The Lyceum: Power & Infrastructure Weekly — Jun 27, 2026

The AI power story just stopped being about megawatts and started being about jurisdiction. FERC put all six federal grid operators on a clock, Tesla and Sunrun pointed 16 gigawatts of home batteries at hyperscalers, and BloombergNEF sketched a future where data centers eat as much electricity as the entire U.S. nuclear fleet produces today. The more interesting action ran underneath the headlines: states slamming the brakes on permits, NVIDIA quietly making cooling water optional, and a Texas queue revealing that "approved" and "actually drawing power" are two very different numbers.

• Tesla–Sunrun–Renew Home Virtual Power Plant (Sunrun, Tesla, Renew Home): A capacity framework aggregating hundreds of thousands of home batteries plus 8 million-plus smart thermostats into a single dispatchable resource aimed at data centers — no new hardware, land, or water required.
• NVIDIA 45°C AI Factory Cooling Reference Design (NVIDIA): A direct-to-chip liquid cooling architecture running warm enough to reject heat through dry coolers, claiming near-zero on-site cooling water versus the ~2.6 million gallons per megawatt-year of conventional towers.
• FERC Section 206 Show-Cause Orders (FERC): Six simultaneous tailored orders directing every FERC-jurisdictional RTO/ISO to justify or rewrite large-load interconnection rules on a 60-day clock.
• Voltus 100 MW VPP for Google (Voltus): A Google-funded aggregation of up to 100 MW from local batteries, thermostats, and flexible business assets — the second hyperscaler VPP deal in a single week.

How a 500-megawatt data center plugs into the grid — and who pays — just got a federal answer faster than almost anyone expected.

On June 18, FERC issued tailored show-cause orders under Section 206 of the Federal Power Act to each of the six regional grid operators under its jurisdiction, directing them to justify or reform the rules governing how large energy users connect. Crucially, FERC skipped the usual Notice of Proposed Rulemaking — a multi-year affair — in favor of customized orders that put operators on a 60-day compliance clock. The stated aim, per the Insurance Journal, is to handle power requests within 90 days, a process that currently can take years.

The tradeoffs are explicit. Per the Institute for Energy Research, data centers will bear the full cost of necessary grid upgrades, and hyperscalers could be required to bring their own power or curtail demand during system stress.

What changes if it works: the interconnection bottleneck loosens for hyperscalers while shielding ordinary ratepayers from the upgrade bill.
What failure looks like: NARUC and state regulators challenge FERC's jurisdiction and the whole thing slides into court. Watch which RTO files a concrete large-load tariff rewrite first — per McGuireWoods, intervention notices are due July 9 and generation-adequacy reports July 20. That first filing becomes the template everyone else reacts to.

The most creative answer to the AI power crunch this week didn't come from a utility or a reactor. It came from millions of American living rooms.

Sunrun, Renew Home, and Tesla announced an agreement to deliver more than 16 gigawatts of flexible capacity to hyperscalers and utilities, aggregating hundreds of thousands of home batteries plus more than 8 million smart thermostats into a software-coordinated virtual power plant — a system that coordinates thousands of distributed devices so they dispatch like a single generator, per StockTitan. No new hardware, land, or water.

The headline number deserves scrutiny. Per Tech Times, analysts say only 4 GW is firm battery capacity, and just 300 MW is deployable now in Virginia. Per Yahoo Finance, this is a capacity framework, not signed hyperscaler contracts — yet Sunrun's stock still jumped more than 25% on the session.

What changes if it works: rooftop solar stops being a retail product and becomes grid infrastructure with a recurring-revenue buyer that has unlimited urgency. The signal that tells you which way it's going: a named hyperscaler signing an actual offtake agreement against this framework. Until then, it's a very large press release.

PJM Interconnection covers 13 states and 67 million customers from Illinois to New Jersey. It is the single most important piece of electricity infrastructure in the country. This week, Bloomberg asked whether it's become too big to function.

Per the Insurance Journal, PJM has drawn particularly heavy criticism for being slow as power bills surge. Bloomberg's reporting, citing the AI data center boom, raises the question of whether PJM's geographic sprawl — 13 states with wildly different regulatory philosophies and load-growth profiles — makes it structurally unable to respond to hyperscale demand at speed.

What changes if PJM splits: separating the Mid-Atlantic data center corridor from the Midwest industrial load would reshape capacity markets, transmission cost allocation, and every hyperscaler's site model overnight. What tells you it's getting real: a governor or congressional delegation formally requesting a FERC study of PJM's structure. That's the trigger that turns a Bloomberg headline into a docket.

Every few months a new forecast lands that makes the last one look conservative. This week's came from BloombergNEF, reported by Utility Dive: U.S. data center power demand reaching 106 GW by 2035 — roughly the entire current electricity consumption of Germany, dedicated to computing.

For scale: the entire U.S. nuclear fleet today produces about 100 GW. Per the Electric Power Research Institute, data centers currently account for 4–5% of U.S. electricity demand, but that share could triple by 2035. The sector is on track to need more power than every nuclear plant in America combined, within a decade.

Why the cooling and water story is inseparable: a 106 GW fleet needs liquid cooling, chillers, and towers at a scale the HVAC industry has never built. If even a fifth of that relies on evaporative water, it becomes a municipal planning crisis in drought-stressed regions. Watch for state water boards in Arizona, Nevada, or Texas citing power-demand forecasts in water allocation proceedings — the moment the power and water stories formally merge in regulatory filings.

While the U.S. debates whether to keep existing clean energy tax credits, Canada wants to build an entirely new nuclear fleet from scratch.

Per CBC News, the Canadian federal government unveiled a national nuclear strategy targeting up to 10 new reactors by 2040 — and explicitly named data center power demand as a driver. That's new language for a national nuclear policy document, which usually frames buildout around decarbonization or energy security. The plan leans heavily on small modular reactors (SMRs) — compact, factory-built plants deployable faster than conventional gigawatt-scale units. Canada already has the furthest-along SMR project in North America at Ontario Power Generation's Darlington site, using GE Hitachi's BWRX-300 design.

What success requires: a pressure-vessel, steam-generator, and instrumentation supply chain that doesn't currently exist at this scale. What tells you it's real rather than rhetorical: a formal procurement process with committed capital. Until OPG or Ottawa issues an RFP for the first BWRX-300 with money attached, this is a policy announcement, not a construction program.

The cooling industry has chased one goal for years: reject the heat without spending the water. NVIDIA just published the blueprint.

NVIDIA's reference design for AI "factories" — its term for GPU clusters running continuous training — operates at a 45°C coolant supply temperature, warm enough to dump heat to ambient air through dry coolers instead of evaporative towers. Conventional air-cooled facilities use towers that consume roughly 1.8 liters of water per kilowatt-hour of IT load; a 100 MW cluster running year-round evaporates around 1.6 billion liters annually — what a city of 30,000 uses. The NVIDIA architecture, per its own announcement, eliminates that, falling back on chillers only in the hottest 1% of hours.

The catch: it's a greenfield architecture, not a retrofit kit. It needs liquid-ready infrastructure from day one and coolant distribution units from vendors like Vertiv or Schneider Electric. The observable signal: hyperscaler RFPs for new AI campuses specifying 45°C liquid cooling as a baseline. That would reshape the entire cooling supply chain within 18 months and make water-stressed regions suddenly viable to build in.

The power story and the water story ran on parallel tracks for two years. This week they converged — and local water utilities aren't happy.

Per Broadband Breakfast, FERC's interconnection action arrives as backlash grows against data centers over the energy and water they consume, plus fears about noise, air pollution, and lost open space. The core problem: evaporative-cooled data centers can draw millions of gallons a day, and most of that demand wasn't in municipal capacity plans written five or ten years ago.

Why this is the binding constraint: a data center that gets grid power in 90 days under FERC's new framework still needs a water supply agreement, a discharge permit, and a capacity study — all governed by state and local authorities FERC has no jurisdiction over. The NVIDIA closed-loop design is the technical answer, but it's a design-stage decision, not something you bolt on after a utility sends a letter. Watch for the first state water board to condition a data center operating permit on closed-loop cooling — the first hard regulatory link between the two stories.

This week: a nation's worth of home thermostats drafted into an AI army, six grid operators handed identical homework with different due dates, and a Texas queue where two-thirds of the "approved" megawatts politely declined to show up to peak. The funniest part is that after all the federal acceleration — the 90-day clocks, the speed-to-power funding, the show-cause orders — the thing most likely to actually stop a data center is a county zoning board in Maryland deciding it would rather not. Build fast, get paused locally.

Forward this to the engineer who keeps telling you cooling and water are someone else's problem.