← Back to feed

Microsoft’s Q3 FY2026 earnings call delivered the clearest look yet at what frontier AI infrastructure actually costs at enterprise scale — and what it is generating in return.

Revenue: $37B AI ARR, Up 123%

CEO Satya Nadella reported that the company’s AI business reached an annualised run rate of $37 billion in Q3 FY2026, up 123% year-on-year. For context: Microsoft’s AI ARR was $13 billion as of Q2 FY2025. The business has nearly tripled in five quarters.

The company hasn’t published a precise definition of what counts as its “AI business” — it likely spans Copilot licensing across M365, Azure OpenAI Service consumption, and GitHub Copilot subscriptions. The number excludes cloud infrastructure revenue that is AI-adjacent but sold under standard Azure pricing.

Infrastructure: $190B Capex, $25B Just for Hardware Inflation

CFO Amy Hood raised Microsoft’s calendar 2026 capex plan to $190 billion. Q4 FY2026 spending is expected around $40 billion, meaning the remaining two quarters of calendar 2026 carry approximately $120 billion combined.

Hood disclosed that $25 billion of the capex increase is attributable directly to cost inflation in DRAM, flash, wafers, and substrates — not to additional capacity. The hardware market is not getting cheaper fast enough to absorb demand, and Microsoft is absorbing the delta in capital rather than slowing deployment.

For calendar 2027, Hood said capex spending would “significantly increase” further.

Capacity: 1 GW Added in Q3, Doubling in Two Years

Nadella confirmed that Microsoft added 1 gigawatt of data centre capacity during the March quarter and reiterated the company’s commitment to doubling total capacity — measured in gigawatts — within two years.

The company’s Azure footprint: 80+ regions globally, 500+ datacenters, 190+ PoP locations, and over 800,000 kilometers of owned network fiber. Total draw is estimated in the low double-digit gigawatts and growing.

Nadella also noted that Cobalt, Microsoft’s homegrown Arm server CPU, is now deployed across half of its regions. The Braga Maia 200 XPU — announced in January and positioned as a homegrown accelerator alternative to NVIDIA GPUs — is live in Iowa and Arizona datacenters.

The OpenAI Dependency Is Unwinding

The earnings call was notable for what Nadella did not say. Microsoft and OpenAI’s exclusive licensing arrangement has been loosening for a year; the exclusivity on OpenAI’s GPT models is gone, and OpenAI is now on AWS Bedrock and Google Cloud. Microsoft will lose a portion of its OpenAI-driven Azure revenue as those workloads diversify.

The counter-move is Microsoft’s own model stack: three in-house models launched in recent months, Phi-family small models for edge inference, and the Maia 200 as a hedge against NVIDIA pricing power. The $37B AI ARR and $190B capex figure represent a company betting its next decade on infrastructure ownership, not resale rights.

For scale: $190B in 2026 capex across the four hyperscalers (Microsoft, Google, AWS, Meta) is approaching $500B combined. That number was $130B as recently as Q1 2026.