On August 10, Nvidia signed memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize more than $500 billion in third-party capital for AI data centres. Almost none of that money will ever appear as Nvidia revenue risk, as hyperscaler capex, or as a single digit in the S&P 500’s forward P/E.
That is not a rounding error in how the AI boom is financed. That is the whole story. AI infrastructure financing has quietly migrated from Big Tech’s balance sheet into private credit, insurance portfolios and securitized paper — and the people arguing about whether AI is a bubble are still staring at the equity market, which is now the wrong instrument.
What actually happened on August 10
Jensen Huang described the move plainly: this is “really the first time that technology chips have become an investable asset class.” Nvidia’s six partners will build compute financing platforms that let hyperscalers, frontier labs and enterprises acquire GPUs and build data centres without funding it from their own cash flow. Nvidia holds an option to backstop up to $125 billion — 25% of the potential deals.
Read that backstop number again. A chip vendor is offering partial credit support on the purchases of its own products. In any other industry we would call that vendor financing and immediately start asking how much of the demand is real.
This is genuinely useful engineering, and it is worth saying so. Compute is a productive asset. Financing it over its economic life instead of paying cash upfront is exactly what capital markets are supposed to do — it is how railways, fibre and the electrical grid all got built. The problem is not that the structure exists. The problem is where the risk lands, and who has been told about it.
The buildout was always going to outrun the balance sheets
Amazon, Alphabet, Meta, Microsoft and Oracle are on track to spend somewhere around $660–690 billion on capex in 2026 — Amazon near $200 billion, Alphabet guiding $195–205 billion, Meta $115–135 billion, Microsoft above $120 billion, Oracle around $50 billion. That is roughly 70% above last year’s already-record $410 billion.
Even these companies cannot self-fund what comes next. Morgan Stanley projects global data-centre investment of about $2.9 trillion through 2028, with roughly $800 billion supplied by private credit, $200 billion by corporate bonds, and $150 billion by securitized products. Add those three together and about 40% of the buildout — $1.15 trillion — arrives as debt raised outside Big Tech’s own balance sheet. That capital has to come from somewhere, and the “somewhere” is pension funds, insurers, and increasingly retail-accessible credit vehicles.
The part that should make you uncomfortable
The migration is already well underway. Meta, Oracle and xAI have moved more than $120 billion of data-centre debt off their balance sheets through special purpose vehicles. Meta’s Hyperion facility in Louisiana is the template: a $30 billion structure called Beignet Investor, owned 80% by Blue Owl Capital and only 20% by Meta, funded with roughly $27 billion in loans from Pimco, BlackRock, Apollo and others. The data centre is Meta’s in every economic sense. The debt is not Meta’s in any accounting sense.
Scale that logic up and you get the number that should be on every investor’s wall. At year-end 2025, the five largest US hyperscalers held $662 billion in undisclosed future data-centre lease commitments outside their reported balance sheets — equal to 113% of the same five companies’ combined adjusted on-balance-sheet debt. The invisible pile is now bigger than the visible one.
These are not fragile companies. Microsoft and Alphabet could service far more leverage than this. But “the borrower is strong” is a different claim from “the risk is visible,” and only one of those is true right now. Private credit lending to AI-related companies has gone from roughly zero to over $200 billion in about three years. We wrote about the early stage of this in The AI Infrastructure Debt Boom Is Here. It has since stopped being a boom and started being the primary funding channel.
The contrarian point: you can be right about AI and still lose money
Here is the consensus framing, and here is why it is wrong.
Consensus says the AI bubble question is a valuation question. Watch Nvidia’s multiple, watch hyperscaler free cash flow, watch whether the revenue shows up. If AI delivers, the trade works.
The structure of the financing says something different. Equity risk and credit risk are not the same risk, and they do not resolve on the same timetable. Equity is patient — a stock can hold a valuation for years while a thesis matures. Credit is not. Credit has coupon dates, covenant tests and refinancing windows. A data centre that earns its return in 2031 still owes interest in 2027.
The market is already pricing this. Coverage ratios on hyperscaler bond offerings collapsed from close to five times in February to under two times by mid-July. That is bond buyers quietly saying they have absorbed as much AI duration as they want. Meanwhile, total AI-related debt is heading toward $570 billion, and data-centre securitization issuance is projected to reach $30–40 billion a year.
So the honest bear case for 2026 is not “AI won’t work.” It is narrower and far more plausible: AI works, and the financing structure built to fund it breaks first anyway — because the assets are long, the debt is short, and the depreciation schedule is a guess.
The depreciation guess that underwrites everything
All three major hyperscalers now depreciate server assets over six years, up from three to four — a change that collectively removed about $18 billion a year from reported depreciation expense. Michael Burry has argued the real economic life of an Nvidia-class training GPU is closer to two or three years given the one-year product cadence, and that understated depreciation from 2026 to 2028 will total roughly $176 billion, overstating profits by more than 20%.
Reasonable people disagree, and the counterargument is decent: older GPUs cascade down to inference and cheaper workloads rather than becoming scrap. But notice what the disagreement is actually about. It is a dispute over the useful life of the collateral backing more than a trillion dollars of new debt. We covered the mechanics in The $200 Billion Line Nobody Reads. Nobody will know who was right until the 2028 refinancing wall arrives.
You are probably more exposed than you think
The retail investor’s instinct is that private credit is somebody else’s asset class. It is not, and it stopped being so a while ago. Non-traded BDCs, interval funds, insurance annuities and pension allocations are all channels into exactly this paper — which is why several large non-traded BDCs and interval funds capped or slowed withdrawals during the redemption spike in late 2025 and early 2026. Software and tech exposure already runs near 26% of direct-lending portfolios. Gross fundings among the top publicly traded BDCs fell to about $5.7 billion, the lowest in two years.
If you own a diversified portfolio, an annuity, and a public pension entitlement, you have AI infrastructure credit exposure. Nobody sold it to you as an AI bet. It was sold to you as yield. This is the same mechanism we described in You Didn’t Buy the AI Trade. Your Bond Fund Did. — only now it is roughly ten times bigger.
What to actually do about it
Not “sell AI.” That is a lazy conclusion and probably a wrong one.
Find your real exposure. Look through your bond funds, target-date funds, any interval fund or non-traded BDC, and your annuity’s general account. The question is not “do I own Nvidia,” it is “how much data-centre credit do I own, and at what seniority?”
Price liquidity properly. The single dangerous feature of this cycle is the gap between a fund that promises quarterly redemptions and assets that take a decade to earn out. Redemption caps are the visible symptom, and they appear exactly when you most want your money.
Distinguish the two AI trades. Owning the equity of companies that will earn from AI is a growth bet with unlimited upside and patient capital. Owning the debt that funds their suppliers’ buildout is a bet you get paid par and nothing more, in exchange for taking the first loss if the depreciation assumption is wrong. Those deserve very different position sizes.
Watch the coverage ratios, not the stock price. The two-times number is the most informative figure of this quarter, and almost nobody reported it.
The AI buildout is real, and the productivity on the other side of it will most likely be real too. But this month the trade changed shape — from an equity story about who wins, to a credit story about who is holding the paper when the collateral gets repriced. The index will not tell you when that happens. Your fund documents might.
Want to track your own path to financial independence? The Wealtharian Wealth Tracker lets you monitor your net worth, FU money progress, and investment milestones in one place. Try it free →