America spent the past three years trying to build a wall around its most advanced semiconductors. DeepSeek has just blown a trillion-dollar hole in it.
The Chinese start-up’s release of an AI model that it said used a fraction of the chips, was built at a fraction of the investment and has lower ongoing costs than its vastly larger US competitors will inevitably trigger recriminations over the failure of the Biden administration’s efforts to restrict China’s access to advanced chips and slow its development of artificial intelligence capabilities.
During its term in office, the Biden administration steadily tightened a web of export restrictions on America’s most advanced semiconductors, or chips, in its attempt to protect and build on the US’ perceived lead in the technology that underpins AI.
Not only did it restrict exports of the chips, but it imposed rules on countries that did have access to them to stymie efforts by China to circumvent the restrictions. It also strong-armed the Dutch government to halt sales to China of the ASML machines that make the most advanced chips.
In the dying days of the administration – on January 13 – it unveiled a new set of rules that would govern the flow of advanced AI chips to the rest of the world, which it divided into three tiers.
The 19 tier 1 countries – the US, Australia, Canada, Japan, South Korea, UK, Taiwan and the major European and Scandinavian countries – would have unrestricted access, provided the entities involved weren’t owned or headquartered in a non-tier one country.
Tier 2 countries – about 150 of them – would have access to only a limited volume of the chips, while the two dozen or so in the bottom tier – China, Cuba, Iran, North Korea and Russia among them – would have no access at all.
The rules will come into force in May unless the Trump administration withdraws or modifies them.
Yet, despite the ever-tightening efforts to frustrate China’s ability to compete with the US on AI, DeepSeek confounded everyone and triggered a tech bloodbath in the US sharemarket on Monday by producing an AI model that is as good as, or better than, all but a couple of the US models that have cost hundreds of millions of dollars to build and train.
It was able to do that despite – or perhaps because of – the restrictions.
America first restricted access to Nvidia’s most advanced chips in 2022. Nvidia, galvanised by the prospect of lost revenue, promptly developed a slightly less advanced chip, but only slightly less advanced, to sell to China – its H800 chip. A year later, the Biden administration also included that chip, along with some similar chips, on the list of restricted exports to China.
It was the H800 – only about 2000 of them, versus the 16,000 or so that its US competitors would use – that DeepSeek claimed it used to build and train a model that cost only $US5.6 million ($9 million) to develop.
Until late 2023, when the Biden administration closed the loophole that allowed Nvidia to sell chips that were only marginally less advanced than its state-of-the-art chips, DeepSeek could have legally acquired the H800s. The start-up, along with other Chinese companies, would inevitably have been stockpiling chips during that time, anticipating tighter restrictions.
It’s also possible that, even after the tougher restrictions set in and Nvidia designed a chip even less powerful than the H800 for China, Chinese companies may have had access to smuggled advanced chips.
The New York Times has reported on the sales of thousands of banned chips in China by dozens of companies. The Biden administration’s plan for tiered access is designed to thwart that smuggling by limiting large-scale access to the chips to trusted allies.
While there is undoubted admiration for the ingenuity that enabled the creation of DeepSeek, as well as a realisation that – because it is open source and therefore its design and the efficiencies it developed will revolutionise and democratise the AI sector – it’s instructive that it was built on US chips.
Despite China pouring the best part of $US150 billion into its semiconductor industry, Chinese companies have not yet been able to make chips as advanced as Nvidia’s or the smaller US companies challenging its dominance. Nor has China been able to build the extraordinarily complex and expensive lithography machines that ASML is now unable to sell to China.
DeepSeek may have challenged America’s leadership in AI software, but US companies retain a significant lead in the hardware.
Unless and until China can develop homegrown chips that are as advanced as the most sophisticated US chips (which is a moving target) and the equipment to manufacture them, that gap in the race to develop AI with the cognitive capabilities of humans could widen.
America’s AI companies are confident that it will be computing power and the value-adding applications it will enable that will define the future of AI, rather than less powerful AI models like DeepSeek.
It was instructive that neither Nvidia (whose share price plunged) nor Donald Trump seemed perturbed by the unveiling of DeepSeek.
Nvidia called DeepSeek “an excellent AI advancement” and suggested it could lead to more sales of its export-controls-compliant chips. Trump described it as a wake-up call for US companies that would encourage innovation at lower cost.
OpenAI’s Sam Altman, whose ChatGPT ignited the fever around AI, said the DeepSeek model was “impressive” but wrote on X that “more compute is more important now than ever before to succeed at our mission” and that “the world is going to want to use a LOT of ai, and really be quite amazed by the next-gen models coming”.
It may be that the massive Chinese investment in semiconductors and chip-making pays off, and it can develop domestic capabilities that challenge those of America and its allies.
Without them, using up a diminishing store of increasingly outdated H800 or similar chips and the limited supply of smuggled chips – no matter how efficiently – will see China increasingly fall behind the pace of AI development.
The Trump administration, with its strong relationships with Silicon Valley billionaires, is committed to entrenching America’s dominance of AI.
On his first day in office, Trump tore up Biden-era rules that tried to establish safety and privacy standards for the use of AI, saying they hindered innovation and imposed onerous and unnecessary government controls.
Less than a week ago, he proudly announced the formation of a consortium planning to invest up to $US500 billion in AI infrastructure to help keep “the future of technology” in the US.
The release of DeepSeek will unsettle the administration (which is stuffed with China hawks) and the US Congress and may lead to yet another layer of export controls and sanctions.
Ultimately, however, while DeepSeek may have demonstrated that the US companies spending hundreds of billions of dollars a year on AI may not lead the world in terms of the efficient use of AI capital, it was their chips and the replication of their existing models by DeepSeek that enabled it to do what it did.
As they move up the AI value chain, and America’s restrictions on access to their essential building blocks – increasingly advanced chips – have had more time to bite, developers of AI in the US will be harder for Chinese companies to emulate.
Discussion about this post