Artificial Intelligence – Debt Debate

Buying AI Infrastructure with debt is a bad idea.

• It looks like all of the hyperscalers who have decided to massively increase capital spending on AI data centres will finance a good proportion of their investment with debt, making the AI boom look more and more like the internet bubble with every passing day.
• Of all the big AI infrastructure spenders of 2026, Alphabet is the one that can actually afford to finance the entirety of its profligacy with cash flow, but it appears that the financial engineers inside the company have decided that debt is a good way to go.
• Normally, they are correct because what all sensible companies and financial advisers will tell you is that a balance of financing with debt and equity is the most efficient way to finance a company.  
• This is because debt interest is deductible from profit before the calculation of corporation tax, making it a cost-efficient method of raising money for investment.
• Furthermore, if one is a large, highly cash-generative company (like the hyperscalers are), then the market will lend one money at an interest rate that is cheaper than the cost of equity.
• This is why, when one looks at how a company is financed (equity and debt), any reasonable person will argue for a blend between the two to optimise the cost of capital and ensure that the company is financed as efficiently as possible.
• I am pretty sure that these arguments are being made in the finance department of Alphabet, and given the demand it has received for its bond offering, it will be able to issue debt at an attractive cost to itself.
• However, it was raising debt to finance the build-out of the internet that caused the internet bubble to be as painful and damaging as it was when it turned out the users did not show up when they built it.
• This time around, the users are coming faster than the companies can build the infrastructure, but unfortunately, hardly any of them are paying to use the compute that is being built.
• RFM estimates that OpenAI has around 1.4bn monthly users (it reports weekly users) of which at least 900m have never paid the company any money whatsoever.
• As a result of this free tier, all of the competitors have had to follow suit, and so now the vast majority of the compute upon which hundreds of billions is being spent has no revenue attached to it.
• Furthermore, when one looks at where OpenAI and CoreWeave do earn revenue, one can quickly conclude that the business model of selling compute is currently compromised and needs to be revised, if any profit at all is going to be generated.
• Against this shaky situation, large amounts of debt are being raised, but the problem with debt is that it does not go away when things go wrong, and it is impossible to write down or write off without the agreement of those one owes money to.
• This means that if the business model of AI is not fixed, then the levels of debt that are being built up in the system mean that the reset will be harder and more painful than it otherwise could have been.
• Given this level of uncertainty, if I were Google CFO, I would finance the whole lot through cash flow from operations, meaning that if things go wrong, I will still have the balance sheet to withstand anything.
• However, the industry is taking the alternative view and is gearing up the balance sheet to pay for some (or in some cases) all of the spending, which I think raises the risk materially.
• At some point, the markets may decline to continue lending money, and it is at this point that it will be those with rock-solid balance sheets and a cash-generative business that will survive.
• Hence, this is not a race over who can get to artificial super-intelligence first, but who has the stiffest backbone to survive the monumental hangover that will surely follow this huge round of spending.
• More than ever, the place to be is in the beneficiaries of this spending, where the best direct investment remains Nvidia, but I continue to prefer Samsung Electronics for the memory super cycle and Qualcomm for inference at the edge.
• Micron, SK Hynix, Broadcom, TSMC AMD and several others are also likely to continue faring well while the current environment lasts.