Capital Conquered: Power and Memory Become the Binding Constraint

Hyperscalers have solved the capital puzzle—$1T+ in financing consortiums—shifting the binding constraint to power generation and memory supply.

Two mega-consortiums this week—Goldman Sachs' $500B asset-backed securitization and Nvidia's parallel $500B financing program with Apollo, BlackRock, Blackstone, Brookfield, and KKR—signal that capital is no longer the constraint in AI infrastructure capex. These programs remove friction from compute and data center purchases, enabling hyperscalers and neocloud providers to move capital at scale without traditional debt-market friction. The Anthropic-Riot $9.1B lease (largest single AI DC lease on record) validates this: $9.1B can now flow to a single project without disrupting the capital stack.

But hyperscalers are voting with capex on what actually constrains. Amazon is building a 7.65-gigawatt natural gas plant in Texas (largest single-operator power commitment to date, authorized for 33 million tons annual CO2) to fuel one AI data center. Nvidia invested up to $3B in Lancium, a Texas power developer, to secure long-term energy supply. Anthropic's $9.1B Riot lease is fundamentally a power deal—191 MW dedicated capacity. Hyperscalers are moving from buying compute from providers to buying integrated infrastructure including power generation, and the power molecule is now the transaction center.

Memory supply is the second structural constraint. SK Hynix' $38.1B fab capex (largest single fab commitment) directly addresses HBM scarcity—a problem that scales with GPU adoption and will persist through 2028. Specialty chip makers are simultaneously carving territory: Broadcom guided $56B AI revenue in 2026 (180% YoY growth), establishing a $50B+ structural market in AI cluster networking and custom switching. Cerebras is shipping inference acceleration (14x faster throughput at 750 tokens per second) to OpenAI's Ultrafast mode, proving that fragmentation of the Nvidia-dominated inference stack is now structural.

CoreWeave's Q2 results (revenue $2.58B, up 112% YoY; backlog $104B) and Nebius' 454% YoY growth prove specialized GPU cloud operators sustain margins at scale. A100 GPUs remain profitable nine years in. IBM's $240M infrastructure deal with Together AI and JLL's 1H 2026 data center absorption of 25 GW (record) with 1% vacancy confirm inference and open-model infrastructure are distinct durable markets; supply-demand tightness persists through 2027.

Core Scientific's $444M acquisition of Polaris and 1.5 GW Muskogee expansion consolidates independent power-constrained compute suppliers; Riot and Core Scientific are now tier-1 providers. Winners: operators who own or control power (Riot, Amazon, CoreWeave partnerships), memory suppliers scaling to constraint (SK Hynix, Samsung), and specialty chip makers fragmenting the Nvidia moat (Broadcom, Cerebras, custom silicon). Exposed: traditional data center landlords without power assets, and undifferentiated GPU cloud providers.

Watch SK Hynix and Samsung's HBM ramp-up through 2027—memory scarcity will throttle GPU utilization if capex doesn't scale. Monitor power grid capacity in Texas, Virginia, Midwest; if regional grids hit ceiling, hyperscaler capex shifts geography or forces PPAs with renewables. Track whether specialty chips (Cerebras, Groq) can scale inference volume beyond OpenAI; if so, Nvidia's moat fragments further and software lock-in weakens.