpenAI for US$5m? 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 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 hyper-inflated tech bubble, and has laid waste to digital darlings across the world, with collateral casualties to match. Uranium stocks fell 11% 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 – downUS$590b ($1.03 trillion) from Nvidia’s in one day – and that company alone is (or was) 7% of the S&P 500 index [Nvidia recovered some ground in Tuesday trading]. You don’t need to buy its elite H100 graphic processing units (GPU) for deep learning at US$30,000 a shot. You can do it with fewer and less-advanced H800 chips.

Though 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.

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

On Jan 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 $5.6m. “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 $60bn this year on AI. Its Llama 4 is expected to cost US$600m and will require 160,000 advanced GPUs to train the model. Open AI is burning through $5bn a year on its close- source ChatGPT architecture. OpenAI’s Sam Altman has been soliciting US$5 trillion. Yes, trillion.

Few dispute about 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.

Marc Andreessen, eminence 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 [New Zealand was 18th].

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, the 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 or not 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.