AI的萬億黑洞：錢燒進去了，但除了兩家虧錢公司，誰在真正買單？

Original Author: Ed Zitron

Original Translation: TechFlow

Introduction: The Bank for International Settlements (BIS) annual report reveals a truth deliberately ignored by tech giants: Trillion-dollar AI capital expenditures have already surpassed these companies' cash flows and profitability, and the ultimate destination of this money is merely to sustain two massively loss-making model labs, OpenAI and Anthropic. If Microsoft, Google, and Amazon decide to stop their quarterly $30 billion GPU purchases, the entire supply chain will collapse.

This month, I published a two-part in-depth series analyzing the "bubble within a bubble" structure of the AI bubble—from the unsustainable reckless growth of semiconductor companies to the personality cult surrounding Sam Altman and Dario Amodei. On Friday, I will publish the long-awaited "Softbank Skeptic's Guide," which you won't want to miss.

On Sunday, the Bank for International Settlements (BIS) released its annual report, saying a lot of things I've been saying:

In the short term, the ongoing AI investment frenzy raises questions about the sustainability of the current economic expansion. The five largest hyperscalers plan to spend over $1 trillion on AI-related capital expenditures between 2025 and 2026. These commitments exceed these companies' profits and free cash flow, leading some to issue debt to raise additional funds.

It's gratifying to see the central banks' central bank say what I've been saying for years. But this part makes me feel both validated and terrible for the whole world:

Disappointing returns could trigger a sudden withdrawal of funding and transform the capital expenditure boom into a prolonged investment slump, with potential knock-on effects on financial conditions... If hyperscalers slow down or stop aggressive capital expenditure deployment, many borrowers in the supply chain may struggle to replace lost revenue and repay their debts.

Nonsense. Last April, I wrote an article titled "AI Is a Systemic Risk for the Tech Industry," outlining how the failure of one model lab, OpenAI, would have seismic effects on its supply chain, dealing blow after blow to NVIDIA, Oracle, Microsoft, and various new cloud providers supplying them compute power (most notably CoreWeave).

Since then, OpenAI's sticky tentacles have reached further into more aspects of the tech industry, signing deals with companies like Google, Amazon, Cerebras, and Broadcom, while also accepting more investments, including Softbank's massive commitments. Softbank can only fulfill these promises by selling precious shares of companies like ARM and NVIDIA and taking on debt.

The concept of systemic risk has never really left my work. I've spent a lot of time thinking about it over the past year—hence, my writing examines the potential consequences of a pullback in AI spending on the financing side of the industry, particularly private credit and the semiconductor sector.

What the BIS cares about is not a revenue crash—which would happen if hyperscalers "slow down or stop aggressive capital expenditure deployment" as it fears—but a revenue crash where borrowers in the AI supply chain cannot repay their growing debt burdens.

Again, this is something I've repeatedly sounded the alarm about. CoreWeave has been a favorite of this newsletter; in March 2025, I published "CoreWeave Is a Ticking Time Bomb," focusing on the company's overwhelming toxic debt pile and its reliance on OpenAI as a customer.

On a larger scale, we have Oracle—a company I covered in detail in the "Oracle Skeptic's Guide" newsletter.

Unlike new cloud providers like CoreWeave, Oracle is an older company that spent most of its existence selling database and ERP software to some of the world's largest companies and public sector institutions. Oracle pivoted to offering AI compute power when its core business lines began to stagnate, and due to its massive size, it could raise insane amounts of debt.

As I pointed out previously, Oracle was a heavily indebted company even before the AI bubble. Coincidentally, due to its dalliance with OpenAI, Larry Ellison felt it necessary to crank the debt dial to eleven.

Oracle's spending has pushed its free cash flow into negative territory—negative $23.7 billion as of the end of fiscal year 2026—and as of the end of May, it had $129.5 billion in outstanding debt. This doesn't include its various lease commitments, totaling nearly $38 billion, nor an additional $260 billion in lease commitments signed but not yet actually commenced.

All this is to say that Oracle is borrowing massively for the benefit of one company, OpenAI, and if that company can't pay its bills, it's over. Oracle's existence—and Larry Ellison's personal wealth—depends on OpenAI honoring its commitment to spend $300 billion on compute power.

This is both the most obvious and least discussed part of the AI bubble—the hyperscalers' over $1 trillion in capital expenditure is fueling a massive semiconductor boom, based on the unlikely assumption that large language models will turn into something completely different.

If Microsoft, Google, Amazon, and Meta decide to stop spending $30 billion or more per quarter on GPUs, RAM, storage, and data center construction, it will tear a hole in the side of what is perceived as a permanent super-cycle.

I need to clarify how foolish it is to think the so-called semiconductor boom is anything but a temporary opportunity to fill your boots before a global stock market disaster. A disaster so severe it would make Futurum Group want to off itself.

The hyperscalers—whose capital expenditure will exceed their cash flow by Q3 2026—are seeing such poor returns on their AI investments that none of them will actually disclose revenue beyond vague "annualized revenue" figures, meaning all these investments are based on the idea that something completely different will happen in the future.

This future must generate at least $2 trillion in entirely new revenue for them by 2030, because if it doesn't, virtually all the capital expenditure will have been used to prop up Anthropic, OpenAI, and whatever Meta is doing with its chatbot.

There is no convincing or rational argument to support continued capital expenditure, at least not one that doesn't implicitly accept that most current spending is wasted, except for inflating stock prices and incubating two different massively loss-making AI labs. Those millions of H100, B200, and B300 GPUs won't usher in a digital god; they won't create recursive self-improvement; they won't be the fulcrum that adds $600 billion or more in entirely new revenue to current services. The only revenue they generate is compute spending from Anthropic and OpenAI, which I estimate accounts for 20% or more of Google, Amazon, and Microsoft's cloud revenue.

I must also clarify that these companies' costs far exceed equity investments. While Microsoft invested $13 billion in OpenAI, Microsoft executive Michael Wetter revealed in the Musk v. Altman case that the partnership has cost over $100 billion, implying OpenAI's infrastructure costs alone are at least $87 billion. I suspect Amazon and Google must spend similar amounts to handle Anthropic's equally voracious compute needs, especially considering Amazon's $11 billion-plus cost for the Anthropic-dedicated Rainier project data center.

This is a severely under-discussed part of the AI bubble. Anthropic and OpenAI have raised less than $300 billion combined since 2019, but I estimate their true costs are at least $500 billion, considering the hyperscaler capital expenditure investments necessary for their existence. This doesn't even account for the $340 billion or more Oracle will spend on building OpenAI's 7.1GW "Stargate" data center. These aren't startups; they are subsidiaries of big tech companies, existing as separate divisions just to boost equity positions and hide the truth: AI capital expenditure is a complete waste of money, even if you count the two fat, loss-making prodigals burning through tens of billions annually.

As I reported two weeks ago, OpenAI spent $17.2 billion on Microsoft Azure in 2025, a year in which it lost $20.9 billion on $13.04 billion in revenue. Even if that were profit (it isn't), it would still be $4.2 billion less than Microsoft's capital expenditure in Q1 2025.

Beyond OpenAI, Microsoft arguably has no AI business. While it boasted in April of having $37 billion in annualized AI revenue (meaning a non-specific month multiplied by 12), that's only about $3.08 billion per month, or less than a tenth of its $31.9 billion quarterly capital expenditure. Worse, Microsoft revealed this figure was "up 12% year-over-year," implying its annualized AI revenue in Q3 FY2025 was $16.59 billion, or about $1.38 billion per month.

However, my November report on OpenAI's inference spending showed it spent $2.947 billion in Q3 FY2025, annualizing to about $11.7 billion, meaning at least in that quarter, OpenAI likely accounted for ~70% of Microsoft's AI revenue. I'd be surprised if this changed dramatically over the year, given OpenAI's Q1 FY2026 inference spending was $3.648 billion.

All this is to say that the only real outcome of all this capital expenditure seems to be propping up two deeply unprofitable companies, Anthropic and OpenAI, and then recouping a small fraction as revenue, which is only possible thanks to hundreds of billions in venture capital subsidies.

Now, OpenAI and Anthropic account for 50% or more of the hyperscalers' remaining performance obligations, approximately $748 billion.

Beyond the mistaken belief that OpenAI or Anthropic can actually afford to pay without Google, Amazon, or Microsoft giving them money, there is simply no logical or rational reason to invest further capital expenditure in AI. The hyperscalers have no meaningful AI revenue in any form other than their own pseudo-startup investments. That A) they continue to invest and B) the market, analysts, and journalists act like everything is fine is both absurd and irrational.

Side note: I'm not discussing Meta because Meta has no AI story. Mark Zuckerberg has wasted every ounce of its capital expenditure, except for anything it could get by reselling capacity to others—but don't worry, he thinks (this is a quote!) Meta has a use for compute! No, sorry, those GPUs haven't driven meaningful advertising revenue growth; I've covered that before.

The record sales of NVIDIA, Micron, SanDisk, SK Hynix, and Samsung are a direct result of a completely speculative asset bubble, driven by the reckless and directionless capital expenditure of some of the world's largest and richest companies.

Anyone investing in data centers is building speculative capacity for demand that doesn't exist outside of Anthropic and OpenAI. If that demand existed, the AI data center new cloud company CoreWeave would have healthy and diversified revenue streams, instead of 65% of its revenue coming from Microsoft (for OpenAI) and NVIDIA, with the rest from Google (for OpenAI), Anthropic, Meta, and, of course, OpenAI itself. There is simply no other large-scale consumer of AI compute power, and the only reason we haven't hit that harsh reality yet is that data centers take 18-34 months to complete.

Even then, I can find almost no evidence of anyone besides OpenAI, Anthropic, and the hyperscalers having the demand or funding required to justify data center construction.

I really need to emphasize this.

If we assume NVIDIA CEO Jensen Huang's prediction of $1 trillion in Blackwell and Vera Rubin sales comes true, that would be approximately 40GW of data center capacity, with about 30GW of IT load. Assuming data centers generate roughly $12 per megawatt in revenue, generously speaking, that would create about $435 billion in annual compute demand by 2030.

Let's be very clear: the only companies that can currently afford to spend money on compute are either hyperscalers or companies subsidized by hyperscalers. Even so, besides OpenAI's projected $50 billion compute spend in 2026 and my estimate of a similar amount for Anthropic, there doesn't seem to be more than a few tens of billions in demand. If there were, CoreWeave, IREN, Nebius, Cipher Mining, and other new cloud providers would have hundreds of billions in remaining performance obligations, not ones that only expand under hyperscaler backing or Meta's Zuckerberg-style AI psychosis.

Let me put it more simply: those hundreds of billions in data centers are being built for no one. The only companies that can "afford" to pay even a fraction of the compute costs are unprofitable AI companies propped up by hyperscalers.

While this might read like a radical stance, I think looking at the current state of affairs and saying "screw it, I think hyperscalers should spend $1 trillion next year" is far more radical.

There is no rational reason to do so, besides fantasy thinking driven by a desperate market afraid to contemplate that tech has no ideas for high-speed growth.

Aside from creating OpenAI and Anthropic, the current capital expenditure is almost entirely wasteful. Microsoft 365 Copilot sucks. GitHub Copilot sucks. Google AI Overviews suck. Google Gemini is a follower LLM and therefore sucks. Meta's LLMs are very dangerous. Amazon Rufus sucks, and Amazon should be investigated by the SEC for implying it drove $10 billion in "annualized revenue" in Q3 2025, because it absolutely did not. Alexa+ sucks. Everything sucks, and it would be just as bad if big tech had spent only a quarter of the capital expenditure.

These products are almost universally despised, generate virtually no revenue, and even in the moderately successful case of GitHub Copilot (about $1.08 billion annualized by the end of last year), that's only because user compute is heavily subsidized, leading Microsoft to move users to token-based billing, angering customers used to paying $39 a month while burning thousands of dollars in tokens.

Yes, All This Money Could Be Wrong

Sundar Pichai, Andy Jassy, Satya Nadella, and Mark Zuckerberg are losers. They may be billionaires, they may run giant tech companies, but they are losers selling a doomed technology based on unreliable, inefficient, and overly expensive tech that isn't suited for the reliable, deterministic, "set it and forget it" nature people actually associate with AI.

The Big Four Losers are the only reason anyone takes these big loser models seriously, marking that the tech industry and our economy are also being driven by losers. Every bit of "progress" we see in LLMs comes from forcibly jamming square pegs into round holes—billions in training costs, hundreds of billions in capital expenditure, endless tools, scripts, wrappers, and layers trying to squeeze out anything resembling the promise of autonomy.

All the king's horses and all the king's men are pouring every dollar and ounce of brainpower into trying to make LLMs into something they are not, and we as a society are expected to coddle these things, act like they're great, and give them credit for things that haven't happened yet. I refuse to accept the premise that LLMs generating code or copying open-source software is evidence they will become powerful autonomous tools, and I think anyone extrapolating to that point is either intellectually bankrupt, deeply cynical, or gullible enough to click on every email claiming their PayPal account has been compromised.

I assure you, all this money could be wrong! Hyperscalers can indeed spend a trillion dollars on something that doesn't do what it says, because these companies are perfectly happy to mislead you, to quote Nik Suresh:

A large part of the economy is driven by people who are, simply put, very suggestible. That is to say, getting them excited and willing to spend money is very, very easy.

Why is everyone investing in data centers? Because the hyperscalers are doing it! Why are Micron and memory companies selling so much memory? Because A) GPUs use a ton of high-bandwidth memory, B) that HBM consumes three times the wafer space of regular DRAM, leaving less room for other, cheaper, lower-margin memory types, and C) because servers for these AI GPUs are also filled with memory!

These data centers aren't being built because creditors have any "insight" into the massive AI compute needs that generative AI tools need and will need. They see the "success" of ChatGPT and Claude (two heavily subsidized products), think that because Anthropic and OpenAI need a lot of compute, everyone will need a lot of compute. And because banks and private credit are desperate for ways to invest, everyone is so excited that it's super easy to get them excited about building big, sexy, expensive things!

It also doesn't help that a lot of information is deeply, deeply flawed.

Exponential View Should Be Ashamed of Itself