A Nikkei Asia investigation uncovered $1.65 trillion in hidden liabilities for Alphabet, Microsoft, Amazon, Meta, and Oracle—eight times higher than four years ago—from leases, GPU contracts, and ventures kept off balance sheets.
I have touched this earlier on META data center buildout discussion. Interesting to see this report released right at earnings season. So, this will be raised by analysts during the call, and we need to see how managements respond.
Of the five, $ORCL’s debt is worrisome. The company took lot of debt to do buybacks at high valuation/ price (basically buying back Larry’s shares) and has much lower FCF compared to others.
S&P Global Ratings cut Oracle’s credit rating to BBB- on July 9, one notch above junk status, citing extreme concentration risk from OpenAI, which represents roughly half of Oracle’s $638 billion backlog, alongside a projected free cash flow deficit of $42 billion in fiscal 2027.
I still think $ORCL will be the first to announce scaling down their DC buildout… but it may not be a decision taken from the position of strength but $ORCL’s hands are forced to make such decision. That would be a shame…
Since this reminded me alarmingly of the off-balance-sheet liabilities of the banks in 2008 I asked Gemini about it.
What These “Hidden” Liabilities Actually Are
These are not illegal or fraudulent hidden debts, but rather contractual commitments disclosed in financial report footnotes under US GAAP/IFRS accounting rules:
Uncommenced Leases: Long-term leases for data centers and facilities currently under construction that do not land on the main balance sheet until the asset is officially handed over.
GPU & Power Procurement Commitments: Multi-year purchase agreements for chips (such as Nvidia GPUs), specialized server hardware, and long-term Power Purchase Agreements (PPAs) with energy producers.
Special Purpose Vehicles (SPVs) & JVs: Off-balance-sheet corporate joint ventures created with private equity partners to fund massive data centers (e.g., Meta’s $50B data center venture with Blue Owl Capital in Louisiana).
Residual-Value Guarantees: Contingent liabilities where tech companies backstop financing for partners or infrastructure providers.
The critical distinction is that Big Tech hyperscalers (like Microsoft, Alphabet, and Amazon) generate massive organic cash flow and maintain clean balance sheets. However, the structural similarity lies in understating economic leverage, making companies look less capital-intensive than they actually are.
1. Shift from “Asset-Lite” to “Asset-Heavy” Models
Tech giants historically commanded high price-to-earnings (P/E) multiples because software and digital ads required minimal capital expenditure to scale. As these companies transform into utility-scale infrastructure providers, Return on Equity (ROE) and free cash flow margins will face long-term downward pressure, potentially requiring valuation multiples to compress.
2. Rapid Hardware Depreciation Risks
Unlike real estate or toll roads, AI infrastructure decays at a rapid pace. Advanced GPUs and hardware become economically obsolete in 3 to 5 years as faster, more efficient chips emerge. If end-user AI revenue fails to scale fast enough to match these fixed lease and purchase obligations, tech companies will be forced to take massive asset impairment write-downs.
3. Credit Disparities Across Tech Giants
The risk exposure is not evenly distributed:
Low Risk (Fortress Balance Sheets): Microsoft, Alphabet, and Amazon hold vast cash reserves and highly diversified revenue bases that can absorb lease commencement costs as they hit the main books.
Elevated Risk: Companies with higher relative leverage ratios or heavy reliance on single-client ventures (such as Oracle or secondary cloud providers) face higher risk of credit rating downgrades or liquidity squeezes if utilization rates lag expectations. [end Gemini quote]
Little by little, these expenses will transfer from footnotes to the income statement. Free cash flow will decline since the bills must be paid.
The “Residual Value Guarantee” reminds me of Proverbs 22:26-27: 26 Don’t agree to guarantee another person’s debt or put up security for someone else.
27 If you can’t pay it, even your bed will be snatched from under you.
I don’t think Alphabet, Microsoft, Amazon or Meta will have to worry about their bed being snatched. But the shareholders should be concerned that these companies will have lower cash flow and profits if the footnoted deals go wrong and if the end-user customers don’t buy the forecasted amount of AI tokens.
The P/E multiples of Alphabet, Microsoft, Amazon and Meta are high because they have traditionally been software stocks with high earnings based on relatively low capital costs. But once they start building huge data centers they become more like debt-heavy factory based utilities. Especially because their chips will have fast depreciation. Utility P/E ratios are much lower than tech P/E ratios.
If the monetization of AI tokens lags behind the massive capacity being built, the tech giants won’t go broke, but their shareholders will likely have to settle for utility-like returns on tech-like risks—a trade-off that is definitely not priced into today’s market multiples.
The issue is not about solvency, but the rapid raise in off-balance sheet items. They are future liabilities. If your RoE is less than your multiples have to go down… And in the case of $ORACLE… while it will take a lot to kill that franchise… I remember folks said the same about Lehman… I know, I know… don’t view everything from Lehman lense… but…
At the height of dotcom bubble, the Lucent’s of the world also had similarly strong balance sheet with some vendor financing and relatively lot less off-balance sheet items. Setting aside that…
Today, everyone is priced like an AI winner. But history tells us, that there will be a shakeout and the survivors may reap outsized gains and some/ many will be bought out/ go bankrupt/ become totally irrelevant, etc.
So, $GOOGL may have a cash generating machine, tons of SW/ AI IP, chips, cloud, frontier model, etc… $ORCL Is not same as $GOOGL. I am not predicting that $ORCL will go bankrupt… I don’t think it belongs to the club of $GOOGL, $AMZN of the world.
Separately, a great game is going on… and I am scared to get into the game, and don’t know exactly how I can protect my portfolio when the music eventually stops.
@Kingran I think we are on the same page. We both know the history of bubbles. We both recognize the danger of the current dynamic. We know that when a bubble pops leveraged speculators sell everything to meet margin calls and the babies (non-AI stocks) will be thrown out with the bathwater (AI stocks). They are all part of the same S&P500 index that will be sold as a block together.
Each of us needs to look objectively at our personal needs and risk tolerance.
I’m a thrifty 72 year old married LBYMer whose household expenses are more than covered by Social Security and income from a bond ladder of A rated and better. I do hold some dividend-yielding stocks because the dividends are tax-advantaged over bond income - and because I could be wrong.
Unlike @intercst I am not interested in maximizing my net worth if that would mean increasing risk and volatility in my portfolio. In a nutshell, I need calm more than I need more money.
I do plan to step in with a nice, big sponge when the music stops to mop up blood in the streets. Patiently waiting since that will take at least a year, probably more.
Of course, everyone is different. As long as a person understands the risks and their own personality – and doesn’t complain when the trade goes against them.
Some of the media discussions seem to suggest that these companies are all fully backstopping these data centers in an ironclad way, just as if they had issued corporate debt. I’m not entirely sure that’s the case, though. Structuring the debt financing of data centers this way doesn’t just keep the debt off their books…it also somewhat increases the chances that the lenders, and not the borrowers, get crushed in a bubble collapse.
The ones who would really take it in the wobblies first (well, after the equity holders in the data center SPV’s) aren’t going to be the folks in that cute little chart. It’s going to be the private credit funds (Blackstone, Blue Owl Capital, Apollo, Pimco, and BlackRock) and all those big insurance company limited partners they work with. They’re the ones who are buying up all this debt being issued by the SPV’s. They’ll try to get a lot of that from the hyperscalers via these guarantees, of course. But that’s commercial litigation over contracts, not debt collection - and if any of that paper isn’t exactly right, look for the hyperscalers’ very good lawyers to fight with the private credit funds’ very good lawyers for a while, and for that debt to be settled at far less than par. And of course, a lot of that paper has already started being carved up and sold into other offerings that are landing in retirement accounts and pension funds.
I don’t know. If things go south, I think I’d be even more scared as a shareholder in Blackstone or Apollo than MSFT or AMZN. Or more specifically, the big PE-affiliated insurance companies that actually provided the money, and won’t get paid back in full. That, and the debt’s going to be so widely distributed throughout the ecosystem to a lot of pension funds and private investors that it might be really, really bad. This doesn’t just stay with the companies that are doing the borrowing - if the debt goes bad, the pain will be felt by the debt owners, not just the borrowers.
@albaby1 thanks for making this excellent point. (I gave you a rec.)
That’s exactly what happened in 2000 when Global Crossing went bankrupt after laying the optic cables for the internet. The lenders were devastated.
Wendy
I think the very fact that the funding and lending and everything associated with this being this strange and convoluted and circular and seeming to be engineered to hide things is a GIANT HUGE RED FRACKING FLAG.
So far, big banks and even smaller ones have limited exposure. Look at the financing of these buildouts, they are primarily PE, equity and debt issued as bonds. When the bubble pops the scale of disaster may be similar to the past, not necessarily the past players will be hurt in the same scale. Banks cannot avoid it but it may not be in GFC scale.
This is no different than in the past. Even the 8x increase is primarily driven by the amount of capex going into DC buildout and not necessarily because suddenly companies are trying to hide things.
These are highly sophisticated, very short-term vehicles. My personal view is most investors have no need for these instruments and should avoid it.
I can try to hedge, but the only true hedge is going to cash. But when things turn the first sell-off will be so violent, it may be practically impossible to get out in quick without suffering a low double digit drawdown.
I was thinking about both these statements… some random thoughts. First of all the rally can go much further than we can imagine, or want. There are folks who are still comparing this to 1996/1997 moment and we have ways to get to 2000… and there are others who have little shorter timeline…
So depending on your energy, trading requires lots of energy, risk tolerance, where you are in life, you can play the current market differently.
One of the advantages of getting old is, you become patient, you are not in hurry and you learn to wait. Waiting is fine, just make sure you have the resolve, courage to pull the trigger when the conditions are ripe. Also, make sure, don’t get sucked into the initial decline. There are lot of folks who were conditioned to buy the freaking dip. Until those folks realize significant losses or lose confidence, the market will not bottom. So patience is going to be the key. None of this is going to happen tomorrow, may take 2, 3, or more years.
My own expectation when this bubble pops QQQ can give back as much as 50% to 65%, i.e., still only getting back to 2022 October level.
The US government, since Reagan, has offset who will pay the price.
Right now, the global economy is falling apart, particularly China and Russia.
We are running with countercyclical economics.
Those of us posting here need to recalculate whether the US can ride this out. It will be a major dose of inflation if the country manages it. It is a bifurcation of the BRICs versus the West. The BRICs will break up, all but in name.
Ah, but the same thing I mentioned in my first post still applies. You can’t just look at where the nominal obligation - the debt - is being held today. You have to see where the risk is living, and it’s starting to get moved out. And in two or three years, it will have spread far beyond private equity.
If the bubble were to pop right now, I think you’d be right. Most of the risk after the data center equity owners get taken out is still living with the big private equity shops and their insurance company LP’s.
However, those guys aren’t just sleepily holding all this stuff on their books. Just as with the lead up to the financial crisis, they’re slicing and dicing and moving the risk around to improve their position. It’s being securitized and sold off. Much of that is just going out as asset-backed securities and corporate bonds. Private equity issues about 90% of data center debt, but only holds about half of it - the rest has been sold out to the public debt markets and are now owned by mutual funds and state retirement plans and the usual holders of corporate debt.
And more scarily, we’re starting to see some of it going out as synthetic derivatives that simply repackage the risk and get it off their books. They might call it SRT (Significant Risk Transfer) instead of a CDO or a CDS, but the idea is the same. The folks originating all this debt to fund data center construction want to get some of the default risk off of their books, and spread deeper into the wider economy. Sounds a bit familiar, no?
Again, if there were a massive collapse in AI now, it there’s a good chance it wouldn’t be a Great Recession problem. It would more likely mirror the 2000 collapse. Direct AI data center debt is now roughly on par with the $500 billion in debt to internet companies (mostly telecomms) that triggered that collapse. The economy is much larger, but there’s also a fair amount of shadow liability associated with this stuff, as noted in the above article (around $1.5 trillion), so we’re probably at the same order of magnitude. And the debt risk is still very concentrated in the lender side - again, half of it is still on the PE shops’ books, and only a tiny bit of the risk has been spread through weird synthetic instruments.
Things get more squirrelly if we go another two or three years according to plan (from the AI folks perspective) and then there’s a collapse. If we actually get to the point where we’ve put $4-5 trillion of debt financing into building AI and then it collapses, it will be “look out below” for the broader economy.
Of course, the absolute magnitude will be higher - it would be bigger than the absolute size of the subprime mortgage market in 2008, though again the economy is bigger. And the debt issuers will have had two or three years to move a lot of this off their books into the larger economy. Either directly (selling the debt or asset-backed securities) or indirectly through new synthetic derivatives like the SRT’s. Because there just isn’t enough money in private equity alone to fund the three-year plan of these companies. You have to start tapping much more of the financial sector if you want to get to $5 trillion. And that process is just now getting started.
