Oracle’s Leases Show the Limits of AI Debt Hedges

$31.5 billion. Amazon, Microsoft and Alphabet spent that much on capital expenses and leases in 2016, up 22% from 2015. Cloud competition already required physical assets; nobody had mistaken a data center for a software download. By November 1, 2026, the reported cost of protecting against default by those three, Meta and Oracle had reached its highest level since 2021.

By 2019, the cost of computing at data-center scale was being framed chiefly as an advantage for the few companies able to afford it. In 2026, credit investors increasingly use company-specific default protection to hedge the possibility that long-dated AI infrastructure obligations outrun the cash flows supporting them. The companies have different balance sheets and different customers. They now share an exposure that creditors can trade.

Reported capex no longer describes the full commitment

The 2016 spending figure provides a baseline, not a claim that cloud finance began with AI. In the first half of 2024 alone, Amazon, Alphabet, Microsoft and Meta spent $106 billion on capex, 50% more than a year earlier. The groups and time periods differ, so $106 billion cannot be divided cleanly by $31.5 billion. Both measures describe money going into capacity.

A coverage analysis estimated that nine major technology companies had roughly $3 trillion in AI-related off-balance-sheet commitments, against about $600 billion in reported capex. The five-to-one ratio is arresting, but it is not a debt-to-capex ratio: multiyear commitments and reported capital spending have different accounting treatment and payment schedules. It identifies what a capex figure omits. Leases, chip purchases and capacity contracts can commit future cash before the associated capacity earns it.

Default protection gives distinct buildouts a common price

In February, Bloomberg reported rising trading in credit derivatives tied to individual technology companies as investors worried about borrowing for AI.

A credit-default swap protects against specified credit events at a named issuer. It does not quote the probability that an AI campus will run below capacity, and the reported five-company high does not establish that AI caused every move in each company’s spread. It does show investors seeking a hedge against obligations that may come due before the buildout’s returns are clear.

Credit buyers have not rejected the sector wholesale. Some AI-infrastructure bond sales were reportedly four times oversubscribed. Investors can want selected bonds and pay more to hedge selected issuers at the same time. Appetite for a bond issue is not a finding that every lease behind it is safe.

Oracle makes the distance between a contract and cash visible

OpenAI reportedly agreed to buy $300 billion of Oracle computing power over roughly five years, requiring 4.5 gigawatts of capacity beginning in 2027. Oracle, meanwhile, reported $248 billion in data-center and cloud-capacity commitments after signing roughly $150 billion of data-center leases in the three months ending November 30. OpenAI’s reported purchase and Oracle’s lease total describe different sides of the business; adding them would count promises, not establish net exposure.

Customer demand is real counter-evidence to an empty-rack scenario. AWS said its 2027 capacity was largely sold out and a good portion of 2028 was sold. A reservation helps a provider plan. Its credit value still depends on the customer’s payment obligations, the capacity arriving on time and the provider’s own lease terms. A sold year of capacity is not the same measure as collected cash.

Moving the construction loan does not move every obligation

An analysis put Oracle’s AI-data-center construction debt moved through special-purpose vehicles at $66 billion. Banks separately marketed more than $56 billion of investment-grade construction loans tied to Oracle’s future leases. The reported amounts may overlap; they are not two additions to Oracle’s corporate debt. They show why project finance requires a different map from reported capex: a bank can sell a construction loan while a tenant remains bound to future lease payments.

Nor does an Oracle credit-default swap automatically protect the buyer of a project loan. The swap references Oracle’s specified credit events. A power delay, an unfinished building or weak hardware resale value reaches that hedge only if it ultimately affects the named issuer under the contract. Loan recourse and lease terms determine where each loss first lands.

The $31.5 billion in 2016 measured what three companies spent building cloud capacity. It did not answer who would still owe money if a rack sat idle. Investors can now buy a price on one part of that answer—the issuer’s default—while the contracts decide the rest.