Hyperscaler capital spending reached $725 billion in 2026, up 77% year over year (Supply Side AI, Aug 2026)
The four biggest spenders in corporate history can no longer pay for the boom out of their own pockets, so they have quietly started borrowing it.
Two things are true at once and cannot stay true together. Microsoft, Alphabet, Amazon and Meta plan to spend roughly $725 billion building data centers in 2026, up about 77 percent from roughly $410 billion in 2025 (ValueAddVC AI Spending Tracker, Aug 11). And Morgan Stanley forecasts that the free cash flow of the five main data center operators will fall to negative $2.8 billion this year, from positive $187 billion in 2025 (Morgan Stanley via Dataconomy, Aug 19). The most cash-rich companies on Earth are spending themselves to a combined cash burn, in a single year, on buildings filled with chips that lose resale value fast.
The trigger for this week's anxiety was credit, not chips. Fund managers polled by Wall Street banks named hyperscaler artificial intelligence spending as the leading candidate to set off the next systemic credit event, with 2026 capital budgets now projected between $725 billion and $760 billion (Dataconomy, Aug 19). Robeco told Bloomberg that these companies are "almost price insensitive" to their own borrowing costs, which means they will keep selling long-dated bonds no matter what yields do, adding supply on top of what Washington, Tokyo and Brussels already need to place (Bloomberg, Aug 19). When the least rate-sensitive borrowers become one of the largest issuers, the bond market stops clearing on fundamentals and starts clearing on volume.
The slow pressure underneath is older than this year's guidance. The buildout outran its fuel: grid constraints have overtaken chip shortages as the top bottleneck, with reports indicating that 30 to 50 percent of new data center projects face delays waiting for electrical capacity (DailyAlpha citing industry reports, Mar 2026), and some analyses project roughly 40 percent of data centers facing power shortfalls by 2027 (informed, clearly, 2026). So the money is committed before the megawatts exist. Capital spent on a campus that sits unpowered earns nothing while its depreciation clock runs anyway.
The actors want different things and only one of them is paying retail rates for money. Alphabet raised its 2026 capex ceiling to $205 billion at second-quarter earnings, and Meta has lifted guidance twice this year (ValueAddVC AI Spending Tracker, Aug 11). Amazon carries the single largest budget at roughly $105 billion to $120 billion of 2026 guidance, ahead of Microsoft at $80 to $90 billion, Alphabet at $75 to $85 billion and Meta at $65 to $72 billion (Buildermuse hyperscaler capex analysis, Jul 2026). Each fears being the one whose model falls behind, because losing the model race costs the cloud franchise itself. Nvidia wants the orders to keep coming and projects total industry spending approaching $1 trillion by 2027 (Nvidia projection cited by remio.ai, 2026). The utilities and grid operators, who never asked for this demand curve, absorb it on decade-long timelines they do not control.
The financing has quietly changed character, and that is where the contradiction lives. Big Tech's AI infrastructure debt hit $159 billion in 2026, more than in all of last year (wealtharian analysis, Aug 2026). The five big spenders will devote about 90 percent of their operating cash flow to AI data centers this year, up from a historical average around 40 percent (Tomasz Tunguz citing industry filings, 2026). Some of the borrowing does not even appear as debt: Meta's Hyperion campus in Louisiana carries about $27 billion of borrowings held by a joint venture majority-owned by funds managed by Blue Owl Capital, with Meta as tenant and minority partner, keeping the liability off Meta's balance sheet (24/7 Wall St., Aug 22). Oracle's large new data-center leases run fifteen to nineteen years, obligations that filing reconstructions suggest sit outside the headline debt figures (WhatJobs News reconstruction, Aug 2026).
Here is the bounded comparison. In the late 1990s, telecom carriers like Global Crossing and WorldCom borrowed tens of billions to lay long-haul fiber, guided by the same logic that demand would catch up to any capacity built. They went bankrupt, but the dark fiber they left behind was later bought for cents on the dollar and became the physical foundation of the modern internet. The buildout was real even when the balance sheets were not. What is different this time: the fiber of 1999 did not depreciate, while GPUs written into today's data centers carry useful lives companies book at roughly six years against chip generations arriving every twelve to eighteen months. The four hyperscalers bought about $434 billion of property and equipment in the four quarters through March 2026 while reporting only around $149 billion of depreciation over the same span, a gap that must eventually close (Silicon Analysts estimate from reported cash-flow statements, Jul 10).
The richest companies in history have engineered themselves into a combined cash burn, and the bill is being sold to bondholders as infrastructure.
The counter-example argues the optimists' case, and it deserves a fair hearing. American railroads in the nineteenth century were overbuilt twice and bankrupted investors repeatedly, yet the cheap freight capacity they left behind powered a century of industrial growth. If AI follows that path, today's record budgets buy national infrastructure that society keeps even if early owners lose. The difference an investor should weigh: railroads had captive local monopolies once built, while rented GPU capacity competes against next year's cheaper, faster chips, and the revenue line depends on customers who are themselves burning venture and borrowed money to buy tokens.
Walk the consequences forward. First order: the suppliers collect now — Nvidia, TSMC, the construction and turbine trades, the independent power producers signing long contracts. Second order: bondholders inherit the risk as equity holders' cash cushions empty, and the credit spreads on data-center-backed paper start pricing a business whose cash flows depend on token prices nobody has defended in a downturn. Third order: if AI revenue disappoints, the losses do not stay inside the tech sector, because Robeco's point holds — this bond supply competes with governments for the same buyers, so a wave of downgrades lands on top of fiscal deficits and lifts every borrower's cost (Bloomberg, Aug 19). Apple, notably, hits all-time highs partly by standing aside, generating cash while rivals compress theirs (TechTimes, Jul 14).
The observable sequence if this read is right: watch the debt lines, not the capex lines. Confirmations include more Blue Owl-style joint ventures moving data-center debt off balance sheets, wider spreads on the new AI infrastructure bond issues relative to the parents' existing senior paper, and further free-cash-flow declines at Amazon and Alphabet in upcoming quarterly reports. The falsifier is simpler: if AI revenue growth catches up to the spending curve — if cloud and token revenues grow toward the depreciation rather than away from it — then negative free cash flow is just investment timing, and the credit alarm was noise.
Who absorbs the consequence if it breaks? Not the executives who approved the campuses; their compensation vests long before depreciation does. It lands first on the pension and insurance portfolios buying the long-dated bonds at thin spreads, then on the towns like those in Louisiana hosting campuses financed by entities that can restructure without their anchor tenant's name on the loan. The chips get replaced, the shells get refinanced, and the people holding the paper at par find out who actually owned the risk.
The judgment this earns: the story stopped being whether artificial intelligence works and became whether this scale of annual spending can be deployed profitably by anyone, which history answers far less kindly.