Microsoft's 2026 capex hits $150B. AI infrastructure now dominates the balance sheet.

Microsoft's 2026 capital expenditure is running at roughly $150B, with the majority allocated to AI infrastructure — Nvidia GPU procurement, data-centre build-outs, networking, and power [Tech Insider analysis].

── What shipped ──

• 2026 capex: ~$150B (most directed at AI compute capacity)
• Anthropic Claude added to Foundry (TX_018) — multi-vendor model hosting
• OpenAI partnership renegotiated under "next phase" framing [Microsoft blog]
• Azure HorizonDB with native vector indexing for AI workloads
• Build 2026 in June expected to detail more agent and infrastructure work (TX_026)

── Why it matters ──

The capex number reframes Microsoft. Five years ago, the company was a software-and-services business with a cloud arm. In 2026, the cloud arm is the dominant capital allocator and software is increasingly the consumption layer that monetises that infrastructure.

Three structural implications.

One — depreciation cost is now load-bearing on margins. $150B of GPU and data-centre capex depreciates over 4–6 years. That is $25–37B of annual depreciation hitting the income statement going forward. Even Microsoft's enormous gross profit absorbs that, but it changes the margin profile materially.

Two — utilisation is the metric to watch. AI compute is high-fixed-cost. Underutilised GPUs are expensive idle assets. Azure's challenge through 2026 is not capacity (Microsoft has bought it) — it is selling enough inference to keep utilisation up.

Three — sovereign-AI competition increases. The U.S. hyperscaler trio (Microsoft, Google, AWS) plus sovereign programmes in EU, India, UAE, and Saudi Arabia are now all sourcing the same Nvidia silicon. Allocation matters as much as capacity. Microsoft's $150B buys access at the front of the queue.

── Editor's take ──

The capex number is the single best signal of Microsoft's confidence in the AI thesis. You don't commit $150B in a single year unless you believe the demand is durable for at least five years. The risk to that thesis is not whether AI demand exists; it is whether per-token pricing power survives the price war that Gemini 3.2 Flash, DeepSeek V4, and the Anthropic-OpenAI competitive dynamic have already started. Margins are the variable. Capacity is committed.