Big Tech's Hidden AI Debt Hit $1.65 Trillion, Don't Bank on Cheap Compute

What Happened

A Nikkei study of recent filings from Alphabet, Microsoft, Amazon, Meta, and Oracle puts their AI-related off-balance-sheet obligations at $1.65 trillion. These are long-term commitments for data-center leases, GPUs, and servers that sit outside reported debt. The total is up eightfold in about four years and now exceeds the roughly $1.35 trillion of debt actually carried on their balance sheets. Meta alone holds around $420 billion off the books, nearly triple its transparent debt. The mechanism is the special-purpose vehicle: an SPV owns or builds the data-center asset, the hyperscaler takes a minority stake and signs a long-term lease or capacity offtake, and most of the borrowing stays off its balance sheet. The Bank for International Settlements devoted part of its June report to this, warning that the buildout has outgrown every prior technology boom and that the debt and circular-financing structures holding it up can unwind fast. The five hyperscalers have earmarked roughly $725 billion of capex for 2026, about three-quarters of it aimed at AI.

What This Means for Founders

Two things belong on a founder’s desk. First, stop pricing in infinite cheap compute. Today’s low inference prices carry a subsidy that Big Tech is funding with debt and off-balance-sheet vehicles. If that financing stalls, the buildout slows, or the circular deals unwind, compute prices can move back up. Run your unit economics twice, once at today’s token prices and once at a level 30 percent higher, and check whether the second version still clears. The lean playbook YC has preached for a decade holds: rent compute, don’t own it, and put the moat somewhere prices can’t erase it.

Second, don’t lock your entire cost base to one closed frontier lab. On OpenRouter, the neutral model router, US-origin models fell from about 70 percent of token volume in June 2025 to about 30 percent a year later, with Chinese open-weight models like DeepSeek, Kimi, and Qwen taking the rest. For bulk, repetitive work they run 60 to 90 percent cheaper than OpenAI and Anthropic. That pressure also complicates the labs’ IPO math. OpenAI, on roughly $20 billion of annualized revenue, is projecting about $14 billion in losses for 2026; CFO Sarah Friar has flagged whether it can sustain its data-center spending, and it is reportedly leaning toward pushing its listing into 2027. Anthropic was valued at $965 billion in its latest round and is reportedly moving toward a confidential filing, yet the two labs lose about $20 billion a year between them. If those listings slip, expect their pricing and free tiers to tighten too.

What You Can Do Now

Build for model portability. Route the high-volume jobs open weights handle well, classification, RAG, summarization, and bulk generation, to the cheapest supply, and reserve frontier models for the work that needs them. Keep at least two providers wired in so a price spike becomes a routing change, not a rewrite. Then watch the leading indicators of consolidation and repricing, which surface in API price sheets and free-tier terms before they show up in any 10-K. Check them every quarter.