What happened
Microsoft, Amazon, Alphabet, Meta, Oracle, Nvidia, Broadcom, AMD, and SpaceX carry roughly $3 trillion in obligations—mostly future spending on data centers, chips, and power—that don't appear as debt on their balance sheets. The figure, based on filings mostly current through June 2026, nearly doubled from the $1.65 trillion estimate published in July. Alphabet alone disclosed $811 billion in purchase commitments and contractual obligations as of June 30, up 152% from $322 billion three months earlier.
Why it matters
This off-balance-sheet debt is roughly triple what these companies report as outstanding leases and long-term borrowings, and it is growing faster than traditional capital expenditures (which totaled about $600 billion over the trailing year). The commitments are mostly unstarted leases and purchase contracts for chips and servers—accounting rules don't require them to be booked as liabilities until the asset goes live. When depreciation on this spending finally hits the books, cumulative depreciation across Microsoft, Oracle, Meta, and Alphabet could exceed $520 billion over three years, which may strain margins unless revenue rises to match.
What to watch
Investor scrutiny should shift from reported debt to free cash flow relative to committed spending. Meta's Hyperion data-center project in Louisiana alone carries about $27 billion in debt that never touches Meta's balance sheet because it is structured as a joint venture majority-owned by Blue Owl Capital, with Meta as tenant and minority partner. Disclosure remains inconsistent: Nvidia disclosed $30 billion of cloud-service commitments and Oracle $10 billion, while Meta declined to quantify the cloud-capacity portion of its $238 billion in total commitments.
Summaries like this, in your inbox every morning.
The AI buildout has created an unusual accounting situation in which Big Tech's largest commitments remain invisible to standard balance-sheet analysis. For most of 2026, investors focused on free cash flow gaps between AI capital expenditures and reported cash generation. The conversation shifted to off-balance-sheet arrangements after the July estimate of $1.65 trillion, but the Wall Street Journal's fresh analysis—citing filings mostly through June 2026—revealed the figure is nearly twice as large. The breakdown shows $1.9 trillion in purchase commitments (contracts for chips, servers, and hardware) plus $1.2 trillion in unstarted leases, a distinction that matters because both fall outside reported liabilities until the assets or services commence.
Structural financing has amplified this opacity. Meta's Hyperion data-center joint venture in Louisiana exemplifies the pattern: by having Blue Owl Capital own the majority stake, Meta's $27 billion in project debt vanishes from Meta's consolidated balance sheet, leaving Meta in the role of tenant and minority stakeholder. Vendor financing and repurchase-style arrangements compound the visibility problem—the same underlying capacity is financed and re-financed across multiple balance sheets simultaneously, a practice the article terms "rehypothecated leverage." Disclosure inconsistency deepens the problem: Nvidia and Oracle quantified portions of their cloud-service commitments ($30 billion and $10 billion respectively), while Meta declined to break out the cloud-capacity component of its $238 billion total.
The real risk emerges when depreciation lands. Construction-in-progress balances at Oracle and Meta have risen roughly 200% and 90% year-over-year; when those assets enter service, depreciation could exceed $520 billion cumulatively across the four largest spenders over three years. That timing mismatch threatens margins—Oracle's depreciation could balloon from 7% of revenue to 28%, and Meta's from 9% to 19%, by fiscal 2028—unless revenue growth accelerates to absorb the charge. The risk compounds if demand shifts toward cheaper Chinese AI models delivering roughly 90% of frontier performance at lower cost, which would undercut both external revenue and the internal equity stakes (Microsoft's tens of billions in OpenAI, Amazon and Alphabet's billions in Anthropic) that helped justify the spending in the first place.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
