Microsoft's AI Strategy Deconstructed – From Energy to Tokens (newsletter.semianalysis.com)

Microsoft’s AI playbook swung from full throttle to a dramatic “Big Pause” and now back into overdrive. After an aggressive 2023–24 buildout to host OpenAI — including massive “Fairwater” training clusters (300 MW GPU buildings housing on the order of 150k GB200 GPUs per building, multi-campus plans >2 GW IT capacity, and an ultra-fast AI WAN design at ~300 Tb/s scalable toward Pb/s) — Microsoft paused large portions of pre-leased and self-built capacity (~3.5+ GW frozen). That gap let rivals (Oracle, CoreWeave, Google, Amazon, et al.) win sizable OpenAI compute contracts in 2025. The company has since reversed course, signing a new OpenAI deal, scrambling for short‑term capacity (self-build, leasing, “Neocloud” rentals, remote sites) and even gaining access to OpenAI’s custom ASIC IP — a potential lever to serve OpenAI models or accelerate Microsoft’s own Maia model training. Technically and economically, this shift matters because demand for accelerated computing is surging and the industry’s margin map is reconfiguring. The report frames Microsoft across the “AI Token Factory Economics Stack” (Chips → System architecture → IaaS/PaaS → Models → Applications), noting Nvidia’s outsized chip margins and high model/API margins (60%+ for leading model makers). Microsoft aims vertical integration to capture more token economics — but faces execution risk, lost contracts (e.g., “Stargate”) and fierce competition in bare‑metal GPU provisioning. Short term, expect accelerated Azure capacity builds, tighter spot markets for GPUs, and intensified chip and networking strategies as hyperscalers race to convert model capability into revenue.