The arithmetic inside the AI bubble

Everyone in the stock market is wondering “Is AI a bubble?”

Quick answer: Yes.

To simplify, there are three sub-sectors in US AI: microchip manufacturers (Nvidia et.al.), tech giants (Google, Amazon, et.al.) which commission data centers, and software developers (OpenAI, Anthropic, etc., which need data centers to run their general-intelligence software.

Some companies do all three, and many of them now invest, lend or lease to each other.

A simple example: A sells B a rock on credit for $1,000. B sells it to C for $1,200. C only has $200 so A invests $1,000 in C, who pays B, who pays A. B records $200 profit and the public doesn’t know what a rock is worth, so all three companies’ share prices rise. But no new money was collected.

Everyone in American AI has been very clever about accounting, so the extent of circularity is difficult to trace. Nvidia, for example, derives half of all its revenues from just four customers, but won’t name them.

However, macroeconomics to the rescue. Simply conceptualize all these companies as a unified enterprise.

Transactions within the enterprise ultimately net out to zero in cash terms. What’s left is a single entity, the “American  AI Syndicate.” It makes chips, borrows money, builds data centers and hopes to get computing usage fees from real end-users.

A “data center” is a warehouse with computers. Data centers have been around a while, but the recent rush to build them is related to AGI, Artificial General Intelligence, which requires enormous computing power. A typical data center costs $500 million, mostly built with borrowed money. Roughly half of that goes to the chip-makers, which promptly re-invest most of the money in software developers, or still more data centers.

So how much does the Syndicate owe to outside creditors, including future lease backstop commitments that were used to attract private lending (a financial sector which is itself another bubble about to pop) that was hidden from their balance sheets?

I asked Google’s AI (ironically) and its answer was $3 trillion.

How much net profit is needed to repay principal and interest on $3 trillion of debt within the (estimated) five-year usable lifetime of the chips in the data centers?

$750 billion.

Chip-makers’ profits mostly came from other parties in the Syndicate so the only “outside-derived” profits are from the tech giants’ traditional activities (Google ads, Windows sales, etc.) — about $250 billion. Meanwhile, the AI software developers lose about $30 billion and show no signs of turning the corner.

So the Syndicate earns only $220 billion, but needs $750 billion annually for debt service.

It needs to more than triple its profits almost overnight  — not illusory profits from selling to each other as in our rock example, but real profits from real end-users.

This is unlikely:

• The public has turned against data centers, and many may not even be allowed to operate or may not be allocated sufficient power and water by unwilling host communities.

• Massive data centers are premised mostly on an artificial monopoly by Nvidia, which makes a specific kind of power-guzzling chip, and licenses the only operating software for that chip, which American AI got “double-hooked” on early. Chips obsolesce fast. Newer chips might not need massive data centers. Implication for the $3 trillion debt? Default.

• Tech giants think data centers are necessary and unavoidable infrastructure (analogous to the tracks for railroads) mainly because the US is fixated on AGI. Newer developers, especially from China, increasingly prefer “edge A.I.” (not “general” but use-specific intelligence, an example being robot welders). Since edge AI only needs to be intelligent about a few things, it can be completely embodied in a few chips in a physical device — no need for data centers, and no stream of data-center toll-road-type earnings with which to repay $3 trillion.

• By the way, a sharp swerve toward edge-type AI would be the simple first step to address the fear of AI destroying humanity (which numerous incidents have suggested it is willing to do). Physically separate the functions and don’t provide any means for the parts to talk to each other. Edge AI is the practical and safe-for-humanity future, not general-intelligence as the US is almost exclusively pursuing. More governments will soon figure this out. China already has.

• Each new wave of American AGI performs better, and charges less. Chinese models do almost as much and charge much less. Far from extracting oligopolistic profits by “cornering the market,” the American AI Syndicate seems to be losing market share and margin at the same time.

• Economists are wrong. Over a sufficiently-long timeframe, economic output=consumption. Period. Not “consumption+investment.” Why? No matter how long the thread, ultimately the only reason for investment is to produce something a consumer will buy (now or someday). Guess what? For consumption, you need consumers. Consumers with money. No future consumers? No present investment. There are many plausible AGI outcomes for mankind, but in most of them only multimillionaires have spending power. Everyone else is jobless or paid just enough to remain in servitude forever, with zero discretionary income. In the 19th century, the technology of railroads raised industrial-nation productivity by 45 percent. By whatever percentage AI increases productivity in the future, the more successful it is, the smaller the number of people who can afford to buy the fruits of that productivity. What’s the outcome of more productivity, but fewer buyers? More supply, but less demand? AGI profit margins, already negative, will get worse yet.

• OpenAI just spent $7 billion to buy stock from its own officers, supposedly with “cash reserves.” How can a company losing billions annually have “cash reserves?” It can’t. The cash is what was left over after March 2026, when investors pumped in $120 billion to cover mounting losses and development costs. This buyout can only be construed as an attempt to bribe key employees to keep silent about how circular American AI profits really are.

Is AI a bubble? Judge for yourself.

González is the owner of Plantation Bay Resort and Spa in Cebu. For full credentials, please see www.plantationbay.com/cred.