Morgan Stanley Predicts 2026 AI Leap

Morgan Stanley’s in‑house “Intelligence Factory” model projects a net U.S. power shortfall of roughly 9–18 gigawatts through 2028. finance.yahoo.com The report says developers are already bypassing grid delays by converting Bitcoin‑mining sites to high‑performance compute, signing private power contracts, and deploying mobile gas turbines and fuel cells to power new data centers. finance.yahoo.com Morgan Stanley documents executives forecasting roughly tenfold increases in training compute at frontier labs, and public filings/announcements show xAI’s Grok 3 pretraining used about 10× the compute of its predecessor on an estimated ~100,000 NVIDIA H100 GPUs. finance.yahoo.com NVIDIA has moved to productize embodied AI with a full‑stack robotics push at CES 2026, unveiling an Isaac GR00T foundation model and simulation tooling aimed at shortening sim‑to‑real cycles for manipulation and humanoid workloads. investor.nvidia.com OpenAI has reconstituted a dedicated robotics effort and renewed partnerships with companies like Figure to bind foundation models to humanoid platforms, while Tesla continues Optimus development with limited internal production runs reported amid delays to original mass‑production targets. venturebeat.com The Morgan Stanley thesis singles out semiconductors, critical minerals and off‑grid/backup power as the primary “infrastructure winners,” naming Nvidia among high‑conviction picks for 2026 and outlining an emerging “15‑15‑15” data‑center economic pattern (15‑year leases, ~15% yields, ~$15 per watt). stackalpha.io Concurrent public projects to watch include multi‑vendor U.S. supercomputer deals (announced NVIDIA partnerships delivering a combined ~2,200 exaFLOPS of Blackwell GPU performance for government systems) and Morgan Stanley’s note that industry voices have raised the prospect of recursive self‑improvement appearing as early as the first half of 2027. tomshardware.com