Artificial intelligence (AI) is reshaping how people work, search, write and build software. Wall Street is betting it will reshape the entire economy. In an October analysis, J.P. Morgan projected $900 billion in global AI capital spending in 2026 and as much as $5.5 trillion in cumulative investment by 2030. The question is whether the returns will come close to matching the bill.
Behind the excitement are some hard problems: expensive chips, data centers waiting for electricity, companies struggling to turn AI trials into profit, and an investment boom increasingly financed through debt and deals among the same handful of firms.
The comparison with the dot-com era is hard to avoid. Investors were right that the internet would change the world, but many were wrong about which companies would profit and how quickly. AI may be a breakthrough and a financial bubble at the same time.
The greatest infrastructure gamble in technology history
The companies that grew rich from software and advertising are now spending like heavy industry. They need vast data centers, processors, cooling systems and reliable power. J.P. Morgan estimates that Amazon, Alphabet, Meta, Oracle and Microsoft could spend about as much on AI infrastructure in 2026 as they had spent since ChatGPT launched in late 2022.
“The surge in AI capex-related expenditures has been the defining feature of our economic times,” said Michael Feroli, J.P. Morgan’s chief U.S. economist.
For the best of our weekly content！
You are now signed up for our newsletter
Check your email to complete sign up
The spending is concentrated among a handful of technology giants. An S&P Global study projected that Amazon, Alphabet (Google), Microsoft, Meta, and Oracle could account for 52 percent of all capital expenditures among S&P 500 companies by 2028, up from just 13 percent in 2020. While those figures include non-AI investments, they show how much of corporate America’s capital spending is becoming concentrated in five companies.
“Investors are underestimating how fundamentally AI is changing the Big Tech business model,” Shay Boloor, chief market strategist at Futurum Equities, told Reuters in July. The old business of selling highly scalable software is becoming much more dependent on costly physical assets.
A Reuters analysis of LSEG estimates projected a $534 billion increase in annual capital spending by five major technology firms between 2025 and 2027, against just $340 billion in additional annual operating cash flow. In other words, investment growth could outpace cash-flow growth by $194 billion.
S&P Global Ratings expects negative free operating cash flow across six hyperscalers in 2026 and 2027. Moody’s has also warned that some AI infrastructure borrowers could struggle to refinance debt maturing from 2027 to 2029.
“As AI infrastructure investment accelerates, the focus is expanding beyond the scale of spending to the funding models, financial commitments and long-term implications that accompany it,” said Naveen Sarma, managing director at S&P Global Ratings. The money is being spent now. Much of the hoped-for revenue remains years away.
Dark GPUs: Expensive chips waiting for electricity
Some of that hardware cannot even be turned on. An October Business Insider investigation reported that Amazon Web Services expected about $5 billion in Nvidia GB300 racks, followed by roughly $3 billion in equipment each month, while some facilities lacked enough electricity to use incoming systems. Equipment was held in warehouses, and one AWS team reportedly lacked “a clear forecasted path” for deploying it.
Amazon called staging equipment near future deployments “best business practice” and disputed claims that it lacked an overall deployment plan. The point is not that every chip is stranded, but that delays can eat into the useful life of equipment bought at today’s prices.
The problem extends beyond one company. The International Energy Agency estimates that grid constraints could delay about 20 percent of data center capacity planned globally through 2030. In Europe, Amazon energy executive Pamela MacDougall told Reuters that a grid connection can take seven years while a data center takes about two.
“There’s a misalignment. We want to expand and grow within two years,” MacDougall said, describing the delays as “challenging our growth aspirations.”
New power plants and transmission lines are not guaranteed to arrive on schedule, either. Some proposed projects will never be built. That does not mean electricity will never reach today’s unfinished facilities; it means investors may be paying for processors that lose commercial value while waiting for it.
The hidden accounting problem
AI’s spending boom also rests on a quieter assumption: how many years the hardware will remain economically useful. Companies spread the cost of servers and processors across estimated lifespans. A longer schedule means lower annual depreciation expenses and, all else equal, higher reported profit.
These estimates already move billions. Alphabet extended its servers’ estimated useful lives from four to six years, adding about $3 billion to 2023 net income. Microsoft made a similar change, increasing net income by about $3 billion that fiscal year. Meta‘s 2025 extension of server and networking asset lives raised net income by $2.59 billion. These accounting changes did not generate new cash.
Amazon then moved in the opposite direction for some servers, shortening their useful lives to five years in 2025. “The shorter useful lives are due to the increased pace of technology development, particularly in the area of artificial intelligence and machine learning,” the company said. The change reduced net income by about $1 billion.
The IMF offered an even sharper warning in April: “GPUs and advanced chips are a major share of artificial intelligence server and infrastructure costs and could face obsolescence within two years.” Its stress test found that shortening the average useful life of hyperscaler capital equipment from seven years to three could cut operating margins by more than nine percentage points. That was a scenario, not a forecast.
In an Oct. 29 Breakingviews analysis, Reuters estimated that cutting depreciation schedules by just one year at Alphabet, Amazon, Meta and Microsoft could lower their combined 2027 net income by about $36 billion. However, older processors may still be useful, especially for less demanding work. The risk is that they may earn less than expected long before they stop working.
The productivity revolution that hasn’t fully arrived
For all the impressive AI demonstrations, the productivity payoff remains uneven. An MIT study found that only 11 percent of S&P 500 companies had deeply integrated AI by the end of 2025, while many others were still running pilots. Heavy AI users did not consistently show higher output per employee.
“Now the question is, What’s the success rate of pilots?” asked MIT researcher Martin Fleming. Installing an AI tool is not the same as changing the workflow around it.
An October revision of a National Bureau of Economic Research study surveyed nearly 9,000 executives in seven countries. Although 78 percent of businesses reported using AI, respondents credited it with only a 0.2 percentage-point increase in annual labor productivity growth over the previous three years. They expected a larger gain in the next three.
McKinsey’s 2026 survey found a similar divide: 80 percent of respondents said AI improved their personal productivity, but only 37 percent reported a positive effect on company earnings. “Organizations’ conviction in AI is growing faster than the immediate financial returns they can attribute to it,” its researchers concluded.
Federal Reserve researchers described the gap this way: “We document a productivity paradox, in which perceived productivity gains are larger than measured productivity gains.” But the news is not all negative. A European Investment Bank study found roughly four percent higher labor productivity among European firms adopting AI. There are real benefits. The question is how widely and cheaply they can be delivered.
The increasingly circular AI economy
Money is not only pouring into AI; some of it is moving in circles. A chipmaker invests in a cloud business that buys its chips. An AI developer raises funding and spends it on computing services from an investor. Each deal may make business sense, but the ties can obscure how much demand comes from paying customers outside the industry’s inner circle.
In its October Global Financial Stability Report, the IMF reviewed 14 disclosed circular-financing transactions involving about $300 billion in equity, $650 billion in debt and $1.4 trillion in future revenue commitments. Those are different kinds of exposure, not money already spent. The concern is that one company’s trouble could hit its investors, lenders and suppliers at once.
Nvidia illustrates the overlap. Under a 2025 agreement, it committed to purchase up to $6.3 billion of CoreWeave’s unsold computing capacity, subject to contract terms. Nvidia also supplies CoreWeave with processors and invested another $2 billion in the firm in January 2026. The company says that investment supports expansion, not direct chip purchases.
In August, Nvidia announced guarantees worth up to $105 billion for an Ohio data center project intended to serve OpenAI. The guarantees are contingent obligations, not $105 billion in cash spending.
Nvidia CEO Jensen Huang said: “We are securing long-lived infrastructure for NVIDIA compute so OpenAI can deploy the most productive AI factories.”
S&P Global’s Mark Habib warned that some AI-related financial obligations may be hard to see in corporate disclosures. “It is safe to assume that there is more (debt) out there that we are not seeing,” he said. Moody’s separately flagged refinancing risks for weaker AI infrastructure borrowers.
During an October briefing, IMF officials warned of a potential “triple whammy”: falling share prices, weaker debt and disappearing future revenue commitments. These deals are not proof of accounting manipulation. But a closely connected industry can transmit a slowdown from one company to many others.
Is the bubble already beginning to separate from the technology?
The October earnings season shows why the answer is not simple. Alphabet reported strong cloud growth but roughly $9.1 billion in negative quarterly free cash flow. Amazon Web Services grew nearly 40 percent, yet Amazon reported negative free cash flow of $42.4 billion over the preceding 12 months as its equipment spending surged. Microsoft’s Azure revenue rose 48 percent, and the company still generated $18.6 billion in quarterly free cash flow.
Amazon CEO Andy Jassy defended the spending: “Our AWS investment in functionality and infrastructure is working,” he said in the company’s earnings release. Microsoft CEO Satya Nadella pointed to customer value: “We are delivering greater capability, choice, control, and cost efficiency for customers across our entire stack.”
Investors have become less patient. An Oct. 30 Reuters report said the Nasdaq 100 had fallen more than 5.8 percent during the month amid concerns about AI spending. Jake Behan of Direxion summed up the mood: “At this point, the market isn’t just asking how fast you can grow. It’s asking what you have to spend to get there.”
AI is already producing real revenue, and some of the largest companies can afford to keep investing. That matters. But revenue growth is not the same as a good return on every data center, loan or GPU purchased at the top of the cycle.
Original article
Comments