When debt is cheap and optimism is high, the distance between vision and repayment risk becomes dangerously thin.
Last week, two headlines, while worlds apart, were cut from the same cloth. Sam Altman of OpenAI declared his intent to spend trillions on AI infrastructure and at the same time, corporate bond spreads fell to record lows. Both signal that optimism is firmly in the driver’s seat. Both also rest on a less glamorous foundation: debt.
Spreads at record lows
Today, the average US corporate bond credit spread is just 0.75%. That’s an all-time low. It’s roughly half the 25-year average and certainly much less than the 2-4% spreads seen during periods of stress like the GFC or pandemic. This represents a tiny gain from what could be achieved by lending directly to the government. The margin for error has all but disappeared. At this level, a single default could destroy six years worth of spread income. This message is clear: investors see no signs of stress for corporate balance sheets, no fears of defaults and no hesitation in lending. It is a clear green light for borrowers.
This highlights the importance of understanding the asymmetry between Altman’s visionary ‘trillions’ and the cold reality of credit investing. Altman’s equity mindset is all about upside: a portfolio can afford multiple failures if it captures just one exponential winner. Credit is the inverse. Returns are fixed - principal and interest, less defaults, so the sole question is whether capital will be repaid. Both equity and credit investors may study revenue growth or EBITDA, but their lenses are fundamentally opposed. For credit investors the most important driver is the return of capital, not the return on capital.
This context is vital when considering Mr. Altman’s AI ambitions. At a dinner with reporters (Bloomberg) last week, he was happy to detail the financial future for OpenAI saying “expect OpenAI to spend trillions of dollars”. This comes just months after launching the $100 billion Stargate supercomputer project, an unprecedented project in terms of scale (at least it was at the time). While it is easy to lose sight of such large numbers, “trillions” is the magnitude of the entire US federal budget. It almost boggles the mind.
AI’s growth fuelled by debt
These huge sums will likely be earmarked for the backbone of AI infrastructure: buying Graphics Processing Units (GPUs), building datacentres, and growing energy supply. Of the trillions spent, a good portion is likely to be funded by debt - the same markets now pricing risk at historic lows.
To illustrate this, we can look at the typical funding model for AI assets today. For a greenfield datacentre development, debt sizing commonly sits around 50% to 65% of build cost. Once complete, refinancing of an operating asset often sees debt levels higher - towards 75% of the total value. Inside the datacentre, Coreweave, OpenAI’s key GPU partner, also uses debt. Recent IPO filings suggest that 60-75% of their GPU budget was debt funded. The takeaway is clear, debt isn’t peripheral to AI’s growth, it’s foundational. And while the debt is likely to come with more spread than 0.75%, it is still subject to the same exuberance that’s permeating markets. That will mean hundreds of billions of dollars borrowed, hundreds of billions to be serviced with interest payments and principal repayments.
This provides important context for additional quotes from Altman, where he acknowledged that his comments may draw criticism from economists, characterising it as “hand-wringing”. Pressed further, when asked about financing, he suggested, “I suspect we can design a very interesting new kind of financial instrument for finance and compute that the world has not yet figured out”.
Viewed through the credit investor’s lens, this is a red flag. The history of financial markets is riddled with periods of speculative excess, each producing its own financial language. They are often spearheaded by individuals who are highly incentivised to justify valuations that traditional metrics cannot support. In the late 1990s, equity analysts defended dot-com valuations with metrics like “eyeballs” and “clicks”. A decade later, WeWork presented the concept of “community-adjusted EBITDA”. These terms may have served to sell a dream to equity investors, but they are clearly irrelevant to a creditor when an interest payment is due. The same is true for the AI infrastructure capex boom: You can’t service debt with GPU tokens.
Financial innovation or alchemy?
It’s possible that Altman is referring to the use of asset securitisations (such as ABS), where borrowing is secured against cashflows from various assets. ABS markets already fund mortgages, credit cards and auto loans. There are also many more creative use cases, from music royalties to hurricane insurance losses (if there is a cash flow, there is likely someone trying to securitise it). Indeed, Coreweave uses GPUs as collateral for its securitised debt obligations. The important concept is that while the asset may be novel, the underlying principles remain the same – there must be cash flow to service debt.
ABS works efficiently when the underlying numbers are transparent and predictable. This is why houses and cars are so easily financed with ABS: the history of payments and values are well established. The opposite is true for AI infrastructure. A new chip generation (e.g. Nvidia’s Blackwell) or new AI model (e.g. Deepseek’s R1) could instantly invalidate the economic assumptions underlying a billion dollars’ worth of ABS secured by GPUs.
At this stage, GPUs simply don’t fit the mould. As an ABS asset class, they’re too new, too dynamic. For example, Coreweave depreciates GPUs over six years – longer than GPUs have even been generating meaningful AI revenue. The risk isn’t in structuring creatively; it’s the lack of reliable cash flows to securitise.
Financial innovation isn’t the problem, capital markets evolve, just like AI technology. The issue is when we are distracted and begin to ignore the basics. When markets are priced to perfection, shiny new investment paradigms, or exotic products can be alluring. But that is precisely the time to become cautious. Altman is right to seize the opportunity to fund his company’s vision at the most attractive rates possible. That’s his job. For credit investors, the opposite is true: Their job is to have the self-control to look through flashy headlines and see the underlying economics that will ultimately repay debt.
Altman himself even seems to concede the point. Asked by Bloomberg during the conversation he said “Are we in a phase where investors as a whole are overexcited by AI? In my opinion, yes”. Coming from a visionary like him, it should give pause to underwriters tempted to back the next innovative credit structure.
History shows that transformative technologies can consume enormous amounts of capital long before they generate stable predictable cash flows. This was certainly true for the early internet (dot-com) bubble of the late 1990’s, even though ultimately, the productivity gains did arrive. But they didn’t arrive in time for many investors, or many businesses, saddled with debt, and no cash to pay the interest bill. The productivity gains from AI will undoubtedly also be real, but they are likely to unfold over decades, not quarters. And timing matters: for a credit investor, coupons (usually) begin immediately, and principal repayment usually within a few years.
Testing credit’s discipline
At its core, the problem appears to be a lack of self-restraint, as credit markets begin to yield to an equity narrative. The very fact that credit spreads are so low will naturally push investors into more exotic, more enticing instruments. And while the largest AI-related businesses possess far stronger fundamentals than WeWork; the financing narrative is beginning to rhyme. The environment of low spreads suggests that credit investors are broadly accepting this narrative, adopting equity-like risks but with fundamentally capped returns. The risk-reward proposition is becoming dangerously skewed.
The solution is not nearly as exciting as the AI vision, but it is necessary: discipline. It’s the role of equity to fund dreams; the role of credit is to finance reality. When trillion-dollar promises are combined with the tightest spreads in history, it is time for prudent investors to remember: equity can afford to dream of trillions, credit cannot.
Talk to us
If you have any questions about your investment or would like to make sure you have the right investment strategy to reach your ambitions, get in touch with us – our team are always happy to help.
An amended version of this article was originally published in the NBR on 26 August 2025 (paywalled).
