Power, Not Chips, Is Now AI's Constraint

Hyperscalers have pivoted from competing for chips to securing long-term power; grid scarcity in Texas and memory bottlenecks are now the real pacing factors for AI deployment.

The capex cycle has inverted. For two years, GPU availability was the choke point; now it is power and memory. Nvidia's $3B stake in Lancium, Amazon's 7.65-GW Texas gas plant, and SpaceX's 10-GW commitment by 2027 show hyperscalers now own generation infrastructure, not rent it. SK Hynix's $38.1 billion fab commitment—targeting AI memory—matches or exceeds traditional compute capex and locks Korea's HBM dominance. Nvidia's $500B financing consortium and Anthropic's $9.1B lease to Riot validate gigawatt-scale neocloud financing. Capital is no longer the bottleneck. Physics is.

Governor Abbott's freeze on new data center approvals has effectively rationed 20% of the US pipeline. This is not political theater—it is grid emergency. Amazon, Nvidia, and SpaceX's parallel bets on on-site generation signal that public grid capacity in high-value regions is spent. Constrained Texas capacity will slip 2027 deployments; the next race is long-term nuclear PPAs (power-purchase agreements) and distributed renewables—both multi-year permitting battles.

The neocloud model is validating at scale. CoreWeave's Q2 revenue ($2.58B, 112% YoY) and $104B backlog prove legacy GPU fleets stay profitable in multi-year cycles when power is locked in. Hut 8's 1-GW commercialization at Beacon Point and IBM's $240M inference commitment to Together AI signal sustained demand for open-source AI infrastructure beyond closed OpenAI/Anthropic ecosystems. But profitability hinges entirely on long-term power contracts. Tighter grids compress neocloud margins if energy access becomes contested.

Memory is the parallel constraint that will extend scaling timelines. SK Hynix's $38.1 billion fab commitment demonstrates that HBM availability is decoupled from GPU speed and runs 18–24 months behind hyperscaler demand. Nvidia ships GPUs; Samsung and SK Hynix set the pace for memory. This cascades into deployment delays for training and inference infrastructure, shifting the critical path from chip design to fab capacity.

The real race is now for power contracts, memory supply locks, and geographic grid diversification. Watch which hyperscalers secure long-term nuclear PPAs and build out regions beyond Texas. Watch SK Hynix and Samsung announce their next fab commitments as memory scarcity becomes undeniable. If Texas remains frozen through late 2026 and into 2027, delaying a 100-MW campus will cost more than the hardware. The winners are those who locked in power 12–18 months ago.