The Power Calendar Became the Chip Calendar — 72 Hours Interconnection Hardened Into AI Capex's Single Bottleneck

The most consequential semiconductor news of the past 72 hours did not happen inside a fab.

Stream Data Centers walked away from an $800M, already-approved campus in South Carolina because the utility interconnection timeline was longer than the AI demand window the project was sized for. The same week, the Department of Justice filed in court to defend xAI's Memphis gas-turbine data center, with the Department of Defense formally certifying the site as critical to national security. The same week, the Department of the Interior paid Invenergy $765M to surrender four offshore wind leases — proceeds explicitly redirected to new natural gas plants in Indiana, Wisconsin, and Ohio that will, among other things, serve incoming AI load. The same week, FERC ordered every US RTO to either justify or reform its data center interconnection framework, a GOP senator introduced a bill to put large-load grid hookups under FERC's exclusive jurisdiction, and Maryland's ratepayer advocate filed a FERC complaint alleging PJM was loading data-center-driven transmission costs onto residential ratepayers.

That is not five unrelated power stories. That is a single fact pattern: in June 2026 the binding constraint on US AI capex is no longer wafers, no longer HBM, no longer foundry capacity, no longer even raw generation. It is interconnection. The on-ramp. The transformer, the substation, the queue position, the calendar a utility hands a hyperscaler when they ask for 800 megawatts.

Foxconn quietly gave the market the unit economics that frame the entire conversation. A single gigawatt of NVIDIA Vera Rubin AI datacenter capex is now ~$47B, with annual power bills of ~$1.3B per GW. Meta's just-signed Crusoe deal alone is 1.6GW. Microsoft broke ground on a 17-building Indiana campus the same 72 hours. Layer that demand onto a US grid where Stream just demonstrated that a fully sited, fully capitalized, fully approved project can die because the power calendar doesn't match the silicon calendar, and the implication is clear: capital is liquid, copper isn't, and the bottleneck has migrated downstream of every chip company in our coverage.

This re-prices the AI stack in three specific ways.

First, it changes who actually shows up to take delivery of NVIDIA's silicon. Vera Rubin shipments through 2027 implicitly assume the buyer can energize their rack. CoreWeave's Vera Rubin NVL72 deployment and the Bull/Foxconn European production line for sovereign AI both got built around hyperscalers that already had grid certainty — not around the long tail of speculative builders. When Stream walks away from $800M because of a timeline, every Tier-2 colo project in the PJM, ERCOT, and MISO queues becomes a marginal NVDA shipment with optionality, not a confirmed bid. NVIDIA's order book is no longer constrained by its ability to fab; it is increasingly constrained by its customers' ability to energize.