There is a story that is often quoted in certain circles, which goes roughly like this: “Nasa spent millions of dollars inventing the ballpoint pen so they could write in space. The Russians took a pencil.”
It’s a pleasing narrative. Regular pens don’t work in zero gravity, but astronauts needed to write. America, the world’s largest and richest economy, threw money at the problem to develop a better pen, but while the USSR was outspent, it found a much more ingenious fix in far less time.
As it happens, the story isn’t strictly true – pencils flake, which isn’t helpful in zero gravity, and they’re a fire risk to boot, meaning both the Soviets and the USA bought in pens. But it certainly chimes with the recent news that a Chinese company has managed to develop an AI model that rivals the best the US produces, with far fewer resources and at a fraction of the cost.
The USA’s big tech behemoths have been engaged in a brute force contest reliant on pure computing power to develop and improve cutting-edge AI systems. This relied on buying the most advanced chips from Taiwan, stacking them up in huge numbers in data centres, requiring massive power draw and water use. The winners of the AI race would be whoever could finance and build enough expensive infrastructure to develop the best models.
Enter DeepSeek. The Chinese developers of the new AI model claim that, thanks to some clever tricks making training much more efficient, it cost only $5.6m to train, and used a relatively small number of affordable chips to do so.
What’s more, DeepSeek seems to be a very good AI model indeed – it’s competitive with the very best publicly available AI models from OpenAI, Google, and others, even beating them on some benchmarks. DeepSeek is cheap to run, efficient, and has been made open source, meaning anyone in the world can download and deploy it, for free.
Overnight, the AI game has changed. The conventional wisdom had been that the huge cost of training cutting-edge AI – coupled with its reliance on hard-to-obtain chips – meant that a small number of huge companies would be competing with one another to dominate what looks likely to be the most important tech of the next era.
It also became immediately obvious what it might look like to have a dominant technology that isn’t American – especially when DeepSeek was asked about Tiananmen Square, the treatment of Uyghurs in Xinjiang province, or China’s Great Firewall. The default DeepSeek app simply refuses to answer, in line with China’s censorship restrictions, but that isn’t the end of the story.
If someone runs DeepSeek on their own computer, or an alternative cloud, DeepSeek will answer all of those questions – but it reveals the bias of its training data, stressing China’s compliance with law and giving a radically different version of the answer to other models. This is because DeepSeek’s training data included much more Chinese media than western models, changing its answers even when censorship is avoided.
The impact of DeepSeek isn’t just that there’s a new model out there, it’s that it opens the possibility that almost anyone, anywhere, could come out with the next breakthrough.
One mooted possibility is that the researchers behind DeepSeek are either lying or exaggerating, and that in reality the model cost a lot more to develop (and/or took much more time) than was disclosed. But generally, the world has taken notice of the development, and is taking it seriously – to the tune of one trillion dollars.
That was the scale of the sell-off in the stock market that followed once investors had absorbed the implications of DeepSeek’s launch – with tech stocks predictably taking the biggest hit. Chipmaker Nvidia lost almost $500bn in one day of trading, the biggest one-day loss in history. Investors thought they knew who would own the next technological revolution. Now they’re not so sure.
But as to what DeepSeek’s launch actually means… that’s where people tend to be much less sure. As often happens in such situations, more than a few pundits stepped forward immediately to say that the unexpected development of DeepSeek proved what they’d been saying all along about AI.
This was perhaps most obviously strange in the case of people who have argued that AI is just the latest overhyped trend from big tech – which has had no shortage of supposedly revolutionary new technologies in recent years that have amounted to very little.
First it was cryptocurrencies, which have become a speculative asset but haven’t actually found any practical use yet. Then for a while, NFTs – essentially online receipts proving you “owned” a particular image or meme – were going to change the world. After that, we were told the “metaverse” would be so big that Facebook even changed its name to Meta – before all mentions of the metaverse disappeared completely.
With such a long line of duds in the run for the next big thing, the sceptics have some reason to be dubious of the huge claims being made for the potential of AI. But the new models do actually work, and are already delivering practical results: they can write decent code, edit copy, do competent research – to the standard of entry-level workers in any of those professions – and they are improving fast. While AI may not live up to the humanity-transforming hype of its greatest fans, there is a real technology here.
Justifiable as their broader position is, sceptics suggesting that DeepSeek somehow validates their view that AI is hype is just bizarre. AI is suddenly cheaper and faster to develop, with new tricks to improve models more quickly than could be done before. That might lower the profit potential for US companies, but it doesn’t suggest the tech won’t work or be transformative – it does the opposite.
The inverse take would be that DeepSeek makes it even likelier that huge AI breakthroughs will happen very soon and radically transform society – perhaps by delivering AIs that can perform any task better than any human within the next few years. This one is much more consistent with the new facts than the sceptical position, but is far from a given – it relies on DeepSeek’s story being entirely true, on development continuing at its fast pace, and more.
This is compounded by the gap between the apparent potential of AI and its practical rollout, which has been slower than the industry seems to have projected. The expectation was that we would have somewhat autonomous AI agents acting on our behalf by now – perhaps as personal assistants, booking our travel or accommodation, sending emails, etc.
In reality, the systems are just not reliable enough for most of us to trust them with our bank details or email inboxes. Similarly, AI has yet to fully replace customer service agents and more. The gap between 95% reliability and the 99.9% reliability that these kinds of functions need is proving very difficult to fill.
The optimistic scenario’s reasoning essentially posits that DeepSeek has found ways to train AI much more efficiently, and to do more with less. So what happens if you combine those techniques with the sheer computing power available to OpenAI and others? It’s certainly an interesting question.
Both Donald Trump and Keir Starmer have reasons to hope that DeepSeek’s innovations don’t make huge computing power irrelevant. In the first week of his second presidency, Donald Trump announced a $500bn AI data centre project, Stargate AI, which was immediately ridiculed by Elon Musk, not least as it involved his adversary Sam Altman and OpenAI. Just weeks ago, Starmer essentially staked the UK’s economic future on AI and investment. Both men have reason to hope DeepSeek doesn’t change the game too much.
There is also the possibility that DeepSeek is just less of a big deal than it appears. A trillion-dollar crash, focused among big tech companies, certainly seems pretty monumental in scale – but in the context of the huge increases in share prices of these companies in recent months, it’s not necessarily all that dramatic.
By a few days after that huge tech crash, most of the tech giants were roughly back where they started. Microsoft’s stock price was fractionally higher than a week before the crash. Alphabet (Google’s parent company) was down 1% on the week, but up 2% on the month. Apple and Meta were both up 8% over the course of the week, and even Nvidia – the company hardest hit – was down 16% monthly, but its share price was still up almost 20% vs six months ago.
Investors are essentially constantly trying to predict which companies will be the winner of the next round of technological change. Innovation and breakthroughs are traditionally quite hard to predict, and DeepSeek made investors a bit less confident they’d picked the winners this time. That confidence was briefly rocked, but by and large they still think US big tech will be fine. They may or may not be right, but that’s just business news.
Where DeepSeek has definitely changed the game – at least for now – is in the geopolitics of AI. Donald Trump was gearing up for a fight with China (and inevitably numerous other countries) over controlling AI, through control of the chips required to develop models.
He was hoping to develop domestic production of chips and control where chips from Taiwan and elsewhere could travel – thus restricting who could develop top-tier artificial intelligence and trying to guarantee US dominance of the next era of tech.
If DeepSeek’s claims are as described, and AI can be developed without specialist high-end chips, any hope of limiting its development are gone. The combination of that with DeepSeek being open source suggests AI development will be less centralised, less subject to control, and less predictable than anyone imagined.
That is a blow to the hopes of Trump or others to pick the winners of the AI race, but it is similarly a blow to the so-called AI doomers, who want a slowdown of AI development – or for regulation to be in place before ever-more advanced models are rolled out.
If that was ever a serious possibility, that moment has passed. AI’s development can no longer practically be slowed, let alone stopped. If regulators want to have any say in its development, they cannot rely on slowing down AI to their pace – they will have to accelerate to meet the changing capabilities of the technology.
It is totally understandable for people to want to make technological change predictable and even controllable – but revolutions by their very nature defy and confound expectations. Almost no dominant technological company of one era has gone on to be a major player in the next. DeepSeek is changing the game in one sense, but in another it’s showing that the old rules still apply – and the world is less knowable than those in charge might hope.
