The GPU Cloud Fractures—Winners Own Substrate, Not Capacity

Hyperscalers entering GPU supply breaks the scarcity moat; neocloud stocks reprice downward while substrate/memory and government-backed plays capture durable advantage.

Meta's $10 billion Anthropic compute lease and SpaceX AI's $28 billion annualized revenue (3-5x traditional cloud) prove the GPU cloud model works at scale—and prove hyperscalers are suppliers, not just buyers. CoreWeave is down ~40% YTD; Nebius and IREN cratered 6-17% on Meta's cloud announcement. The scarcity premise that justified 2024-2025 neocloud valuations is extinct. GPU capacity commoditizes; margins compress.

NVIDIA Vera entering production with confirmed HBM4 supply (SK Hynix, Samsung, Micron shipping Q3) is repricing inference on cost-per-token, not raw MW. SpaceX AI's 3-5x revenue premium works because Vera cuts token costs 2-3x. TSMC's $100 billion incremental US capex and SK Hynix's $713 billion plan aren't capacity builds—they're substrate, packaging, and memory plays. Whoever owns the middle layers owns margin through 2027. Neocloud generalists without architecture differentiation face structural compression.

Moonshot AI matching Anthropic's Opus 4.8 on a 300-person team under US export controls eliminates the Western compute-moat story. If Chinese labs compete on efficiency plus open-weight models, the "$10 billion to stay ahead" capex thesis collapses. Demand peaks sooner. Western AI lab margins compress. This directly undermines neocloud growth assumptions—capex demand deceleration plus commodity pricing equals valuation reset.

Japan's Noetra (140 MW, 27,500 Rubins) and TSMC's $265 billion US expansion are geopolitical hedges, not cost-competitive builds. Hut 8's second $9.8 billion lease (Beacon Point) proves brownfield-to-AI customer lock-in works; CoreWeave's generalist model does not. Watch: (1) TSMC US fab yields—does $265 billion actually hit Taiwan cost parity or embed a 20-30% geopolitical premium? (2) Chinese lab capex growth—do Moonshot or peers raise $10 billion for domestic fabs before export controls tighten? (3) Hyperscaler utilization—if Meta's own gigawatt capacity saturates before 2027, lease upside evaporates and neocloud margins stay compressed.