AI’s Cash-Flow Test: A Lesson From the GFC

In brief

• Like in the run up to the global financial crisis (GFC), the AI buildout depends on continued access to capital and on borrowers making their future payments.
• Infrastructure providers borrow against customer commitments that may not hold up if funding tightens or demand shifts.
• Active managers can underwrite the conditions necessary to potentially separate the winners from the losers.

A Narrow Comparison to 2008

A lot of comparisons have been made between today’s AI cycle and the technology, media, and telecom (TMT) boom of the mid to late 1990s. While there are obvious similarities, namely a new and powerful technology driving immense spending with large question marks around who will earn acceptable returns, there are differences, too.

A less common, but appropriate, comparison is the US housing boom that preceded the global financial crisis of 2008. This is not to say AI resembles housing or that we see a problem with the banking sector. The analogy is narrower and, as always for us, about cash flows.

In 2006, US housing appeared healthy. Prices were still rising and credit was readily available to borrowers. But home price appreciation had slowed. For many marginal borrowers, the loan worked only if rising home values let them refinance before payments reset. As home price appreciation slowed, then reversed, that option disappeared. Payments rose, defaults increased, distressed sales followed, ultimately leading to the banking crisis.

The Borrower and the Payer

There are vast differences between AI and ’08, but one similarity is the dependence on future payments and continued access to capital. It helps to understand how one part of AI financing works.

A frontier model company or large AI customer commits billions of dollars to a hyperscaler or neocloud for compute. Investors treat those commitments as future revenue and earnings. The cloud providers then use those commitments to buy chips, memory, power, cooling, land, labor, and more. Increasingly, because of the growing capital intensity and price inflation from bottlenecks, the cloud providers borrow against the assets and contracted customer payments.

This is where the distinction matters. In housing, the homeowner was both the borrower and the source of repayment. In AI, those roles are split. The infrastructure provider is the borrower. The AI lab, model company, or general compute customer is the offtaker (a buyer who agrees to purchase a producer’s goods or services before they are produced) whose payments help support the investment.

This arrangement works, but it requires capital to remain abundant and affordable, customer demand to keep growing, and contracted payments to be honored. But backlogs are not cash. Their value depends on the customers’ creditworthiness, the contract terms, and whether the capacity can be redeployed if the customer cannot pay.

If funding conditions tighten, if private-market capital becomes less available, if monetization disappoints, or if customers opt for cheaper models, the promised cash flows may prove less valuable than investors assume. The provider may still owe lenders, landlords, equipment vendors, and power suppliers even as the revenue supporting those obligations weakens.

The largest hyperscalers are cash-rich, so this isn’t likely to be a solvency problem, which is an important distinction from 2008. Their risk is different. The problem is whether the next dollar invested earns an adequate return — the critical KPI for this cycle.

Conclusion

None of this means AI isn’t real. Adoption is real and productivity benefits are likely. But as we’ve written often, investors don’t own productivity, they own cash flows.

Benchmarks own the entire AI ecosystem from model builders, cloud providers, data center owners, scarce suppliers of chips and memory, capital goods beneficiaries, but also potential victims. Similar to 2008, they don’t distinguish between high-quality and weak backlogs, between reusable capacity and stranded capacity, or between businesses earning returns above their cost of capital or those dependent on tomorrow’s financing.

Active managers can. They can underwrite customer concentration risk, contract quality, leverage, debt maturity schedules, pricing power, capital discipline, and so on. As financing becomes more selective and the AI profit pool more diluted, these factors will start to matter more to financial results — and thus asset prices.

AI may become the most important technology of our lifetime. But that doesn’t mean every company funding it will earn an acceptable return on invested capital. The market and the benchmark may price the promise, but active managers will underwrite the payer. 

The information included above as well as individual companies and/or securities mentioned should not be construed as investment advice, a recommendation to buy or sell or an indication of trading intent on behalf of any MFS product.

Keep in mind that all investments carry a certain amount of risk, including the possible loss of the principal amount invested.

The views expressed are those of the author(s) and are subject to change at any time. These views are for informational purposes only and should not be relied upon as a recommendation to purchase any security or as a solicitation or investment advice. No forecasts can be guaranteed. Past performance is no guarantee of future results.