Google, Meta, Amazon, and Microsoft are collectively projected to spend $750 billion on AI infrastructure in 2026, according to Bloomberg data cited by City AM. That figure is roughly half the annual spending of the entire UK government and represents the largest single-year capital expenditure surge in technology industry history, per Tech Insider’s analysis.
The individual commitments, as detailed by Tech Insider: Amazon at $200 billion, Google at $175 to $185 billion, Microsoft at approximately $150 billion annualized, and Meta at $115 to $135 billion. Share prices across the four companies have doubled since 2023, but quarterly capex budgets have roughly quadrupled over the same period, according to City AM.
The Depreciation Problem
The headline spending obscures a growing maintenance burden. Annual depreciation of property and equipment across the four companies has nearly doubled over the past two years to $116 billion, per City AM. Data center servers typically last three to six years before replacement, and the pace of AI hardware innovation pushes that toward the shorter end.
Amazon has already acted on this. The company’s Q1 2025 10-Q filing reduced the expected useful life of its data center assets from six years to five, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.” Meta, Microsoft, and Alphabet still use six-year depreciation schedules, but City AM notes it appears to be a matter of time before they follow.
Data center equipment accounts for roughly two-thirds of build costs, according to City AM. Layering replacement costs onto current capex projections pushes total infrastructure obligations substantially higher than the headline figures suggest.
Financing Under Strain
Alphabet illustrates the financing pressure. Google’s parent company has raised $85 billion in debt over the past year and plans an additional $80 billion equity raise, according to City AM. These are unprecedented levels for a company that historically funded operations from cash flow.
The spending is concentrated on inference infrastructure rather than training, per Tech Insider. As agent workloads scale, a single task can trigger multiple model calls, query vector databases, write to external APIs, and loop dozens of times before completion. That workload pattern demands sustained compute capacity, not burst training runs.
What Shortening Depreciation Means for Agent Compute Costs
For teams building on hyperscaler infrastructure, the depreciation dynamics create pricing uncertainty. If all four companies shorten useful life assumptions to match Amazon’s five-year schedule, the resulting depreciation increase flows into cloud pricing calculations. If the actual replacement cycle runs closer to three years, as the AI hardware refresh pace suggests, the cost basis for inference shifts further.
The hyperscalers currently absorb these costs while competing for AI workload share. The question is whether that absorption continues as depreciation compounds on $750 billion in annual spending, or whether inference pricing adjusts to reflect the true replacement cost of the hardware running it.