Big tech capex accelerates: $1.1T spent on AI infra since 2023, $745B more expected in 2026 alone

Amazon, Google, Meta, and Microsoft have collectively spent more than $1.1 trillion on AI infrastructure—data centers, GPUs, memory, and power—since 2023. Financial Times analysis of earnings reports shows that the big four alone will add another $745 billion in capex during 2026, alongside Amazon's recently raised guidance to $220 billion. This acceleration reflects the industry's conviction that frontier AI training and inference demand will sustain massive capacity buildout for years.

The capex surge is reshaping downstream markets. Memory and storage chip makers—Micron, Samsung, SK Hynix—have been forced to prioritize AI hyperscalers due to their willingness to pay premium prices for high-bandwidth memory. This repricing has created consumer-grade DRAM shortages that rippled into PC, automotive, and smartphone markets. Apple, historically dominant with suppliers, was forced to raise consumer prices due to memory scarcity in 2025 and beyond.

Power and cooling have become structural constraints. The White House instituted a 'ratepayer protection pledge' in July 2026 asking hyperscalers and utilities to shield residential consumers from electricity cost increases, but so far no state has codified the pledge into law. Oregon's 2025 POWER Act showed the political risk: it imposed a 30% surcharge on users consuming over 20MW while cutting residential bills by 1.3%, demonstrating the tension between hyperscaler demand and customer protection.

The capital intensity of AI buildout has exposed hidden 'off-balance-sheet debt.' RBC Capital analysts estimate that big tech's contracted future capex obligations total around $1.65 trillion, annotated in quarterly statements as future liabilities. This amount equals 122% of the actual debt on their balance sheets, potentially understating the true leverage. Investors may be underestimating total financial obligation if these future capex commitments are treated as binding.

For practitioners, the lesson is that capex-driven capacity scarcity will persist into 2027–2028. Memory, power, and cooling will remain above historical pricing, and utility partnerships will become table-stakes for new data center sites. Hyperscalers are willing to operate at near-zero free cash flow to capture share, meaning competitive pressure will drive further capex escalation as rivals match spend to stay in the game.