Q1 2026 / capex $725B

Why it matters

Three macro updates for the firm-equivalent-mw underwrite. (1) The $725B combined 2026 capex is a $315B YoY increase from 2025's $410B — that's a ~$26B/month incremental construction-and-equipment outlay across the four hyperscalers, and it lands on a US data-center supply chain that's already absorbing the Tom's Hardware-disclosed 'half of 2026 pipeline canceled or delayed' bottleneck (covered Tuesday). The math implication: capex *commitments* are growing 77% but actual *deliverable* capacity is constrained by power infrastructure, transformer / switchgear lead times, and now memory-chip pricing pass-through; the gap between committed-capex and built-capacity is widening, which means the marginal site that actually *can* be built and powered in 2026 has a higher willingness-to-pay than 2025's marginal site. Cliff's submission-readiness wedge is selling speed-to-approval into a market where speed-to-energization is the binding constraint, and the 77% capex-growth signal validates the wedge's pricing power upward. (2) Microsoft's $25B-of-$190B memory-chip disclosure is the first time a hyperscaler has cleanly separated 'capex for capacity' from 'capex for component-cost inflation' in earnings disclosure. This matters for green-loan-data-center-financing.md because lenders underwriting against capex-as-collateral need to discount the inflation-driven portion (it doesn't add operational capacity) versus the capacity-driven portion (it does). The 13% inflation share at Microsoft is a starting benchmark; Meta's similar 'higher component pricing' disclosure suggests the share is similar across the four. Bilateral-DC-debt deals (Meta's $13B El Paso, Oracle's $300B Stargate-restructured commitment) should price against capacity-only capex, not gross capex. (3) Google Cloud's $20B Q1 (growing fastest by percentage of the four) is the strongest signal yet that hyperscaler-DC demand is being driven by *cloud* AI inference, not just LLM training — Q1 2026 is the first quarter where Google Cloud's growth rate decisively outpaces AWS and Azure on a percentage basis, and the underlying demand mix is shifting toward GPU inference (Anthropic, OpenAI competition) plus vertical-specific AI applications rather than pure compute-and-storage. The site-selection implication: latency-to-end-user starts to matter for inference workloads in a way it never did for training, which favors metro-adjacent siting (Northern VA, North Texas, Phoenix metro, Atlanta) over remote stranded-asset siting (West Texas, Wyoming, rural Utah). Watch whether 2026 hyperscaler announcements skew metro-adjacent vs remote-asset compared to 2025; if so, the BTM-gas / retired-coal-site geographic-arbitrage thesis (covered Wed/Thurs) needs to weight metro proximity more heavily in the firm-equivalent-mw model.