Hyperscalers Consolidate GPU Cloud; Memory is the Bottleneck

Meta's launch of GPU cloud pricing competition crushed CoreWeave 25% in two weeks—confirming that hyperscalers are capturing AI infrastructure returns internally while neocloudss face margin squeeze from both directions: compute commoditization and HBM shortage until 2030.

The neocloud thesis inverted this week. CoreWeave closed a $20 billion funding round (plus $24.9 billion in debt) and signed a $19 billion deal with Anthropic, positioning itself as a genuine competitor to hyperscaler compute. Then Meta announced Meta Compute: selling GPU capacity to external customers at scale. CoreWeave immediately fell 14% that day, Nebius 17%, IREN 6.5%. Two weeks later, CoreWeave is down 25% and Bernstein downgraded sector sentiment. The market has repriced the entire neocloud thesis from "enterprises waiting for hyperscaler capacity will pay margin" to "hyperscalers will underprice us to fill their own infrastructure." Meta spent $50 billion building Louisiana's largest AI data center; instead of running to capacity internally, they're capturing external margin. That makes CoreWeave and Nebius not infrastructure providers—they're commodity compute brokers.

The capital flow now runs through hyperscaler balance sheets and equity, not through neocloud rounds. Alphabet secured $80 billion from Berkshire Hathaway explicitly for AI infrastructure capex, framed as strategic equity conviction. TSMC committed an incremental $100 billion to US fabs (now $265 billion total), not as venture but as foundational capex to build out AI chip supply. Hyperscalers are raising at equity multiples and deploying capital from balance sheet; neocloudss are raising at debt multiples (CoreWeave's $24.9 billion financing is debt-for-capex, a fragile structure). If Meta undercuts neocloud pricing and takes margin, neocloud debt service becomes unsustainable and equity investors bail.

The actual constraint is memory, not compute. SK Hynix's $26.5 billion raise (the largest foreign IPO on US markets ever) came with a CEO warning: 2027 will be "the worst year" for memory shortage, with acute HBM supply crunch lasting until 2030. ASML's Q2 guidance raised again on "unrelenting AI chip production demand," but HBM scaling isn't a fab problem—it's a supply chain problem. This squeezes neocloudss twice: they can't compete on compute margin (Meta has unlimited capex), and they can't differentiate on memory (everyone is starved for HBM through 2030). CoreWeave's Anthropic deal looked like sectoral validation until the week Meta entered the market. Now it looks like a customer locking in capacity before realizing they could build cheaper.

The secondary capex wave—Japan's 140 MW sovereign AI factory with Nvidia, SoftBank's SB Neo ten-gigawatt investment, Alphabet's $80 billion from Berkshire—signals government and strategic capital moving AI compute off US hyperscaler dependency. But none of that competes with the core dynamic: hyperscalers are vertically integrating AI infrastructure, neocloudss are repriced as junior debt with unsustainable models, and memory shortage is the binding constraint through 2030. CoreWeave's stock moves tell you everything about who captures value in this cycle.