The Gazillion-Dollar Oops
An update
Three months ago I wrote about the AI spending binge and called it “The Trillion-Dollar Oops”. It felt cute at the time.
I owe you an apology.
I was being generous.
A whole year back, back all the way to December, the big five hyperscalers were spending $405 billion a year and I thought that number was already insane. I said it was accelerating. I said No1 knows what anything costs anymore. That these estimates kept getting revised upward so fast that by the time you finished reading the sentence, someone would have revised it again.
I thought I was joking. But increasingly this dimension has other ideas with reality. I was being prophetic.
The 2025 final tally came in at $443 billion for the top five, up 73% year-on-year. Fine. That’s the world we live in now. But the 2026 projections? CreditSights estimated about $600 billion in November. Then Q4 earnings calls hit and - in their own words - the guidance “blew away” their estimates. They raised it to $750 billion. Moody’s independently pegged it at $700 billion. This is the third consecutive year of growth exceeding 60%. And Moody’s projects $870 billion for 2027.
In what universe is this rational capital allocation? Well... in this one, apparently. Google co-founder Larry Page was quoted saying “I’m willing to go bankrupt rather than lose this race”.
At least someone’s being honest.
Let’s revisit my earlier article and do a company-by-company damage, because this REALLY starts to feel like performance art.
Amazon: $200 billion. Ok, pocket change for Bezos I guess? Consensus had it at $147 billion and he basically said: ‘hold my beer’. This is the same guy that called it “kind of an industrial bubble” just a few months ago [1]. He’s now spending $200 billion on the bubble. I need a drink - and by now probably a consult with AA.
Alphabet went from $92 billion to somewhere between $175 and $185 billion - nearly double 2025 and 50% above what analysts expected. Meta jumped from $72 billion to $115-135 billion. Microsoft is tracking toward $120 billion or more, having already burned through $37.5 billion in a single quarter. Oracle - the company I dedicated an entire section to mocking last time - lifted its capex guidance to $50 billion, more than tripling from 2025. Capex as a percentage of revenue will hit 86% at Oracle this year. Eighty-six percent of revenue spent on infrastructure. Meta’s at 54%. Microsoft 47%. Alphabet 46%. These are utility-company ratios for companies that used to be software businesses.
Each of these companies’ 2026 budget is expected to match or exceed their spending for the past three years combined. COMBINED!
I wrote in December that the spending had exceeded the telecom boom as a percentage of GDP. Well, the hyperscalers’ capex is now set to make up 2.2% of GDP. Goldman Sachs calculated that spending would need to hit $700 billion to match the peak of the 1990s telecom cycle. We just blew past it. We matched the most famous infrastructure bubble in modern history in the space of about eighteen months.
For scale, the Kobeissi Letter calculated that in 2025 alone, tech capex at 1.9% of GDP nearly matched the combined scale of every major American infrastructure project of the 20th century. And 2026 will exceed the lot.
Congratulations? I guess?
Let's talk about Oracle, because Oracle is now my favourite financial horror story - which is saying something, given the competition.
The $248 billion in off-balance-sheet lease commitments I reported last time? That’s grown to $261 billion as of February. Still insane. Still off the balance sheet. Still not reflected in the numbers most investors look at.
But the numbers they CAN see are already bad enough. Oracle finished its fiscal year with $92.6 billion in total debt, then ratcheted that up to $108.1 billion after issuing $18 billion in new notes in September. Since then it’s issued another $43 billion in senior notes, created a $10 billion revolving credit facility, expanded its commercial paper programme, and issued mandatory convertible preferred stock. In February, Oracle announced its intent to raise up to $50 billion in debt and equity financing in a single calendar year. Within days, it raised $30 billion. The order book was “substantially oversubscribed”. Oracle’s $25 billion bond sale alone attracted $129 billion in buy orders.
Five times oversubscribed… For.. this..??
This bubble doesn't need a pin - it needs a mirror.
Remaining performance obligations hit $553 billion, up 325% year-on-year. Free cash flow went from a modest $2 billion deficit to a $10 billion shortfall by November, and management expects it to stay negative until 2030. The stock has fallen over 50% from its September high of $345, wiping out roughly $460 billion in market value.
And in January, bondholders filed a class action lawsuit in Manhattan claiming Oracle knew it would need to raise significantly more debt when it sold $18 billion in bonds last September - and didn’t disclose it. Sure enough, seven weeks later Oracle went back for another $38 billion. The optics, as one analyst delicately noted, “aren’t great”.
Meanwhile, Oracle simultaneously announced a restructuring plan costing up to $1.6 billion. Primarily for employee severance. They’re firing people, being sued by bondholders, watching their stock halve, sitting on $261 billion in off-balance-sheet commitments, and raising $50 billion in new financing - all while telling everyone this is going according to plan.
I want to frame this and hang it on my wall. “In memoriam: fiscal sanity, 2024-2026. Gone but not forgotten. Except by Oracle, who never met it in the first place”.
And Oracle might be a basket case, but in aggregate? It’s not even the scariest part yet.
Hyperscalers raised $108 billion in debt during 2025. Morgan Stanley estimates borrowing will hit $400 billion in 2026. This could push total high-grade US corporate bond issuance to a record $2.25 trillion.
And then Alphabet did something extraordinary. Google - the company sitting on $126 billion in cash - raised $32 billion in debt in less than 24 hours. Including a 100-year bond. Anyone who’ll buy this, will be dead before it matures. Denominated in sterling. At least someone in that boardroom has a sense of humour.
The last tech company to issue a century bond was Motorola. In 1997. Remember Motorola? Exactly. I'm sure it'll be fine.
Alphabet’s long-term debt has quadrupled from $11 billion in 2024 to $47 billion in 2025. Meta’s has doubled from $29 billion to $59 billion. Bank of America reports that capex now consumes 90% of operating cash flow across the hyperscalers, up from 65% in 2025. Share buybacks - the thing these companies used to do when they had money left over - fell 33% last year. There’s nothing left anymore.
Investors demanded record protection against these bonds via Credit Default Swaps. So get this: the market is simultaneously lending these companies hundreds of billions of dollars and buying insurance in case they can’t pay it back.
I'm not saying it's incoherent. I'm just saying insurance exists for a reason.
J.P. Morgan still estimates $1.5 trillion in investment-grade bonds needed over the next five years. Bill Blain, CEO of Wind Shift Capital, looked at the debt binge and described it as an “AI hyperscaler debt-fest” pointing to “late-cycle frothiness in credit markets”.
Late-cycle. Those are the words you use right before you use different words.
So what are they getting for all this money?
Still not much, as it turns out.
The MIT study I cited in December - 95% failure rate for enterprise AI projects - still holds. Nobody’s contradicted it. Nobody’s even tried to contradict it with better data. Gartner found that only one in five AI initiatives achieves measurable ROI. One in fifty delivers what they call “disruptive value”. Seventy-four percent of organizations are breaking even or losing money on AI. IBM’s numbers are even more depressing: pilot-stage ROI of 31% collapses to 7% at scale. Seven percent. Below cost of capital. You’d get better returns from a savings account.
42% of companies abandoned most of their AI projects in 2025. Not “scaled back”. Not “restructured”. Abandoned.
But wait! METR [2] ran an actual randomized controlled trial. Sixteen experienced developers working on 246 real-world coding tasks with the assistance of frontier AI tools.
The developers predicted that AI would make them 24% faster. They even felt ~20% faster. But akshually? 19% SLOWER.
METR tried to run a follow-up study in late 2025 and early 2026. But they couldn’t get reliable data. Why? Because an increasing share of developers now refuse to work without AI, even when paid $50 an hour. The tools make you feel productive. The data has hallucinated a different conclusion.
Meanwhile, a Stanford study found that employment among junior software developers fell nearly 20%. So the tools that are making seniors slower, are getting juniors fired.
Progress.
Meanwhile, the thing these data centres actually run on - GPUs - is collapsing in price. And not gently.
In December I mentioned GPU rental prices had dropped 23%. That was quaint. H100 cloud pricing has now fallen 64-75% from late 2024 levels. From $8-10 per hour to $2.99. On marketplace providers like Vast.ai you can rent an H100 for $1.49 an hour. The chip that was supposed to underpin the AI revolution is now cheaper per hour than a cup of coffee in most Starbucks.
A100s? $0.75 an hour on some platforms.
And it’s about to get worse. Blackwell B200 GPUs are shipping now, which means the entire H-series is sliding toward commodity pricing. Previous-gen flagships typically see about 15% list-price cuts within six months of a new generation launch. The H100 hasn't looked down yet.
Remember that I mentioned that below $1.65/h the rental economics stopped working? Well, we’re way below that on the cheaper providers. The very infrastructure these companies are committing trillions to build is becoming uneconomical whilst they’re still building it.
It’s like when hotel rates are cratering and somehow you decide the solution is to build more hotels.
Any good GPU needs power. So last time I talked about the grid. Guess what? It’s getting worse. The 2026-2027 PJM capacity auction added another 22% on top of last year’s 833% jump, hitting the FERC-approved cap - and would have gone 18% higher without that cap. They project now a full six gigawatt shortfall against reliability requirements in 2027. 1.21 gigawatts used to sound like a lot.
Virginia, America’s data centre capital, is building a 944-megawatt gas plant just to keep the lights on. Data centres are expected to consume 41% to 59% of all Virginia’s electricity by 2030. Up from 25% today.
AEP Ohio then introduced a novel concept: making data centres pay for power they reserve even if they don't use it. Paying for things you order. Revolutionary stuff. The moment that rule went into effect, half the pipeline - 17 gigawatts - vanished. Half the grid crisis evaporated the moment consequences were introduced. I would make a joke but I think the universe already did.
The Stargate project - the $500 billion joint venture between OpenAI, SoftBank, and Oracle that I was mildly sceptical about - has actually moved from announcement to construction. Two buildings are already operational at the Abilene campus since September last year, with six more expected by mid-2026. Five additional US sites have been announced. It’s happening. There are cranes and concrete and 6,000 construction workers in a Texas town that didn’t ask for any of this - and is now dealing with double-digit traffic collisions and residents being run off the road by construction vehicles.
OpenAI claims over $400 billion in investment is committed. They’re planning sites in Michigan, New Mexico, Ohio. They’ve gone international too. UAE - I guess this one is postponed. Norway. UK. Argentina. South Korea.
Sam Altman’s long-term plan, and I quote this from his blog because it deserves to be quoted, is “to create a factory that can produce a gigawatt of new AI infrastructure every week”. A gigawatt. Per week.
A gigawatt is roughly enough to power 750,000 homes.
Per week.
Yup, never one for thinking small that guy.
There are some cracks now though. Tiny ones…
Amazon’s stock fell 8% after they announced their $200 billion capex plan. Microsoft dropped 11% after quarterly results. Oracle has halved from its September high, wiping out a cool half trillion in market value. Funny thing about markets - they eventually ask where the returns are.
Yann LeCun left his role as Meta’s chief AI scientist in late 2025. His argument was that LLM capabilities were fundamentally limited. Great at regurgitating old knowledge. Not so great at coming up with new knowledge. Remember: Meta is about to spend $135 billion on that.
Larry Ellison, Oracle's co-founder, looked out at the smoking wreckage of the last six months and told analysts that the SaaS apocalypse "applies to others but not to us". Others. Not us. I don't know what Larry is smoking, but I want it.
Three months ago I wrote that someone was going to pay for this. That either it would be shareholders or the rest of us. I laid out the circular financing, the Enron comparisons, the 19-year leases for 1-year-obsolete hardware, the 95% failure rate, the grid that can’t cope, the CEOs who admit it’s a bubble while pouring money into it.
Every single number got worse since then.
I called my last article “The trillion-dollar oops”.
I don’t think “trillion” cuts it anymore. We’ve left a trillion behind. We’re in territory where the numbers are so large they’ve stopped being numbers and started being theology.
You don’t analyse them anymore. You believe in them, or you don’t.
…
Amen.
💯💯💯 A.M.E.N. I’m sure to sound like a bumpkin… but how on Earth can AI successfully grow and be ‘profitable’ when the society that “depends” (🤮) on it is going to be so poor that we can’t afford to EAT much less engage in the wonders of regurgitated recursive programming? I don’t have much in the means of capital.. but all that I have is going to old-school AI = Agricultural Intelligence, so that I can somehow get off of the pharmaceutical-industrial food complex that is going to kill us one-way-or-one-way!
",,,the market is simultaneously lending these companies hundreds of billions of dollars and buying insurance in case they can’t pay it back."
I guess people have forgotten how well insurance like this worked when companies like AIG were the underwriters. Can you spell "counter-party risk"?
Makes perfect sense to borrow rather that put perfectly good cash into this. If the whole thing blows up (as it most likely will), they'll be playing with OPM (other people's money). Between now and then, how much do you want to bet that their existing capital gets firewalled off somehow?