The first test is cash flow. Can this investment generate enough revenue to service the capital behind it? The answer depends heavily on price, and here the US faces China. China has built models with roughly 10 per cent of the US capital spend while reaching about 90 per cent of US performance. The cost of using a Chinese model is often only 10 per cent to 20 per cent of an American one.
That matters. Most users will not pay ten times more for a small quality gain. Frontier researchers may need the very best systems, but most firms do not; they need cheap, reliable productivity. On that basis, the Chinese offer is hard to beat. It also travels well. Many Chinese models are open source, so users can adapt them freely. They are built for practical use across the economy, not only for frontier display. The US approach, by comparison, is far more capital intensive, which leaves it more exposed if prices fall.
The second test is financing. China has less capital at risk and lower AI equity valuations, so a correction there would hurt but would be less systemic. The US is different. Its hyperscalers and labs have raced ahead of their own cash flow, and they now need debt to fund compute, power, and data centres.
By my estimates, AI hyperscalers and labs carry US$356 billion of long-term debt and US$248 billion of lease liabilities. Those are just the visible obligations. Off the balance sheet sit about US$900 billion of additional lease commitments and US$1.5 trillion of purchase commitments for data centres and advanced chips. The real exposure is therefore much larger than reported debt alone would suggest.
This structure is fragile. Circular shareholdings increase borrowing capacity, but they also spread losses around the system. If one participant weakens, other balance sheets get damaged in turn. The risk appears dispersed across many players, but in substance it is concentrated.
The borrowing structure adds another layer of weakness. A hyperscaler, think Meta, may guarantee the borrowing of a special purpose vehicle. That vehicle buys the chips and builds the data centre. A smaller lab such as Anthropic then signs a 15- or 25-year lease to use it. The debt itself is priced on the guarantor’s strength, but the cash needed to service it comes from the lab’s product. A large liability is therefore pinned to a narrow, unproven revenue stream. If that product misses expectations, the whole structure breaks.
That mismatch becomes lethal if US providers are forced to cut prices by as much as 80 per cent to match China on cost. At those prices, the capital stack simply cannot earn enough. Debt service becomes doubtful. Equity values would fall first, and credit would follow close behind.
A US AI collapse would then become an economic winter, for four reasons. AI is already deeply embedded in capital spending, corporate strategy, and market optimism, so its effects would not stay contained to one sector. A break of this kind would hit credit markets directly, not just equity multiples, spreading the damage into the broader financial system. Policy would lag behind events; denial would come well before any real repair effort begins. Households now own the AI trade through their portfolios and pensions, whether they realise it or not. Losses on that scale would cut household spending and lift precautionary saving.
The scale of the AI buildout is not, by itself, a reason for optimism or pessimism. It simply raises the stakes on the outcome. A technology this large, financed this aggressively, and this tightly woven into portfolios, credit markets, and corporate strategy, leaves little room for a soft landing, if the underlying economics do not hold. The world may still get to the good ending, where AI lifts productivity broadly and prosperity spreads. But the more likely route there, on the evidence of price, debt, and financing structure, runs through a crisis.