Big Tech's $755B AI spending spree shifts from buybacks to bonds

Amazon, Alphabet, Meta, Microsoft and Oracle are on track to spend $755 billion on capital expenditures in 2026, an 83% jump from a year earlier, as the AI infrastructure buildout forces the biggest reallocation of corporate cash in a generation.

"The funding sources to me are normal funding sources, because I see them in infrastructure all the time," John Medina, an infrastructure finance analyst at Moody's Ratings, said. "It's just that this sector is now being seen as infrastructure."

The spending has already reshaped how Big Tech raises money. Meta and funds managed by Blue Owl Capital closed a $27.3 billion private placement for the Hyperion data-center campus in Louisiana in October, the largest deal of its kind on record, using a 144A bond structure that S&P Global Ratings assigned an A+ rating. Developers including CoreWeave, Applied Digital and Related Cos. have added more than $40 billion in such placements since November, according to a Bisnow analysis. Amazon's free cash flow fell to $1.2 billion in the 12 months through March from $25.9 billion a year earlier, as property and equipment purchases rose by $59.3 billion.

The shift changes what shareholders own. Hyperscaler buybacks fell by nearly two-thirds in the first quarter, Goldman Sachs data show, with Alphabet repurchasing no stock after spending $15.1 billion in the same period a year earlier. Free cash flow for the five companies is expected to drop 91% in 2026 to about $16 billion, the Wall Street Journal reported, even as net income rises 25% to $506 billion. That gap — between accounting earnings and actual cash — is the AI buildout in a single number.

Debt Markets Become the Backstop

The bond market has filled the void quickly. Compass Datacenters, owned by Brookfield, raised $830 million in February by securitizing six fully leased buildings in Phoenix and Toronto, a portfolio with 198.2 megawatts of capacity. Moody's graded the $500 million senior slice AAA, its first top mark for a hyperscale securitization, with the remaining tranches rated Aa3 and A2. DataBank Holdings followed with a $665 million securitization covering 36 data centers, more than half its portfolio, that was 84% leased by floor area to about 1,750 customers.

The mechanism works because the leases are hard to break. Relocating racks of computing equipment is expensive and disruptive, so tenants tend to stay put. In the Hyperion deal, Blue Owl-managed funds took an 80% stake and Meta kept 20%, with Meta leasing the whole campus once built and absorbing construction risk. The site is expected to draw up to 5 gigawatts of power, roughly the demand of four million homes. The contract structure lifts the project's credit close to that of Meta itself.

J.P. Morgan estimates AI capital spending could reach $5.5 trillion by 2030, with $4.1 trillion financed through debt. Nvidia sold its own bonds this week, showing the funding call is spreading beyond the cloud platforms. Eight companies borrowed roughly $4.1 billion of data-center asset-backed securities in the opening weeks of 2026, the fastest annual start since at least 2014, according to Bloomberg data cited by Datacenter Knowledge.

The Risk That Hasn't Been Tested

The Financial Stability Board warned on May 6 that private credit at its current size — $1.5 trillion to $2 trillion — has not been tested through a severe downturn, flagging the sector's concentration in technology and its deepening links to banks. Blue Owl's own shares are down more than 30% over the past year after a sell-off in March rattled investors worried about how much AI risk private-credit lenders have taken on.

The reclassification of data centers from industrial property to infrastructure is doing much of the work in these ratings. It assumes full occupancy and treats rent as close to guaranteed. Should AI demand fall short of the forecasts now baked into lease projections, that assumption would be the first thing to give. Medina draws a careful line on this point. "There is a difference between the risk of the debt and the risk of the market," he said, arguing the 144A tranches would be among the last places any distress showed up.

For now, pension funds and insurers are happy buyers, drawn by long-dated income that matches their liabilities. But the question the FSB raised lingers: a financing model built for a sector that barely existed five years ago, resting on leases whose value depends on a technology boom continuing, has yet to meet its first recession. Until it does, a high rating tells you how well the deal is built, and very little about whether the boom underneath it will hold.

For investors, the calculus has changed. Alphabet, Microsoft, Meta, Amazon and Oracle are starting to look more like infrastructure companies — heavy capital spending, long payback periods, rising financing needs and less excess cash to hand back. That does not make the bet wrong. Railroads, utilities and telecom networks built valuable businesses this way. But it does mean the shareholder contract has been rewritten. Anyone still pricing these stocks as though the 2017-to-2022 buyback machine is still humming is working from an outdated model.

This article is for informational purposes only and does not constitute investment advice.