Wall Street’s trillion-dollar wipeout sets off alarm bells

The vulnerability of the US sharemarket to its highly concentrated reliance on a small cluster of mega technology stocks has been exposed by the release of a large language model from a Chinese artificial intelligence start-up that didn’t exist 18 months ago.

The US market fell modestly on Monday but the fall in the big tech stocks – the so-called “Magnificent Seven” – was anything but modest.

More than a trillion dollars was wiped off the value of tech stocks on Monday. Credit: Bloomberg

An index of those stocks – Alphabet (Google’s parent), Amazon, Apple, Meta Platforms (Facebook’s parent), Microsoft, Nvidia and Tesla – slumped more than 3 per cent. The market capitalisation of the flagship for America’s AI sector, Nvidia, was slashed by more than $US590 billion ($938 billion).

The trillion dollar-plus savaging of the AI-related stocks was sparked by last week’s release of a large language model by a Chinese start-up, spun out of a Chinese hedge fund, and then Monday’s release of a detailed paper on how the model was built.

DeepSeek’s AI assistant rocketed to become the top-rated free application in Apple’s app store on Monday, overtaking OpenAI’s ChatGPT.

The reason that DeepSeek sent tremors throughout the US AI sector, and punctured the share prices of those most heavily invested in it, is that the company claims to have built its model using a fraction of the Nvidia chips that the US companies have used to construct their models and at a fraction of their costs.

DeepSeek claims to have built its latest model, whose performance is similar to or better than most of its US competitors, for $US5.6 million, using only about 2000 chips. Similar chatbots in the US contain up to 16,000 chips and Open AI’s Sam Altman has said it cost more than $US100 million to train the ChatGPT-4 model.

While there is some scepticism over the modest cost of building the model and the quantity of chips involved claimed by DeepSeek, testing of the model suggests that only OpenAI’s and Google’s most powerful models outperform it. DeepSeek also appears to have built a far more efficient model that maximises the computing power of a limited cluster of less-than-leading-edge chips.

That throws into question the concept of American AI exceptionalism and the conviction that the massive investments US companies are making in AI – $US224 billion last year and $US280 billion this year, according to UBS – will give it a widening lead in AI developments over China.

Only last week Donald Trump proudly announced that a consortium of OpenAI, Oracle, Japan’s Softbank and United Arab Emirates’ tech invest fund, MGX would invest up to $US500 billion to build AI infrastructure in the US, while Mark Zuckerberg said Meta would spend $US60 billion to $US65 billion on its AI business.

The US efforts to restrain China’s development of AI and other advanced technologies appear to have forced Chinese companies to develop innovative workarounds.Credit: Bloomberg

DeepSeek’s apparent ability to do what the US companies have invested hundreds of billions of dollars to accomplish plays on pre-existing fears of investors in those companies that they have massively over-invested and that the returns on those investments will be too distant to make commercial sense.

With the Magnificent Seven accounting for nearly a third of the US market’s value, the entire US market was leveraged to those companies’ share price performance.

The companies themselves were trading at prices that capitalised perfection – Nvidia, for instance, was trading on a multiple of about 55 times its earnings – so they were very exposed to the first indication that the moats that protected them – their intellectual property and the sheer scale of their investments – had been breached.

The other protection the US industry was supposed to have – the Biden administration’s layers of US sanctions and restrictions on access to the Nvidia chips that dominate the AI industry, with a market share of about 90 per cent, and the machines that manufacture them – has also proven to be imperfect.

DeepSeek appears to have stockpiled earlier generation Nvidia chips and may also have had access to some of its more powerful chips via third parties. (There’s a big trade in smuggling Nvidia’s latest chips to China).

However, its ability to produce a model that competes with the best from the US says something about the counterproductive nature of those trade measures.

As Huawei was forced to do when it lost access to the latest US semiconductors, DeepSeek had no alternative but to innovate and fund more efficient ways to use its limited store of earlier generation chips.

Huawei, which has been showered with US sanctions, shocked the US in 2023 when it released a smartphone whose capabilities showed it had been able to completely design and manufacture a relatively advanced chip without access to the technology of the US and its allies.

The US efforts to restrain China’s development of AI and other advanced technologies appear to have forced Chinese companies to develop innovative workarounds. Rather than throttling China’s ability to compete, the US may have stimulated Chinese innovation.

In DeepSeek’s case, it has developed an AI model at a fraction of the conventional cost, both to build and operate, and therefore with a reduced need for the vast and energy-intensive data centres that the US industry is building.

It could be that DeepSeek’s ability to do what it has done with fewer and less advanced chips has been facilitated by the existence of ChatGPT and other models, particularly Meta’s Llama open source model. It may have simply replicated existing models.

In any event, its success in developing innovative and more efficient ways to train models and then sharing its breakthroughs threatens to commoditise large language models.

That won’t stop US companies from trying to move up the AI curve in pursuit of AI with expert human capabilities, or Nvidia developing ever more powerful chips to create the computing power that would enable them to achieve that, and between them opening up a new gap over their newly-emerged Chinese rival.

The emergence of a cheaper, open-source alternative that developers can play with and build tools for might, however, help accelerate the deployment of AI and help show other companies outside the US that they can compete in AI without the same scale of investment as the US mega techs.

Competition in technology drives innovation and take up. DeepSeek might have shaken the US industry up this week, but that might prove to be a positive if, with relatively unsophisticated models commoditised, the US companies feel the need for more urgency in developing ever-more sophisticated and valuable – and profitable – models and applications as they move up the AI value chain.

The Market Recap newsletter is a wrap of the day’s trading. Get it each weekday afternoon.