Tech giants chasing artificial intelligence have loaded up on debt at a pace that now worries even the bank that once cheered their ambitions. Goldman Sachs analysts and traders have issued pointed alerts this summer. The numbers tell a story of accelerating obligations. The six biggest AI spenders issued $244 billion in bonds so far in 2026. That total is 14 times the level seen in 2024.
Microsoft, Amazon, Alphabet, Meta, Oracle, Nvidia and SpaceX sit at the center of this borrowing wave. Their combined leverage ratios have climbed from 0.9 times to 1.8 times in just six months, according to a Yahoo Finance report. Credit markets show the strain. Spreads on Goldman’s AI-focused bond basket widened sharply. The amount of new supply needed to rattle investors has fallen from $75 billion to $25 billion. One Goldman derivatives trading head described conditions on the credit desk as “carnage.”
But the spending keeps rising. Wall Street analysts now project $527 billion in capital expenditures by AI hyperscalers for 2026. That figure sits well above the $465 billion consensus at the start of the third-quarter earnings season last year. Some models point even higher. Goldman Sachs Research laid out a baseline scenario that calls for $765 billion in annual AI-related capital spending in 2026. The number climbs to $1.6 trillion by 2031. Add it up and the cumulative tab reaches $7.6 trillion between 2026 and 2031. Those projections cover compute, data centers and power infrastructure.
Jim Covello, Goldman Sachs’ head of global equity research, sounded skeptical in a June 2026 podcast. “The economics of artificial intelligence are more questionable today than two years ago,” he said. He pointed out that enterprise buyers, model developers and hyperscalers have yet to demonstrate clear returns on their outlays. “At some point you got to make money,” Covello added. “We’ve gotten further away from that over the last couple years.” His colleague George Lee, co-head of the Goldman Sachs Global Institute, described the challenge in starker terms. “$7 to $8 trillion spent here… big hill to climb.”
And the market has noticed. Stocks of pure-play AI infrastructure companies have diverged from the broader hyperscaler group. Investors rotated away from names where operating earnings face pressure and capital spending relies more on borrowed money. Credit default swap spreads for AI-linked debt have blown out relative to the wider market. That divergence signals doubt that these investments will pay off quickly enough to service the growing debt load.
The physical demands compound the financial ones. Data centers that power modern AI models consume electricity at rates once reserved for entire cities. Global data-center electricity use hit roughly 415 terawatt-hours in 2024. That represented about 1.5 percent of worldwide consumption. The sector’s appetite has grown at a 12 percent annual clip since 2017, more than four times the pace of overall electricity demand. Forecasts suggest U.S. data centers alone could reach 400 to 600 terawatt-hours by 2030. In the most aggressive scenarios the figure climbs higher still.
Power availability has become the binding constraint. Nearly half of U.S. data centers planned for 2026 face delays or outright cancellation because local grids cannot support them. Grid modernization will cost trillions. Utilities have already boosted their own capital spending plans. The 47 largest investor-owned utilities expect outlays to jump 22 percent this year to $212 billion. Over the next five years that cumulative utility investment could top $1 trillion. Hyperscalers plan to spend even faster. Their combined AI-related capital expenditures could hit half a trillion dollars annually by the early 2030s.
Energy concerns stretch beyond electricity. Training and inference for frontier models require massive cooling. Water usage has surged in arid regions that host new facilities. Land demands for these campuses add another layer. One recent analysis from the United Nations University warned that AI’s infrastructure carries large carbon, water and land footprints that raise equity questions across global supply chains.
Yet the bets continue. Hyperscalers argue that falling inference costs will unlock new applications and drive productivity gains that justify today’s outlays. GPU rental prices and per-token inference costs have dropped dramatically. Frontier-equivalent inference costs fell from around $30 per million input tokens in 2023 to under 50 cents by mid-2026. That thousandfold improvement sounds impressive. Compute demand at leading labs still grows faster than efficiency gains. One recent analysis noted demand outrunning the cost curve by a factor of roughly 2.4 times per year.
Comparisons to earlier technology booms surface often. Goldman Sachs has flagged five signals from the dot-com era that investors should watch. A peak in investment spending ranks near the top of that list. So far AI capex shows no sign of cresting. Consensus estimates have been revised higher every quarter for two years running. But the bank’s own traders see limits in credit markets. Absorption capacity for new investment-grade supply has shrunk. Private credit and consumer finance already show cracks elsewhere in the system. A slowdown in AI returns could widen those fissures.
Some analysts remain optimistic about long-term payoffs. The International Energy Agency notes that affordable and reliable electricity will decide which countries capture the most value from AI. Nations that solve the power puzzle fastest stand to gain a strategic edge. In the United States, data centers could lift productivity enough to ease federal debt concerns over decades. That optimistic view assumes the current spending wave eventually translates into measurable economic output. Covello and his colleagues want evidence sooner. “Do the enterprises make or save money implementing AI?” he asked. “If they do, this technology is going to fulfill its promise.”
For now the market grants a long leash. Hyperscaler balance sheets remain strong enough to absorb higher leverage. Cash flows from cloud businesses still cover much of the spending. Yet the trajectory points toward greater reliance on external financing. If returns arrive later than expected, rating agencies could act. Downgrades would raise borrowing costs and force capital allocation choices that today’s optimistic forecasts ignore.
Recent commentary reinforces the tension. A Goldman Sachs video published at the end of June asked directly whether AI-driven corporate debt would strain credit markets. The discussion highlighted that the largest tech companies have already issued more than $170 billion in bonds this year, surpassing their full-year 2025 total. Expectations for further increases in AI outlays only add pressure. Supply could eventually overwhelm demand if monetization lags.
Power bottlenecks have already forced strategic shifts. Hyperscalers chase sites with existing grid capacity or renewable power purchase agreements. They have become the world’s largest corporate buyers of clean energy. In 2024 Big Tech accounted for 43 percent of all global clean-energy power purchase agreements. Prices for those contracts rose 35 percent that year. State governments compete aggressively with subsidies. Texas alone plans to offer more than $1 billion in incentives for data centers in 2025. The scramble reveals how location now dictates project viability as much as technology.
Private equity and infrastructure funds have stepped into the gap. Firms such as Apollo, KKR and Energy Capital Partners back developers that build and lease campuses to the hyperscalers. This layered ownership model spreads risk but complicates accountability for long-term energy and environmental impacts. Project-level debt finances many of these builds. That structure keeps the burden off hyperscaler balance sheets in the short term while still feeding overall corporate leverage statistics.
The debate will not resolve quickly. Technological progress races ahead. Model capabilities improve even as unit economics remain murky. Semiconductor makers continue to capture the bulk of economic value created so far. Covello expects that imbalance to correct eventually. “All of the economic value has continued to accrue to the semiconductor companies,” he observed. “At some point that has to rectify itself.” Until it does, debt markets will price in higher risk. Credit traders already feel the difference. Stock investors have started to differentiate between winners and those simply riding the wave.
So the AI build-out proceeds. Trillions more will be spent. Power plants will be built or repurposed. Grids will expand. Whether the returns arrive in time to service the obligations remains the open question that Goldman Sachs has now placed squarely in front of its clients. The bank that helped finance much of this expansion now warns that the bill is coming due faster than many assumed.
WebProNews is an iEntry Publication