The Open Web Is Barricading Itself

How AI is pushing Silicon Valley into the infrastructure economy

In February 2023, John Hennessy, then chairman of Alphabet, Google’s parent company, told Reuters that a conversation with an artificial intelligence likely cost ten times more than a classic search. He thought the gap was temporary: “a two-year problem, at worst.”

Almost four years later, Wall Street has sent the bill. On Thursday, October 29, Alphabet shares fell nearly 13%, its worst session since its IPO in 2004, with some $475 billion wiped out. The cause: disappointing results and, above all, cash flow in the red. For two quarters — a first in its history as a public company — Alphabet has been investing more than its operations generate.

That same evening, Amazon confirmed the diagnosis: $77 billion invested in three months, $33 billion more than its operations generated. Its stock tumbled as well.

Hennessy was right on one point: the unit cost has collapsed, by 78% in 2025 for Gemini, according to Google. But he was reasoning about the price of an answer, not the number of answers. As early as 1865, economist William Stanley Jevons observed that the more efficient steam engines became at burning coal, the more coal England burned. The same goes for compute: the cheaper an answer becomes, the more we demand. Add to that the race for giant models, equally hungry for machines.

A broken alchemy

For a quarter of a century, Silicon Valley seemed to prove Milton Friedman wrong, when he liked to repeat that there is no such thing as a free lunch. Software obeys a singular economy: the first copy costs a fortune, the billionth costs almost nothing; each additional sale then turns almost entirely into profit. Google, Microsoft and Meta have drawn outsized margins from it.

For them, alas, AI breaks that alchemy: an answer cannot be copied, it has to be computed in data centers that consume electricity and wear out their chips. Cost no longer disappears with volume: it follows it. Alphabet, which invested $31 billion in 2022, will spend close to $200 billion this year, and even more in 2027.

While sales at Amazon, Microsoft, Alphabet and Meta are still growing by 12 to 26%, the four groups have invested more than $200 billion in three months — $26 billion more than their operations generated.

Google is paying to replace itself

Google’s case is the most troubling. Admittedly, the number of searches is growing at its fastest pace in five years, but search revenue is slowing: +12%, versus +19% at the start of the year. Its chief business officer acknowledged that the rise of “AI Mode” — conversational search — weighed on results, because it monetizes less well. Google is therefore spending record sums on a product that is cannibalizing its own cash cow. Management replies that most of these queries are in addition to classic searches, and it is now placing ads inside the answers.

The bet is rational: if Google doesn’t replace its engine, someone else will. But there is nothing free about it anymore. Above all, it breaks the old pact of the Web. For twenty years, a tacit symbiosis governed the circulation of knowledge and money: publishers let Google crawl their pages for free; in return, the engine sent them billions of visitors they could monetize. But a search engine directs, while an AI answers: it absorbs the substance of sites while confiscating the click.

The admission came right after Alphabet’s rout. In front of his shareholders, Reddit CEO Steve Huffman gave a clinical account of this rupture: “AI Mode does not bring traffic. Not to us, nor to anyone as far as I know.” Despite sharply higher revenue, Reddit saw its number of American users decline, before its stock collapsed by more than 25% during the session.

By computing the answer instead of directing the user, the tech giants are sawing off the branch on which their raw material rests. The counterattack is already underway: Amazon is blocking AI bots from its product pages to protect its advertising business; Reddit is pushing visitors behind a login wall and now charges top dollar for access to its archives. The doors are closing: after losing the free copying of software, Silicon Valley is losing the free fuel that powered it. The barricade is up.

The money that built railways and power plants

The economics of AI has a precedent: railways and electricity, where you pay for the line years before traffic makes it profitable. In July, Amazon boss Andy Jassy admitted it: cash flow will suffer as long as its data centers are not up and running, profitable and in use for a few years.

One difference makes the picture worse, however: an electric utility sells power with stable demand and often regulated prices. AI offers neither, and its servers age so fast they have to be replaced before their return is proven. Yet nearly two-thirds of Alphabet’s investments go to servers and networking.

The most telling symbol dates from June: Alphabet announced an unprecedented $84.75 billion equity raise, including $40 billion to be deployed over the coming months. Berkshire Hathaway, the conglomerate built by Warren Buffett, subscribed for $10 billion. Berkshire also owns one of America’s great railroad networks and power utilities: the money that once financed rails and power plants is now financing chips.

The decisive advantage therefore no longer lies solely in engineering talent, but in the ability to mobilize hundreds of billions. And the stock market is starting to draw the consequences. Admittedly, demand is real: Google’s cloud doubled in one year, Amazon’s is accelerating. But investors are no longer asking whether AI sells; they are asking how much each dollar invested returns. They were paying for these groups as software publishers; they are now judging them as infrastructure operators.

Yet the history of infrastructure is not kind to those who finance it. In the 1840s, Britain was covered with railways at the cost of a multitude of ruined shareholders. Around 2000, fiber optics brought down operators before benefiting others. For today’s giants, it remains to be seen who will capture the value created.

For now, despite the race by Alphabet and Amazon to design their own chips, their suppliers are taking a good share of it, while Web publishers are losing their audience. This summer, Amazon raised its annual investment budget by $20 billion, citing the rising price of memory chips.

Silicon Valley used to sell code that copies for nothing, fed by a Web that gave itself away for nothing. It now sells compute that has to be paid for with every answer, drawn from sites that are closing their doors. Milton Friedman ended up being right twice over.