Over the last couple of weeks, AI risk seems to have gone mainstream. Social media was set abuzz when some current and former Anthropic employees noted that they genuinely believed that there was a real risk of an AI disaster wiping out the human race. That led, understandably enough, to all sorts of coverage across the wider media and brought greater public attention to the innocuously named estimate of ‘P(doom)’, or the chance that AI will lead to an extinction-level event. Cheery stuff indeed.
But whether one takes this scenario seriously or not, and some cynics can certainly see why the frontier AI labs might be keen to talk up just how powerful their new technology is ahead of likely IPOs, there are plenty of other AI-related risks that few can afford to ignore.
Over the last 18 to 24 months, financial markets have shifted their attention from one flavor of AI worry to another. For a long time, the bear case seemed to be that the technology would simply not work as well as hoped and that the vast resources poured into it would turn out to be dud investments. For a while towards the start of this year, the opposite fear gripped markets: that the technology worked too well. That saw the value of software firms plummeting as investors worried that AI would eat their business model.
Now though, worries are focused on a far more traditional risk: that of old-fashioned overvaluation and overconcentration, served with a side dish of fears about complicated circular financing structures.
As any student of market history knows, even the most genuinely transformative of new technologies – whether that be railroads in the nineteenth century or the internet at the end of the twentieth century – can still find themselves associated with overtly generous valuations. Investors can, and frequently do, get overexcited by the promise of vast profits in the future and push valuations out of touch with reality. Some of the assumptions implicit in the pricing of SpaceX seem little short of heroic.
The real worry is not that some over-optimistic investors will end up losing money but that the scale of the AI boom, and its related capital expenditure, is being a concentrated risk to the macroeconomy as a whole.
The numbers are now staggeringly large. The so-called hyperscalers look set to invest around $800B this year. This August, though, Goldman Sachs argued that even this huge number is an understatement.
Goldman Sachs Research extrapolated its preferred estimates for AI investment to forecast total AI investment as a share of US GDP through 2028. Those estimates imply that AI capex will rise from 1.8% of GDP in the US (with global AI investment totaling 0.9% of global GDP) in 2026 to 2.5% of GDP in the US (1.3% globally) in 2027, with a further increase to 2.8% (1.4% globally) in 2028.
“These levels are consistent with the 2%-5% of GDP peak investment impulses observed in prior general-purpose technology buildouts,” Briggs writes. “And while our US portfolio strategy team has flagged that consensus capex projections for 2027 are likely too conservative, even significant upward revisions would leave the level of AI investment as a share of GDP comfortably within the historical range observed in prior technology cycles.”
Asked whether ‘The AI-led boom in investment and growth has substantially increased the exposure of the economy and the stock market to a single concentrated source of tail risk’? there was widespread agreement. 30% of respondents, weighted by confidence, strongly agreed and another 64% agreed. That is one of the most decisive polls ever conducted of the finance panel.
Of course it is hard to argue that when AI-related capex alone is now worth something like 2 to 3% of GDP that concentration risks have not risen. That does not necessarily mean that this is a mistake or a bad thing. As Steve Kaplan of Chicago Booth argued, ‘answer is probably yes, but expected benefits are very high. So risk-return is extremely favorable.
Other respondents were concerned. As Nancy Wallace of Berkeley Haas stated, ‘I have spent a lot of time evaluating the off balance sheet debt deals and the heavy PE use of wholly owned insurance co. to park the debt. The omni presence of Nvidia, Open AI, and Anthropic as counter parties in these deals is very concerning!’.
The omnipresence of a handful of counterparties in what has become a major driver of growth is the next obvious point of concern coupled with the existence of circular funding systems and a large dash of vendor financing.
Asked whether ‘The use of financing from technology suppliers such as Nvidia to support investments in AI infrastructure substantially increases the financial risks associated with the AI-led boom’, 10% of respondents, again weighted by confidence, strongly agreed and 59% expressed agreement. That was less decisive than the first question but still rather striking.
Stijn Van Nieuwerburgh of Columbia Business School argued that ‘Nvidia is lending its balance sheet to facilitate more investment into GPUs by financially weaker companies. The guarantees it has provided are in the hundreds of billions. Entire AI ecosystem has become vulnerable to Nvidia shock’. While Loretta Mester of Wharton added that, ‘Suppliers are funding their customers, which creates concentration risk. Reports suggest very high leverage, but there is opacity around the amounts, which increases the risk’.
Jonathan Parker of MIT Sloan offered a different viewpoint, arguing that ‘Financing the AI investment boom by lending from suppliers of chips to users of chips keeps more of the risks within the sector and out of banks, pension funds, etc. If AI turns out to be less profitable than expected, the financial fallout has less systemic impact’.
In the view of the panel, then, there are two concentration risks worth taking seriously – that of the stock market and economy as a whole to the AI investment complex and within that complex itself, the key role of Nvidia.
In other words, there are plenty of things to worry about when it comes to AI before you even start considering the end of the world.
