Hyperscalers will spend $450 billion on AI-specific infrastructure in 2026, approximately 75% of their total capital expenditure, according to CreditSights and Futurum Group estimates. A growing share of that spending is financed with debt, not earnings.
The Debt Behind the Data Centers
Goldman Sachs estimates that AI-linked debt issuance across the technology sector has already reached $412 billion in 2026. Big Tech will fund more than a third of its AI investments with debt by 2027. The capex-to-revenue ratios tell the structural story: Oracle is spending 86% of sales on capital expenditure. Meta sits at 54%. Microsoft at 47%. Alphabet at 46%.
Independent research pegs Meta's off-balance-sheet AI commitments at roughly $420 billion, nearly triple its reported debt. These are not startup gambles. These are the largest companies on Earth leveraging their balance sheets on a single thesis: that AI inference revenue will eventually justify the buildout.
Mohamed El-Erian warned last week that AI financing strains are "much broader" than any single company. When $450 billion in private sector infrastructure spending layers on top of persistent government deficits, the systemic credit pressure compounds. If AI monetization timelines slip by even 12 to 18 months, the interest burden becomes the story.
The Memory Wall
The physical constraints are tightening alongside the financial ones. Deloitte reported this week that costs for AI server DRAM roughly doubled during Q1 2026, with prices expected to quadruple for the full year. SK Hynix, the leading supplier of high-bandwidth memory chips, has seen its AI memory revenue more than triple since 2024.
New supply is not coming quickly. Deloitte's analysis, citing SemiAnalysis data, projects that memory supply tightness may persist until 2029 or 2030, even as chipmakers increase capital expenditure. The HBM market alone is forecast to reach $28 billion in revenue this year, with SK Hynix holding 53% share.
This is the constraint that most AI bulls underweight. You cannot build inference capacity faster than the memory supply chain can expand. Every delayed fab adds months to the gap between capex deployment and revenue generation.
The Telecom Parallel
Jim Chanos, the short seller who called Enron, has compared the AI capex boom to the 1999-2000 telecom buildout. The shape of the capex curves is nearly identical. Less than 5% of the fiber buried during the 1996-2001 telecom buildout was ever lit. Most of the builders went bankrupt. The technology was real. The demand eventually came. The companies that built the infrastructure did not survive to see it.
The critical difference: the early AI buildout was largely funded by earnings from existing businesses. That is changing. As the debt share rises, the AI trade begins to carry credit risk, not just execution risk. High-yield spreads in tech remain calm today. They were calm in early 2000, too.
For Bitcoin, the macro implication is indirect but structural. A credit cycle driven by AI overinvestment that forces central bank intervention would expand global liquidity. The same pattern that made Bitcoin a macro asset in 2020 could repeat, triggered by a sector most Bitcoin analysts are not watching.
Follow @UTXOMacro for daily Bitcoin, macro, and AI breakdowns.