Meta's Compute Cloud Ends the Neocloud Moment; Power Becomes the Real Moat

Hyperscalers moved from AI capex constraint to market competitors this week, collapsing the neocloud growth thesis—but frontier labs are insulating themselves via long-term dedicated capacity and power locks that are the real 5-year strategic moat.

Meta's entry into the AI compute cloud market (selling excess GPU capacity externally) signals the end of the neocloud relief-valve narrative. CoreWeave and Nebius crashed 14–17% in a single session on the announcement, erasing $120B+ in market value. The implication: hyperscalers don't have a capex ceiling—they've crossed it and are now using absolute scale to capture long-tail enterprise demand that was supposed to be neoclouds' wedge into the market. CoreWeave's $20B funding round and $24.9B debt raise now read as the cycle peak, not a floor. Circular financing (neocloud debt backed by hyperscaler customer lock-in) works until the hyperscaler itself becomes the competitor.

Anthropic's $19B 20-year lease with TeraWulf is the counter-move: frontier labs are securing dedicated capacity outside hyperscaler control, betting that hyperscalers can't be trusted suppliers for training if they're optimizing their own competing models in parallel. TeraWulf stock jumped 10–16% on validation that long-term model-provider commitments support neocloud capex. But this model scales only if power is available. SoftBank's launch of SB Neo and Bloom Energy + Brookfield's $25B AI power partnership signal the real constraint shift: power, not compute, is the binding resource. Grid capacity is the durable moat.

SK Hynix's $26.5B IPO (largest foreign Nasdaq listing ever) and CEO warning that 2027 will be the "worst year" for memory shortage until 2030 lock in a third structural constraint: HBM supply. AI inference and training both demand HBM at exponential scales; the supply gap forces long-term contracts and pricing hedges across hyperscalers, frontier labs, and chip makers. HBM shortage isn't transient—it reshapes capex ROI for three years and forces everyone into allocation wars now.

Clear winners: TeraWulf (Anthropic lock), power infrastructure (Bloom, Brookfield, SoftBank), SK Hynix (HBM monopoly tightens). Clear losers: general-purpose neoclouds (CoreWeave, Nebius) that positioned themselves as hyperscaler alternatives rather than specialized capacity. Exposed: any neocloud holding debt concentrated in hyperscaler customers; if Meta's move spreads to Alphabet and Microsoft, financing structures break fast. SambaNova's $1B Series F at $11B valuation (with JPMorgan inference backing) is a hedge against GPU and memory monopoly, but unproven at scale.

Watch whether Meta's compute cloud becomes industry standard (Alphabet, Microsoft follow) or remains Meta-isolated. Watch TeraWulf capex execution under hyperscaler price pressure. Watch HBM spot pricing for 2027: if memory shortages materialize, chip price premiums will reverse years of semiconductor gains and force capex triage. Finally, watch whether power becomes the gating factor over compute or chips. Bloom + Brookfield's $25B marks the inflection—if grid capacity is the constraint, the entire capex playbook shifts to renewable + battery + site selection, not GPU procurement.