Where Do Microsoft, Amazon, and Meta Get the Money for AI Infrastructure as Their Debt Keeps Growing?

Where Do Microsoft, Amazon, and Meta Get the Money for AI Infrastructure as Their Debt Keeps Growing?

Ondřej Barták
Ondřej Barták
Entrepreneur and Programmer
27. 5. 2026
5 minutes reading · 12 views
Where Do Microsoft, Amazon, and Meta Get the Money for AI Infrastructure as Their Debt Keeps Growing?

    At first glance, the balance sheets of the largest US technology companies look exemplary. JPMorgan calculated that 28 leading companies associated with artificial intelligence account for half of the S&P 500's market capitalization but carry only five percent of its total net debt. Microsoft, Amazon, Alphabet, Meta and Apple look like financially stable fortresses. That is how the market reads it and how analysts present it. But there is a catch.

    The Bank for International Settlements, an institution described as the "central bank of central banks," published a study in March that changes the way the situation is viewed. Large technology companies issued $120 billion in corporate bonds in 2025. That is five times as much as in 2024. At the same time, they moved another $120 billion off their balance sheets into so-called special purpose entities. These are legally separate entities that finance assets and assume debt without that debt appearing in standard financial statements.

    The result is $240 billion in new borrowing by five companies in a single year. Net debt remains low on paper, but total debt is growing at a pace unprecedented in the history of the technology sector.

    AI costs astronomical sums

    The five largest US technology companies plan to spend approximately $660 billion to $690 billion on infrastructure in 2026. That represents a year-over-year increase of one-third, with roughly three-quarters expected to go directly toward computing capacity for artificial intelligence. Amazon expects to spend more than $200 billion. Microsoft nearly $190 billion. Alphabet is in the same range, while Meta plans to spend between $125 billion and $145 billion. Altogether, that amounts to approximately 2.2 percent of total US GDP, concentrated in four companies.

    Goldman Sachs estimates global annual spending on AI infrastructure at $765 billion for 2026 and predicts that this figure will exceed $1.6 trillion by 2031. By the end of the decade, total investment in AI infrastructure is expected to reach $5.2 trillion. Even hundreds of billions of dollars in their own cash are not enough to keep up with the pace. Hence the bonds, hence the special purpose entities, and hence the volumes that have shaken credit markets in recent months—volumes previously seen only once a decade.

    Banks face a natural problem. Every institution has limits on how much exposure it can have to a single company. When Meta, Alphabet or Amazon issue bonds in such volumes, banks quickly approach those limits. And this is where credit derivatives come into play.

    Specifically, credit default swaps, or CDS for short. A bank buys protection against the potential default of a company's debt, thereby freeing up room in its portfolio and allowing it to lend to the company again, underwrite its bonds or trade other products with it. A derivative is simply a tool that allows banks to do business even when their limits would otherwise be exhausted.

    The volumes are staggering. Trading in CDS linked to Microsoft, Amazon and Oracle surged to a notional value of $4.6 billion in the first quarter of 2026, compared with $759 million in the same period last year. Bank of America recorded a tenfold increase in monthly hyperscaler CDS trading volumes since the beginning of 2025. CDS on Meta, launched only in October last year, reached a volume of more than half a billion dollars in a single quarter.

    Hedge funds see an opportunity

    Bank demand for protection is driving CDS prices higher. In Meta's case, five-year contracts are trading at roughly 0.73 percentage points annually. Anyone who sells protection on $10 million of principal receives $73,000 a year. Yet Meta has an AA- rating from S&P Global and an Aa3 rating from Moody's, the fourth-highest possible grade.

    For comparison, selling CDS on the average North American investment-grade index yields only $52,000 a year, even though the average rating of that index is four notches lower. In other words, selling protection on Meta offers a higher return for better credit quality. "It is the best opportunity in the AA segment of credit derivatives in a very long time," said Andrew Weinberg of Saba Capital Management. The market is, as he himself puts it, inefficient. And inefficient markets attract money.

    Debt is growing, but this is not a solvency crisis

    It would be easy to turn this into a catastrophic story. It is not that simple. At the beginning of 2026, Microsoft held more than $58 billion in cash and marketable securities. The group as a whole is sitting on hundreds of billions. Their revenues are growing, the financial system is willing to lend to them, and bankers, asset managers and private credit funds are allocating capital with full confidence that they will get their money back.

    Goldman Sachs also noted that the real constraint on spending will not be cash flows or balance-sheet capacity, but rather the supply chain or investors' willingness to finance further expansion. In other words, these companies have more borrowing capacity than most people realize.

    If more than $660 billion is spent on infrastructure in 2026 and part of that financing is structured through instruments that do not appear in conventional financial statements, then "low debt" as a headline no longer reflects the true exposure. The market is becoming increasingly aware of this. Hence the record volumes of derivatives, hence the banks buying protection, and hence the hedge funds that are happy to take the other side of the trade.

    Sources: ainvest.com, bloomberg.com and finance.yahoo.com

    Advertisement

    Content created with help from UpTier.

    SEO and GEO on autopilot. UpTier’s multi-agent systems write and optimize content for search engines and AI answers.

    Discover UpTier ↗

    Category:AI
    Did you enjoy this article?
    Discover more interesting posts on our blog
    Back to blog

    Related posts

    Anthropic explicitly bans election interference in Claude usage policyAnthropic explicitly bans election interference in Claude usage policy
    Anthropic has updated its model usage policy to prohibit voter deception, deceptive campaigns and persistent verbal abuse of models. The company says the abuse ban excludes ordinary frustration, disagreement and research.
    1 min read
    10. 10. 2026
    xAI releases Grok Imagine Video 1.5 Lite to create videos with audioxAI releases Grok Imagine Video 1.5 Lite to create videos with audio
    The lightweight version of xAI’s video model creates clips from text or images with native audio. It supports lengths of 1 to 15 seconds and is available through an API and developer platforms.
    2 min read
    10. 10. 2026
    Microsoft releases Decision-1 to score predefined choicesMicrosoft releases Decision-1 to score predefined choices
    Microsoft-Decision-1 returns a probability score for each supplied option. Available in Microsoft Foundry, the model is designed for ticket classification, request routing, and evaluating AI responses and agent actions.
    3 min read
    10. 10. 2026
    Přihlaste se k odběru našeho newsletteru
    Zůstaňte informováni o nejnovějších příspěvcích, exkluzivních nabídkách, a aktualizacích.
    CodedTrip

    Operated by CodedTrip LLC, USA.

    YouTube
    TikTok