Microsoft

SemiAnalysis: scaling / reasoning / agentic eras; open weights close gap in ~half the time each era (Kimi K2.6 vs Opus 4.5 ~4.8 mo), but benchmarks ≠ product. See AgentX v3, CUDA moat, NVDA.

Apache 2.0 AgentX: $3M dataset, 1M+ context, 393 Claude Code traces, 95%+ KV reuse; ~2MW/1000+ chips full-stack test. DCP/PCP composability favors NVDA; post Aug 21 B200 perf/$ passed MI355X. See TileRT, open models note, AMD.

SA sees SpaceX ~10GW cumulative by YE2027 as credible: speed-first builds + pricey short contracts. MSFT signed 10GW+ YTD with a near-term gap; ~$100B/GW/yr inference economics → likely largest offtaker. See MSFT, NeoCloud.

2026-08-05 DeepMind leadership overhaul; SA sees frontier odds collapsing as compute tilts to GCP/external TPU sales. Gemini ARR ~$12B in 2Q26 vs ~$200B YE2027 GCP external AI+TPU systems narrative. See GOOGL.

Microsoft FY26 Q4: revenue $90.0B (+18%), Azure +43%, Microsoft Cloud $59.3B (+27%), RPO $678B (+84%). FY26 revenue $331.8B, OI $155.2B, diluted EPS $17.95. CY2026 CapEx ~$175B. See Amazon, NVDA.

GPUs buy silicon; returns hinge on utilization. Collectives, HBM, schedulers, power/thermal, and inference batching decide effective output from hundred-billion AI CapEx. Hub NVIDIA; complements AVGO/ANET; constraints TSM/MU; physical layer VRT. See CPU industry, AgentX v3 (agentic KV/routing).