Hyperscaler Glut Collapses Neocloud; GPU Supply Shifts Scarcity to Commodity

Meta's entry into GPU cloud services triggered a 6-17% neocloud selloff because hyperscalers can undercut specialists with loss-leader pricing after cornering excess capacity.

The story is supply glut, not demand scarcity. An unnamed billionaire CEO announced $25B committed capex backlog; TeraWulf locked $19B with Anthropic; Hut 8 carries $19.6B in leases; TSMC committed $100B more US capex (now $265B total). Japan's Noetra consortium with NVIDIA deploys 140 MW and 27,500 Rubin GPUs as sovereign alternative. Memory bottleneck persists: SK Hynix's $713B plan locks HBM pricing through 2026-27. Yet Hut 8, Nebius, and CoreWeave fell 6-17% in one day when Meta announced Meta Compute cloud services. This was not demand weakness. It was disintermediation.

When Meta, Google, and Amazon each built 2-3x their internal GPU consumption to hedge shortage, they created 500+ MW of idle capacity. Neocloud specialists bought at premium prices (2023-2024) betting on persistent scarcity. Now hyperscalers rent that capacity to Anthropic and OpenAI at $1.5-2M/MW/year—below neocloud cost basis—because they monetize the AI apps, not the infrastructure. CoreWeave, Nebius, IREN monetize only the VM. They cannot compete. CoreWeave is down 40% from 2026 highs; IREN hit new lows. SpaceX AI's gigawatt sustains $28M/MW/year revenue, but that is inference-optimized premium, not training commodity where hyperscaler discounting controls.

Geography is fracturing US data-center hegemony. TSMC's $265B US commitment, Japan's sovereign Noetra consortium, and China's AI summit signal polyvalent competition. Moonshot AI's Kimi K3 reportedly matches Anthropic's Opus 4.8 on a 300-person team despite export controls—narrative risk if talent and open-weight models offset Western hardware advantage. Brownfield data centers validated the capex arbitrage: TeraWulf's $19B Anthropic deal and Hut 8's $9.8B second lease prove fossil-fuel plant conversion. Hut 8's 1 GW Beacon Point is fully leased on $19.6B in contracts. The neocloud crisis is triage, not collapse. Survivors are those with captive workloads or structural capex edges.

Watch three signals for the final verdict. First: does SK Hynix sustain HBM pricing as Samsung and Micron ramp HBM4 and NVIDIA Vera enters production Q3. Second: M&A velocity—will hyperscalers acquire neocloud operators or will private capital back independent survivors. Third: does Moonshot's efficiency narrative hold, triggering a valuation rerating off compute-scarcity assumptions. Neocloud is a bellwether. If week two brings another selloff, markets have priced structural oversupply permanent.