AI Investments - the Largest in Human History - Going Blind

Five U.S. Companies Will Spend Nearly Three-Quarters of a Trillion Dollars This Year, and None of Them Can Say How Much They're Getting in Return.

Credit analytics firm CreditSights estimates that capital spending by Amazon, Microsoft, Alphabet, Meta and Oracle will total about $750 billion in 2026, up 67% from last year and the third straight year of growth of more than 60%. The same firm estimates that about three-quarters of that will go toward AI infrastructure. In 2024, the five companies' combined spending was $256 billion.

Here's a comparison.

The CGIAR is the largest publicly funded agricultural research network in the world: fifteen centers, nine thousand researchers, eighty-nine countries. The varieties that come out of its labs feed hundreds of millions of people today. Its annual research budget is just over $900 million. In its fifty years of existence, about $60 billion has passed through the CGIAR, with $10 of measured benefit for every dollar spent, and with returns that have remained high throughout.

Everything that this network has received in fifty years is now being spent on AI infrastructure in less than six weeks.

This does not prove that money is being spent in the wrong place. It proves that it is being spent in one place, with a haste that the proven returns on other investments do not justify.

• What is being invoked

The companies signing these budgets have a better argument than it seems on the surface.

The potential market for AI should not be compared solely to the market for computer software, because part of its value may come from the cost of the work it takes to do so. Software sells for about $1.5 trillion a year worldwide, according to Gartner. Wages paid worldwide exceed $40 trillion. A tool that replaces hours of human labor is measured in terms of the second dimension, not the first. This is the labor replacement hypothesis, the only one that can support budgets of this magnitude. The labor replacement hypothesis is not a mere promise. Microsoft has officially announced an annualized AI revenue rate of more than $37 billion for the third quarter of fiscal 2026, up 123% year-over-year. Nvidia reported $89 billion in quarterly data center equipment revenue for the quarter ended July 2026. The cost of running a model with comparable performance to GPT-3.5 on the MMLU test has fallen more than 280-fold in about eighteen months, according to the Stanford AI Index. And AlphaFold solved the problem of protein structure prediction, which had been open for over fifty years.

The labor substitution hypothesis, however, also has a measured counterargument. The Boston Consulting Group estimates that about 12% of jobs in the United States fall into the category where artificial intelligence can directly replace humans in basic tasks, without demand compensating for the substitution. The rest remains assisted labor, not taken over labor-and assisted labor is sold for the price of a subscription, not the price of a salary.

The labor substitution hypothesis also justifies the size of the sums spent, but not the speed with which they are spent. The rush has four causes that arise not from the calculation of the return on investment, but from the structure of the competition and the incentives of the participants.

The gain is divided unequally. If first place is worth several times as much as third place, each company separately calculates that the loss of falling behind is greater than the loss of spending too much. The calculation is correct for each company individually and dangerous for all of them together: capacity is being built for a top position that not all can occupy.

The risk of the director is not the same as the risk of the shareholder. A director who spends too little risks being replaced before it is seen that he was right. One who spends too much, in an industry that is all wrong, shares the responsibility with the others.

Part of the demand is created by itself. The processor manufacturer invests in the buyer, and the buyer uses the money received to buy processors from the very person who invested in them. The sale is included in the manufacturer's turnover, although the demand was financed, in part, by the capital he gave to the buyer.

Finally, debt now covers 32% of these expenses, compared to 9% in fiscal year 2024, according to FactSet. The size of the bet remains the same. Who pays if things go wrong changes.

• First crack: money started coming out

Until last year, spending was done from the own cash of some extremely profitable companies. That's not the case everywhere anymore.

Alphabet reported, for the second quarter of 2026, an operating cash flow of $39.1 billion and capital expenditures of $44.9 billion. The result is a negative free cash flow of about $5.8 billion. Reuters reported on July 22 the general pressure on the cash flows of the major operators.

The second flaw is in accounting. Investor Michael Burry argues that the useful economic cycle of some artificial intelligence processors is closer to two to three years than the five to six used in accounting depreciation, and estimates that the difference could underestimate depreciation by about $176 billion between 2026 and 2028. It is a hypothesis, not a finding: old equipment can produce value even after the emergence of new generations. If the hypothesis is verified, however, part of the profits reported today are deferred expenses.

• An impossible calculation

One can attempt a calculation, and one must be clear where one stops.

The consulting firm Bain estimated, in its annual report from September 2025, that the infrastructure under construction would have to support about two trillion dollars of annual revenue by 2030 to cover its cost. Sparkline Capital repeated this estimate in October 2025, adding it to the revenue from artificial intelligence at that time - about a hundred times smaller.

The estimate is not a measurement. It depends on three quantities chosen by the one doing the calculation: how many years a processor lasts, how fast revenues grow, what return is required on capital. Moved within reasonable limits, the same three quantities give results that range from full coverage to a deficit of more than half.

Another path can be tried. The Epoch AI research institute has found that the amortized cost of equipment and energy to train a top-tier program has increased by about 2.4 times each year since 2016. That's one of the few solid metrics in the field.

But it doesn't say much on its own. Cost is a fractional term that lacks the second, the only one that matters: how much performance has been gained from one generation to the next. Herein lies the fundamental difficulty. There is no single, stable, and generally accepted unit of measurement that can translate the progress of these programs into a comparable measure from one generation to the next. There are many tests, and the Stanford AI Index shows that many of them saturate within a few months; a 2026 study of sixty language model tests found nearly half in a state of advanced saturation. The test that separated two programs last year no longer separates them this year, and is being replaced by another.

Here lies the most important fact of this field: you know exactly how much you are paying, and you cannot know, in comparable terms, what you are buying.

• What has been seen before

The closest precedent is the fiber-optic boom of the 1990s. After the overinvestment in telecommunications, most of the installed capacity remained unused for a long time, and on many international routes, data transport prices fell by more than 90% in two to three years, according to specialized publications of the time. The five largest manufacturers of telecommunications equipment were worth about a trillion dollars together and lost about 90% of their stock market value in a year or so, and the industry had borrowed about two trillion in the five years before the crash.

But fiber remained in the ground and was later used by others. A pattern seen in both rail and fiber is that a significant portion of the economic gain has subsequently been transferred to the users of the infrastructure, not to its builders. Who pays today is not necessarily who collects tomorrow.

The difference with fiber is that, then, the capacity has been unused for years. Today, capacity is sold before it is built, and the constraint is physical: electricity, memory, and labor, not a lack of customers.

• What remains standing

A single observation, which does not depend on any estimate and remains true even if the bet turns out well.

Companies publish investment budgets, demand estimates, and revenue expectations, but they do not publish verifiable thresholds at which construction would stop. Not at what level of revenue. Not at what date. Not under what condition.

An investment whose stopping condition is not publicly known is less like an economic calculation and more like a bet without a criterion for refutation. Things would border on the irrational - if there were no other explanation.

• You buy time, not products

The statutes of the University of Paris, issued in 1215, required at least six years of liberal arts study before one could teach, and for theology they set a minimum age of thirty-five, at the end of a course that could last fifteen years. The harshness had a basis that we have since lost. At that time, the body of scholarly knowledge recognized by the European university could still be contained in an individual path, provided that man did nothing else.

The rupture becomes visible around 1500, and it was also caused by a technical change. It is estimated that, in the fifty years since Gutenberg, European printing presses had produced between eight and twenty million copies. In the same years, ships brought from America and Asia plants, animals, languages, and peoples that no ancient author had written about. Erasmus and Pico della Mirandola still belong to the Renaissance ideal of the scholar who can aspire to encompass all knowledge; after them, the ideal dies out.

Today's specializations do not, therefore, come from a narrowing of the human mind. They come from the explosion of the quantity of knowledge. For five centuries now, no one can know everything, and each generation has received this constraint as an irretrievable loss.

What is bought today with hundreds of billions of dollars is not a computer program. It is possible that that loss is not definitive: an intelligible access to the totality of human knowledge, with the possibility of linking together fields that have not spoken to each other for five hundred years. A technical change then broke the possibility of encompassing everything; now one pays for a technical change that one hopes will restore it.

And from here on, a hypothesis begins, which I give as a hypothesis. At the least confessed end of the same desire lies the extension of life and, at its limit, immortality. None of the above figures prove it. But it explains the rush better than any model of profitability: one does not buy products, one buys time. For those who think that they are paying to not die, the question of what amount to stop at makes no sense.

The people who built the Tower of Babel did not want a tall building. They wanted to reach the sky. The greatest mobilization of bricks and arms of their era did not stop because the materials ran out, but because those who were building stopped understanding each other.

Today, concrete is poured and cable is pulled for the same purpose, with the same haste, and a common language is missing from the beginning: there is no comparable measure of what is being built, so there are no words in which one could say that enough has been built.

There is only one question left to ask those who sign these budgets, and it is not when the investments become profitable. They have the answer ready for that.

What would have to happen for you to stop?

(NOTE: text prepared with AI assistance)