Compute has crossed from procurement into territory. Chevron’s 20-year agreement to supply Microsoft’s Project Kilby in Reeves County, Texas, is designed to deliver about 2.67 gigawatts at full build-out, with first power targeted in 2028 and final investment decision expected by year-end 2026 [E2]. That is not a cloud customer buying capacity after the grid appears. It is a cloud customer underwriting a power plant because the grid may not arrive on time.
Washington moved the same argument into nuclear supply chains. The Department of Energy announced $17.5 billion in conditional loan commitments for long-lead-time components tied to 10 large commercial reactors, across five projects of two reactors each, targeting Westinghouse AP1000 units of about 1.1 gigawatts per reactor [E1]. The stated objective is 10 new large reactors with complete designs under construction by 2030, but the practical timetable stays slow: these loans revive supply chains before they deliver electrons, and first power is unlikely before the early-to-mid 2030s even if acceleration works [E1].
Grid access now functions like a second chip allocation system. Large loads can face multi-year interconnection waits, with Northern Virginia and PJM-type timelines stretching to five to seven years or more, while US interconnection queues hold well over 2,000 gigawatts of generation and storage, more than current installed US capacity [E3]. The hyperscaler metric now runs well past cost per token and GPU count. The metric is speed-to-power: the time between choosing dirt and energizing megawatts [E3].
Chevron’s West Texas structure shows why gas is back in the AI stack. Project Kilby uses Permian Basin gas, a co-located natural-gas-fired facility, a phased modular build, selective catalytic reduction for NOx and non-potable brackish groundwater [E2]. The industrial bargain is direct: dedicated power now, emissions politics later. For AI campuses, a dispatchable gas plant near the load can beat a cleaner interconnection that arrives after the market window closes.
Nuclear carries the opposite duration profile. The DOE program is a capacity bet for the next build cycle; it will not cure data-center demand now. Energy Secretary Chris Wright said the conditional loans would help revive the supply chain for large-scale commercial reactors and accelerate construction timelines by up to three years [E1]. That matters strategically because nuclear offers large, steady blocks of power. It does not solve the 2028 campus problem unless the campus already has another source.
Capital kept chasing the other physical bottleneck: memory. SK Hynix said it plans to raise up to $29.4 billion through a Nasdaq ADR listing, with proceeds aimed at chip factories and equipment for high-bandwidth memory capacity [E5]. The listing converts AI demand into fab financing, which is the correct accounting treatment. High-bandwidth memory is not an accessory to the model boom. It is one of the places where the boom becomes concrete, steel, cleanrooms and tool orders.
Software moved as well, but the strategic layer is now attached to hardware heterogeneity. Qualcomm agreed to buy Modular for nearly $4 billion, acquiring software that Reuters described as running AI models across chips [E4]. The transaction points at the CUDA moat without pretending the moat is only silicon. If power decides where capacity sits, and memory decides how much of it can be fed, cross-chip inference software decides how much stranded silicon can become usable capacity.