For two years, the AI-infrastructure narrative was dominated by GPUs. The composition of hyperscaler spending now points to a broader reality: an accelerator becomes revenue only after the building is ready, power is stable, networking is connected and cooling works.
Microsoft spent $31.9 billion in capital expenditures in the March 2026 quarter. Roughly two-thirds went to short-lived assets such as GPUs and CPUs; the rest went to long-lived assets, including data centers designed to support monetization for fifteen years and beyond. The company added another gigawatt of capacity during the quarter, yet said demand continued to exceed available supply. Microsoft FY2026 Q3 earnings call
Four numbers point to the same physical stack
Expected calendar-2026 capital expenditures; the latest quarter was $31.9 billion and the following quarter was expected to exceed $40 billion.
Official callAlphabet's 2026 CapEx guide; first-quarter 2026 capital expenditures were $35.7 billion.
Official callSEC 10-QMeta raised its 2026 range, citing higher component pricing and additional data-center costs for future capacity.
Official releaseAmazon's first-quarter 2026 cash CapEx, up from $24.3 billion a year earlier, primarily reflecting technology infrastructure that supports AWS growth.
Official releaseAccounting note: reported CapEx, cash CapEx and finance-lease treatment differ across companies. The figures show direction and magnitude; they should not be summed into a clean efficiency comparison.
CapEx now travels through six bottlenecks
- 01Accelerators and servers
GPUs, CPUs, custom silicon, memory and networking create the short-lived asset layer.
- 02Campuses and long leases
Land, buildings, finance leases and long-term capacity contracts lock demand in for years.
- 03Grid access and energy
Interconnection, substations, gas and renewable projects determine when a campus can energize.
- 04Cooling and density
As rack density rises, thermal design limits how much compute can actually be deployed.
- 05Commissioning
Installation, testing and software optimization turn physical assets into billable capacity.
- 06Depreciation test
If revenue ramps more slowly than depreciation and operating cost, returns come under pressure.
Do not stop at the headline CapEx number
Microsoft
Watch the short- versus long-lived asset mix, delivery speed and when incoming supply becomes Azure-ready capacity.
Official scheduleMeta
Watch the $125–145 billion range, additional data-center costs and the relationship between depreciation and AI monetization.
Official scheduleAmazon
Watch AWS growth, the full-year spending frame, Trainium and GPU deployment, and the free-cash-flow impact.
Official scheduleLong commissioning cycles amplify the investment risk
The thesis weakens if cloud growth slows while CapEx keeps rising, energization and equipment delivery slip, utilization disappoints, or depreciation and data-center operating costs grow materially faster than the revenue they support.
The leading indicators are therefore interconnection progress, usable megawatts, time to revenue-ready capacity, backlog conversion and depreciation growth relative to revenue.
IDC ATLAS VIEWThe AI infrastructure value chain now extends from a single accelerator to the entire campus. High-density capacity that can be powered and commissioned on schedule will set the pace of supply.