AI Infrastructure Capex Hits a Power Wall—Vertically-Integrated Players Now Ahead

The US AI datacenter boom has hit hard grid constraints; Texas just froze 20% of new pipeline capacity, forcing hyperscalers and frontier labs to own power generation while memory becomes the secondary bottleneck locking Korea into a two-year premium.

The capex numbers mask a tightening resource constraint. Hyperscalers committed nearly $600 billion to AI infrastructure this cycle, with individual deals now in the $10–50 billion range (Anthropic–Volta, TeraWulf–Anthropic $19 billion, Amazon–OpenAI $50 billion). But Texas Governor Abbott's freeze on new data center connections—triggered by 1,800 pending projects requesting 474 gigawatts, five times historical peak—signals that power supply, not capital or chips, is now the binding constraint. Duke Energy's 7.6 GW in new data center agreements and the Department of Energy's $100 billion Kentucky infrastructure project are not abundance signals; they are desperation signals. The grid cannot absorb uncontrained demand.

Vertical integration and power ownership are now the competitive moat. Nvidia's $3 billion Lancium stake, SpaceX's 10 GW deployment commitment by 2027 (with Microsoft as largest offtaker), and Anthropic's $19 billion TeraWulf lock-in all pivot away from pure compute procurement toward owning land, generation, and transmission. SpaceX's $18.4 billion quarterly capex—among the highest in corporate America—signals a hyperscaler-scale operation financed by cloud revenue. Players without power optionality face siting delays rippling through 2027 and cost premiums that compress margins. CoreWeave's Indonesia expansion and neocloud stock pressure both reflect this: traditional cloud providers cannot match integrated power-plus-compute capex or negotiate from siting flexibility. Consolidation will accelerate; fragmented players will either fold into hyperscalers or exit.

Memory supply tightens as the secondary constraint through 2028. SK Hynix's $38.1 billion fab commitment—the largest single HBM capex allocation outside Nvidia—locks Korea's dominance and ensures HBM premiums through the forecast window. TSMC is capacity-constrained and Samsung's older nodes are less suitable for AI memory; Hynix becomes the marginal supplier. Expect high-end training clusters to adopt memory-optimized architectures (longer context windows, fewer distributed GPUs) rather than raw scaling, and early-stage labs to face 6–12 month HBM lead times.

Watch three signals for 2027 headwinds: (1) CoreWeave's APAC revenue mix—geographic fragmentation pressure and rate of consolidation upward; (2) SK Hynix Q3 HBM shipments—supply signal for the memory constraint; (3) Texas pipeline recovery timeline—how many projects relocate to Kentucky, Wyoming, or overseas versus abandon. The next phase of AI infrastructure is no longer about capital abundance or GPU chip yields; it is about power access and regional consolidation. Hyperscalers with generation assets and multiregional optionality will outpace peers by 2027.