Silicon Valley’s $1 Trillion AI Bet Is Starting to Look Fragile

The AI boom is evident in almost theatrical ways when you drive through Santa Clara on a Tuesday morning. Overlooking new data center plots are construction cranes. trucks transporting server racks on roads not intended for such traffic. On office campuses, engineers fervently assert that they are creating the most significant technology in human history. The energy is genuine. The aspiration is genuine. A question that has been getting louder and more unsettling throughout the year is hidden somewhere among the quarterly earnings calls, analyst forecasts, and presidential tweets mentioning GDP figures: is any of this truly operating as people believe?
There is no question about the size of the investment. In 2024, private AI investment in the United States reached $109.1 billion, nearly twelve times that of China and twenty-four times that of the United Kingdom. Investment in corporate AI has increased thirteenfold since 2014 and increased by 44.5 percent in just one year. In 2025 alone, big tech companies like Microsoft, Amazon, Meta, Google, and Nvidia spent hundreds of billions on data centers and AI infrastructure. Global AI capital expenditure is estimated to be around $700 billion by 2026. These sums have no modern equivalent, and the perpetrators are not blatantly foolish.
Over the course of 2025 and 2026, it has become debatable what precisely all of that spending is doing for the US economy and whether the headline narrative—that investments in AI are the heroic force keeping GDP together—survives contact with how economists actually measure these things. The figure that sparked a thousand breathless articles was provided by Harvard’s Jason Furman: in the first half of 2025, investments in information-processing hardware and software accounted for 92% of GDP growth, despite making up only 4% of the US GDP. He calculated that annualized growth would have been 0.1 percent if that category had been eliminated. Not even a pulse. Wall Street and the White House both cited the number, which was interpreted as proof that Silicon Valley’s wager was, in fact, keeping the nation afloat.

The story became much more complicated when Jan Hatzius, chief economist at Goldman Sachs, began to explain how GDP is actually calculated. Domestic production is measured by GDP. The final figure is reduced by imports. Additionally, the dirty secret of the AI infrastructure boom is that about 75% of a data center’s cost comes from imported components. This fact was hidden for months before anyone well-known made it clear. Taiwanese semiconductors. South Korean memory chips. specialized hardware that is shipped to campuses in Texas, Oregon, and Virginia after being assembled elsewhere. “A lot of the AI investment that we’re seeing in the U.S. adds to Taiwanese GDP, and it adds to Korean GDP,” Hatzius stated in a telephone interview, “but not really that much to U.S. GDP.” The net effect on GDP is almost zero when you record a significant capital expenditure in the investment column and an equally significant import in the net exports column. Austan Goolsbee, president of the Chicago Federal Reserve, put it more bluntly: investment in AI “is buying imported goods,” and it has “not been as big a driver of the economy as some have portrayed.”
In light of these accounting realities, Goldman Sachs analyst Joseph Briggs noted that the story had been “very intuitive” and that its intuitive appeal had “prevented or limited the need to actually dig deeper.” That’s a very depressing observation. MRB Partners researchers attempted to combine the two perspectives and came to a middle-of-the-road conclusion: AI investment contributed between 20 and 25 percent of total growth over the first three quarters of 2025, or about 40 to 50 basis points to real GDP growth after accounting for imports. Significant, but not the pillar of support that the prevailing narrative demanded.
It is difficult to ignore how much more disturbing the productivity story is. Despite spending billions on AI, more than 80% of businesses report no discernible increases in productivity. As of right now, Hatzius admitted, there is “no reliable way to accurately measure how AI use among businesses and consumers contributes to economic growth.” The speed at which the technology is being implemented is astounding. As of yet, there has been no systematic appearance of the aggregate payoff in the data. Economic historians refer to this as the productivity J-curve, a well-established pattern in which transformative technologies exhibit costs right away but benefits only after businesses figure out how to reorganize around them. In the 1990s, it occurred with computers. Businesses purchased computers for years before the productivity gains appeared in GDP figures, as Brynjolfsson and associates meticulously documented. The quantifiable productivity impact of AI is expected to start around 2027, not 2025, according to Goldman Sachs’ own forecast. It’s possible that the infrastructure currently being constructed will serve as the basis for something truly important. It simply hasn’t occurred yet.
The current growth story’s concentration is already apparent and concerning. In 2025, J.P. Morgan Asset Management observed that AI capital expenditures had emerged as the primary driver of investment growth, “offsetting declines elsewhere.” Other significant industries, such as manufacturing, retail, real estate, and services, either made minimal contributions or actively reduced output during the first half of the year. America has become “one big bet on AI,” and “AI better deliver for the U.S., or its economy and markets will lose the one leg they are now standing on,” stated Ruchir Sharma, chair of Rockefeller International, in an unusually straightforward manner.” That is a truly concerning statement from a truly serious analyst, and it has not gotten nearly the attention it merits.
The bubble question is no longer considered rude. In March 2026, Benchmark general partner and venture capitalist Bill Gurley told CNBC that a hard reset was imminent. “We’re going to have an AI reset one day because interlopers cause waves to create bubbles. We have bubbles because when people become wealthy quickly, many others enter the market and want to become wealthy as well. Concerned, he continued, “I just think it’s harder to land the plane.” Overseeing Goldman Sachs during the 2008 subprime crisis, Lloyd Blankfein said he “smells” a similar building now. Over a billion dollars has been wagered on the decline of Nvidia and Palantir shares by Michael Burry, the investor who successfully shorted the housing market in 2008 and was made famous in The Big Short. Speaking in Italy, Jeff Bezos was blunt as usual: “This is a kind of industrial bubble.” There will be a drawdown, a reset, and eventually a check. He appears strangely content with that result. A lot of people wouldn’t be.
There is a slight but significant difference between the current state of affairs and the dot-com collapse. Nvidia is offering real products at incredibly high profit margins. Over the course of two years, its stock increased by about 1,700 percent, but the company’s revenue and chip demand are real. OpenAI recently raised $110 billion at a $780 billion valuation, with a revenue target of $12.7 billion for 2025. The companies at the core of this internet boom are not directly subject to the vapor-ware critique that aptly characterized the internet boom of the late 1990s. The self-reinforcing logic of competitive spending does apply, with businesses making large investments not because returns are guaranteed but rather because it seems riskier to fall behind if a rival gains an advantage than the investment itself. In order to gain priority access to its own chips, Nvidia reportedly invested up to $2 billion in xAI. The self-referential dynamics of this moment are demonstrated by that circular arrangement, in which a chip company finances a startup to guarantee that it purchases chips.
As this develops, there’s a sense that an honest assessment of Silicon Valley’s AI wager necessitates juggling two genuinely uncomfortable things at once. Goldman Sachs is not discounting AI; rather, it predicts that its effects will start to manifest around 2027. The infrastructure being built is real and may eventually support productivity gains that justify the investment. Despite the catastrophic dot-com bubble, there was a real computing boom in the 1990s. One does not negate the other. However, the short-term narrative that AI investment is valiantly maintaining U.S. growth and that the wager is currently paying off does not hold up to the accounting. Before it is even included in the GDP calculation, a large portion of the money leaves the nation. The increases in productivity have not materialized on a large scale. Furthermore, an economy that depends so heavily on a single technology cycle—the majority of which are produced overseas—is more vulnerable than the media has been willing to acknowledge.
Beneath a January 2026 report, the St. Louis Fed came to the conclusion that “A definitive empirical assessment of the long-run productivity effects of generative AI remains premature.” This may be the most accurate summary that is currently available. That does not mean that AI is unimportant. With trillions of dollars riding on the outcome, it accurately captures the state of the evidence, which is somewhere between enormous promise and unproven reality.