I imagine that many readers will have had to endure anti-money laundering training. I sat through dozens of AML sessions and was often reminded of the unnamed character in The Hitchhiker’s Guide to the Galaxy who survived a poetry reading by gnawing off one of his own legs. The single message that stuck in my mind was that if a transaction looked too good to be true, it was quite possibly suspect.
This maxim is often true of vendor financing, if the terms look suspiciously generous. There is nothing inherently wrong with lending your customers the money to buy your products – that’s how most new cars are sold – but there should be an underlying business and financial logic.
An estimated US$500bn will be invested in AI data centres in 2026 alone, with about half of this being spent on Nvidia processors. It’s not obvious whether this investment will deliver a satisfactory return, but it is increasingly obvious that markets are getting nervous about the sheer scale of spending. I still remember the tech bust of 2000, and the mood music today is similar – to me at least.
Even the richest tech companies are turning to unusual structures to fund investment
The pre-2000 bubble was largely financed by public equity and bond markets, leaving both retail and institutional investors with huge losses when the bubble burst. Trillions were invested in fibre-optic cabling, and tech IPOs often had big first-day pops. The expected returns never materialised and total losses were roughly US$5 trillion to US$10 trillion, depending on the dates you choose.
Rise of the SPV
What I find fascinating about today’s AI data-centre boom is the source of the financing and the opacity of the related transactions. Locating robust numbers is not easy. I have made heavy use of AI to trawl the internet and unearth some of the plumbing, but the results are indicative, not precise.
Even the richest tech companies are turning to, shall we say, unusual structures to fund this scale of investment. Just like ships, many of these data centres are individually funded via special purpose vehicles (SPVs), each with its own features. One common feature is minimal public disclosure – both of the accounts and of the full credit rating analysis. Another is the non-appearance of these SPVs on any listed company balance sheets.
Trying to analyse the true scale of exposure and risk is virtually impossible
A typical SPV will have a blend of equity (20%-35%) and debt (65%-80%) financing, supported by user-lease commitments, residual-value guarantees, revenue guarantees and straight debt guarantees. Some of these numbers will be visible to investors but others, such as residual-value or revenue guarantees, are harder to trace. Diligent credit and equity analysts may be able to incorporate these risks into their reports but the exercise is neither robust nor straightforward.
Sting in the tail
The two features that particularly caught my eye were the residual-value guarantees and the financing provided by the equipment vendors, primarily Nvidia. When lease deals go wrong, residual-value guarantees often provide a nasty sting in the tail, especially if there is a widespread surplus of the assets in question. This applies equally to simple things like cars and to complex things like data centres.
Unpicking Nvidia’s exposure to data-centre financing is even trickier. It has direct equity investments, partnerships with investment banks and private equity, debt guarantees and more. In July Bloomberg published a report stating that Nvidia was working on a fresh round of US$750bn data-centre financing.
Let me be clear: I am not stating or implying that some sort of bubble implosion is likely or imminent. What I am saying is that trying to analyse the true scale of exposure and risk is virtually impossible. Retail investor involvement may be minimal but the numbers are so large that any collateral damage would be widespread.
