A model that seemed to match the performance of top US systems at a fraction of the training cost was released in January 2025 by DeepSeek, a Chinese AI lab. About $600 billion in Nasdaq market value vanished within a day after the revelation reached Western markets. The biggest single-day decline in market capitalization by any firm in stock market history occurred when NVIDIA’s stock dropped by around 17% in a single session. The response was striking and illuminating: the US AI buildout’s whole investment thesis is predicated on the idea that increasing computing, chips, and data center capacity is essential to maintaining competitiveness. That notion might be incorrect, according to DeepSeek.
It’s really hard to remember how much money the big US tech companies are spending on AI infrastructure. In just the fiscal year 2025, Microsoft pledged $80 billion to building data centers. Alphabet plans to spend over $75 billion on capital projects in 2025. $60–65 billion was indicated by Meta. Amazon is worth more than $75 billion. When you add those up, you get a figure that is close to $300 billion in a single year—committed before the AI software industry has proven it can produce returns at even the scale needed to support such expenditures. In the middle of 2024, Goldman Sachs released a research note titled “Gen AI: Too Much Spend, Too Little Benefit?” which posed the question that the investment world was starting to voice.
In investor communications, Alphabet CEO Sundar Pichai, who has staked the company’s capital allocation strategy on AI infrastructure, said that there were “elements of irrationality” in the present cycle. That’s a cautious way of expressing what a more forthright person could say: the industry is creating more than it needs now, assuming that demand will come in to fill the capacity. That wager has proven accurate in the past; the fiber optic overbuild of the dot-com era, which led to the demise of hundreds of businesses, built the infrastructure that enabled broadband internet and ultimately powered Google, Amazon, and everything that came after. There was nothing wrong with the infrastructure. It was the timeframe. Businesses that took out loans to construct it ran out of time before the demand materialized.
Venture capitalist Bill Gurley and others have been publicly drawing comparisons to 1999–2000, but it’s important to be clear about where they hold and where they don’t. Businesses with virtually no revenue were trading at exorbitant valuations during the internet boom. The money being invested in AI infrastructure comes from the massive, lucrative core businesses of the current AI giants, including Microsoft, Alphabet, Meta, and Amazon. Because AI takes longer to monetize than the present capital expenditure cycle predicts, they won’t go bankrupt. The bears are genuinely worried that the AI investment thesis will result in years of below-expected returns for these stocks, that if chip demand slows, NVIDIA’s premium valuation will compress sharply, and that if the public market narrative changes, the funding environment for the private AI companies, currently valued at billions without substantial revenue, will look drastically different.
Daron Acemoglu, an economist at Goldman Sachs, provided a precise and unconventional estimate that sparked a lot of debate: he contended that AI would actually automate only roughly 4.6% of work tasks in the near future, which is significantly less than the estimates being used to support current infrastructure spending. Bulls respond that his calculation is cautious and that even 4.6 percent of global economic activity is a huge amount.
However, the discussion itself highlights the fundamental uncertainty: no one truly knows how rapidly enterprise AI adoption will result in income at the scale required to support an annual infrastructure expenditure of $300 billion. Although the adoption rates are indeed encouraging—McKinsey found that 65 percent of large organizations were employing generative AI by 2024—the majority of these installations are still in the pilot level, producing efficiency advantages that are hard to monetize and, at best, early-stage monetization.
