08-31-2026 
|
Getting your Trinity Audio player ready...
|
The views expressed in this article are those of the author and do not necessarily reflect the position of CoinGeek.
A short clip has been making the rounds on X in the last week of August 2026. Ricardo (@Ric_RTP) posted 161 seconds of Ed Zitron, the tech industry’s most relentless bookkeeper of artificial intelligence (AI) losses, walking through the actual economics of the AI industry, and it is that rare clip where the scary numbers check out. The AI bubble’s real problem is not the chatbots or the science; it is the infrastructure financing. The industry is paying for this buildout as if it were long-lived infrastructure (twenty-year buildings, multi-year debt), while the productive asset inside the building is silicon with the shelf life of a gaming laptop.
- The AI bubble’s real problem is its financing
- Apple found a way to monetize depreciation
- When does the AI math finally work?
- Didn’t Anthropic prove AI can be profitable?
Railroad iron outlived its bankruptcies, and so did telecom fiber. GPUs will not.
Here is the post:
The numbers Zitron is working from are not made up; they come from OpenAI’s audited 2025 financials, leaked to Zitron and independently verified by the Financial Times, showing $13.07 billion in revenue against $34 billion in total costs and expenses, resulting in a net loss of $38.53 billion. That is roughly seven and a half times the $5.09 billion OpenAI lost in 2024 on $3.7 billion of revenue, and the R&D line alone ran $19.18 billion, more than everything the company earned.
And the compute bill is the monster in the basement: by the leaked documents’ own accounting, $17.2 billion has flowed from OpenAI to Microsoft (NASDAQ: MSFT) for Azure capacity, more than OpenAI’s entire annual revenue, handed to a single vendor.
Microsoft, for its part, paid OpenAI $303 million.
Cash making round trips, not revenue…
One correction to the clip, in fairness: its $47 billion run rate for Anthropic is already stale, because Bloomberg reported on August 17, 2026, that the run rate topped $65 billion in late July, so it’s been a sevenfold jump since the end of 2025. The bulls will tell you that pace proves the rockets are real.
That pace is the best card in their deck, and I am going to play it face up before the end of this article, so strap in.
First, the machine under the machine.
I made the long version of this argument in my data center piece in May 2026: the boom is financed on twenty-year assumptions against hardware with a three-to-five-year useful life. When Penn Central collapsed, the rails stayed in the ground; when WorldCom and Global Crossing imploded, the fiber stayed in the conduit, and a generation of cheap internet was built on the carcass. The creditors ate their losses, but the asset outlived the balance sheet. Unfortunately, a GPU will not extend that courtesy to anyone.
The scale is the part that is probably keeping bond underwriters up at night: the five big hyperscalers (Amazon [NASDAQ: AMZN], Microsoft, Google [NASDAQ: GOOGL], Meta [NASDAQ: META], Oracle [NASDAQ: ORCL]) are forecast to spend over $600 billion on capex in 2026, roughly 36% more than 2025, and they raised $108 billion in debt during 2025 to help cover it.
Goldman Sachs (NASDAQ: GS) projects global AI investment at over $1 trillion in 2026.
Some of the demand underneath all that spending is synthetic: buried in an SEC filing in the fall of 2025 sat a deal signed quietly back in 2023, obligating NVIDIA(NASDAQ: NVDA) to buy CoreWeave’s UNSOLD cloud capacity, up to $6.3 billion of it, through April 13, 2032.
NVIDIA is CoreWeave’s chip vendor, its investor, and its customer of last resort, all at the same time. I have been known to sarcastically say “probably nothing” about various things. This one is hard to chuckle about because this time, it is a lot less market-clearing and more a gallery owner bidding on his own paintings to keep the appraisals up.
Apple just turned depreciation into a product
Now watch what the one company that truly understands hardware depreciation is doing while the rest of the economy burns money.
Apple’s (NASDAQ: APPL) new Apple Upgrade program lets you lease an iPhone, iPad, Watch, or Mac through Klarna, with Mac terms of 24 or 36 months. A $1,999 MacBook Pro becomes $38.99 per month on a 36-month lease, and at the end, you can upgrade, walk away, or buy the machine out. The enthusiast debate on Reddit is mostly about who the lease math favors, and Macworld’s verdict was that it “makes getting a new Mac easier but not necessarily cheaper,” because the leaser hands Apple the resale value.
That last clause is the whole story, because Apple just converted hardware depreciation into a profitable subscription for the company. It can absorb the residual-value risk because Apple silicon holds resale value and Apple controls the refurbished channel that recaptures it. The hyperscalers are carrying the same depreciation risk on GPUs that hold nothing.
Same risk, opposite balance sheets.
Curiously, the Mac mini is not on the lease list. MacBook Air, MacBook Pro, iMac, Mac Studio: all leasable, while the mini stays a buy-outright box. And the box is not getting cheaper, because Apple raised the mini from $599 to $799 in June 2026, then announced the new M6 mini on August 25, 2026, starting at $899. Two price hikes in one year, on the one Mac Apple will not lease you. I argued in my Apple balkanization piece that the mini is becoming the sovereign-agent box under your desk, and Tim Cook himself has said that mini and Studio shortages were driven by AI developers running agent platforms at the edge.
Apple will rent you the fashion object, but they expect you to own the useful infrastructure for your AI productivity.
So how does the math ever close?
There are only three ways; I have run this from every direction I know, and every path home runs through a miracle, so we should all start praying.
Miracle one is hardware: an unprecedented acceleration of Moore’s Law on the AI silicon side, chips so much better that a dollar buys multiples of today’s compute. Possible! NVIDIA ships a new flagship roughly every year. But the trap is built in: every leap reprices the mountain of silicon you already bought toward zero, faster, while training costs per model generation have been rising three to five times rather than falling.
A hardware miracle rescues next year’s buyer and executes this year’s balance sheet.
Miracle two is software: unprecedented efficiency gains, more work per token and per watt, and this one is genuinely underway. Anthropic’s compute cost fell from 71 cents per dollar of revenue to 56 cents in a single quarter, per the Wall Street Journal’s figures, and inference prices collapse year after year. But efficiency cuts both ways: the capex was underwritten against revenue per token, and efficiency shrinks that very metric. Jevons paradox says cheaper compute summons more total demand. Maybe. Jevons is a hope; the depreciation schedule is a contract. And no efficiency curve fixes the data problem: models feeding on model output while the provenance of the training corpus decays.
Miracle three is the real economy: AI productivity gains actually showing up as revenue, at the scale Bain has said the industry needs, roughly $2 trillion in annual AI revenue by 2030 to justify the math. When I wrote the data center piece in May 2026, AI product revenue sat around $45 billion. Billion, with a B, against a two-trillion target, in four years.
Name one industry that has ever crossed that distance on schedule.
I’ll wait…
Now let’s shuffle the deck, stack them with some more luck and prayers, and be a little more than generous with converging probabilities: chips improving on their historical curve, efficiency climbing the way Anthropic‘s just did, real revenue growing faster than any enterprise software wave in history. Run all three together, and the gap still does not close. As far as any audited number in the public record can show, it does not even get close: not one leaked financial statement, not one SEC filing, not one verified quarterly figure in front of us makes this math work.
If somebody is holding the spreadsheet that does, please publish it, because the industry insisting the miracle already happened keeps declining to show receipts.
Didn’t Anthropic just prove it works?
Anthropic just projected an operating profit. Doesn’t that prove the model works?
Here’s the best card I could find in their deck, face up on the table.
Two real numbers first: the Wall Street Journal reports that Anthropic projected its first-ever quarterly operating profit of $559 million for the second quarter of 2026, and Bloomberg reports a $65 billion run rate. Serious figures, from serious reporters.
Ed Zitron read the same numbers and argues the margin turn leans on a discounted SpaceX compute deal, roughly $1.25 billion a month with reduced fees during the ramp, plus a non-GAAP presentation and run-rate claims that do not reconcile cleanly with the quarterly ones. Read him and decide for yourself.
Here is mine: one projected quarter at one company with disputed accounting does not underwrite a trillion dollars of industry investment this year. And notice where the profit came from: cheaper compute. The bulls’ best exhibit is miracle two firing, and miracle two shrinks the revenue per token that everyone else’s capex was underwritten against, even while it fires.
Their proof of health is my mechanism of collapse. Heck of an exhibit.
So here is where I land, and it is not the same place where the pure doomers land.
Humans are industrious, clever, and calculating. We built our way out of the Dark Ages, the Dutch collapse, the British Empire’s wind-down, the railroad bust and the fiber bust, and every overbuild since the canals, because the species has a habit of turning its wreckage into foundations.
Do not bet against people building their way out of this hole in the long term. I will not.
But the railroads left rails, and the telecoms left fiber; those bankruptcies endowed the future. This bust would leave warehouses of five-year-old silicon nobody wants, and twenty years of debt everybody still owes. The next boom would inherit nothing but the empty buildings…
Miracles happen. I believe that more literally than most people writing about technology. But nobody sane signs a twenty-year mortgage against three miracles landing at once.
This opinion piece is published to encourage discussion. The author’s views are their own and do not constitute legal, procurement, or policy advice, nor do they represent the positions of CoinGeek or its partners.
In order for artificial intelligence (AI) to work right within the law and thrive in the face of growing challenges, it needs to integrate an enterprise blockchain system that ensures data input quality and ownership—allowing it to keep data safe while also guaranteeing the immutability of data. Check out CoinGeek’s coverage on this emerging tech to learn more why Enterprise blockchain will be the backbone of AI.
Watch: AI is becoming a game-changer for iGaming
