Hyperscalers Seal the Deal: Compute Glut Crushes Neocloud, Memory Becomes the Moat

Hyperscale GPU glut from Meta and Alphabet displaces independent cloud providers, while memory shortages through 2030 become the true infrastructure bottleneck.

The AI infrastructure market just bifurcated. Meta announced this week it will sell excess GPU capacity externally; CoreWeave crashed 14% in response, Nebius 17%, as investors realized that hyperscaler-subsidized compute has made independent GPU clouds economically unviable. Meta's $50B Louisiana complex and $13B Canadian campus are being built 2-3x oversubscribed for internal use, creating pricing leverage that pure-play compute brokers cannot replicate. Alphabet's $80B Berkshire backing, TeraWulf's $19B Anthropic partnership, and SoftBank's entry with SB Neo at 10GW all point to the same conclusion: scale wins, and only captive capacity or long-term contracts create defensible returns. The neocloud thesis—that independent compute would diversify the stack—has collided with economics.

Memory is the binding constraint. SK Hynix raised $26.5B this week (the largest foreign IPO on US markets), with its CEO explicitly warning that 2027-2030 will be the "worst years" for HBM shortage. With only three memory makers globally and AI scaling demand already exhausting supply, chipmakers will be forced into forward contracts at 30-40% premiums. Hyperscalers absorb these lock-in costs through scale; independent neocloud operators face margin compression that no pricing power can offset.

Power grids are the secondary constraint. Meta's site selection for Canada and Louisiana was driven by gigawatt-scale grid access. This signals that regional power infrastructure, not chip supply, will throttle capacity growth through 2027. Hyperscalers can negotiate bilateral grid partnerships; standalone operators cannot. Result: infrastructure concentration in clustered regions, defensible moats for hyperscalers with power contracts, and higher execution risk for broadly distributed workloads.

Winner: memory makers (SK Hynix), hyperscalers with integrated power partnerships (Meta, Google), model providers with capacity locks (Anthropic). Exposed: neocloud brokers (CoreWeave, Nebius), chipmakers without secured memory contracts (AMD), enterprises overcommitted to neocloud capacity. The critical watch: whether SoftBank's 10GW entry can disrupt through subsidy or operational efficiency—if not, neocloud shakeout continues into 2027.