The Lyceum: Power & Infrastructure Weekly — Mar 21, 2026

This was the week the physical world started handing the AI boom its invoice — and the invoice came from three directions at once. California regulators told data center developers to pay upfront for their own grid upgrades. An Ohio utility announced a $4.2 billion transmission project funded entirely by SoftBank's developer subsidiary. And Microsoft formally ended the use of NDAs that kept these deals in the dark. The common thread: power, water, and local politics are no longer externalities that hyperscalers can engineer around. They're the negotiation.

The White House secured signed commitments from seven major hyperscalers — the companies building the largest AI data center campuses — to build or purchase enough generation capacity and cover grid upgrade costs to avoid passing those expenses to residential ratepayers. This is a political document, not a regulatory order, but it matters because it establishes a public baseline: these companies have now formally acknowledged, in writing, that the grid costs of their expansion are theirs to bear. If the pledge holds, it gives state regulators like the California Public Utilities Commission (CPUC) and the Federal Energy Regulatory Commission (FERC) a reference point when designing cost-allocation rules. If it doesn't — if hyperscalers later argue in rate cases that some costs should be socialized — the pledge becomes ammunition for consumer advocates. Watch for the first state regulatory proceeding that explicitly cites this commitment as precedent.

The California Public Utilities Commission approved transmission upgrades for a data center in Sunnyvale on March 19, 2026, but required developer Menlo Equities to pay the full upfront cost — with refunds only if the facility generates enough revenue to justify them. This is the third time the CPUC has used this "distribution refund" model, and it's now considering making it standard. The CPUC's Public Advocates Office is pushing further: special contracts, upfront grid-impact studies, and potential "bring-your-own-grid" requirements so a single hyperscaler campus can't silently shift billions in upgrades onto retail customers. If California — the largest U.S. electricity market and historically industry-friendly on siting — codifies cost-causation into standard practice, expect Virginia, Georgia, and Texas to follow within 18 months. The failure mode is familiar: regulators create the rule but grant enough exceptions that it becomes decorative. The signal to watch is whether the CPUC's next large-load interconnection proceeding applies the same upfront-payment structure without negotiation.

AEP Ohio will build new 765-kilovolt transmission lines in Appalachian Ohio to power a planned 10-gigawatt data center campus at a former Department of Energy site in Piketon. The bill — $4.2 billion — would be paid by SB Energy, a SoftBank subsidiary, not AEP's existing ratepayers. This is the clearest example yet of a "private grid for private load" structure: the developer funds dedicated transmission, the utility builds and operates it, and residential customers are contractually insulated from the cost. If the Ohio Power Siting Board approves it and the model works financially, it becomes a replicable template for utilities trying to accommodate historic demand growth without political backlash. If permitting stalls or SB Energy's financing wobbles, it demonstrates that even willing-to-pay developers can't outrun the physical timelines of high-voltage construction. Watch the siting board docket — that's where the real timeline lives.

Microsoft announced on March 18, 2026, that it's terminating existing non-disclosure agreements with local governments worldwide and will no longer use NDAs as a default mechanism for new data center projects. The company will still use confidentiality agreements for private land purchases, but public-sector deals go transparent. This is part of a broader "community-first" initiative that includes commitments to pay full power costs, replenish more water than facilities consume, and deploy closed-loop cooling in water-stressed regions. The practical consequence: local officials now have a named corporate framework to cite when negotiating permits, and competing hyperscalers face pressure to match or explain why they won't. If Microsoft follows through operationally — publishing site-level water and energy data, not just pledges — it becomes the standard other companies are measured against. If the commitments stay vague, community opposition doesn't go away; it just gets better organized. The tell is whether Microsoft's next five site applications include published water-use-efficiency targets.

Eaton finalized its purchase of Boyd Thermal on March 12, 2026, for $9.5 billion — 22.5 times Boyd's expected 2026 EBITDA — to build an integrated "grid-to-chip" infrastructure offering for AI data centers. The thesis: as GPU racks blow past 100 kW (a level where air cooling physically cannot remove enough heat), direct-to-chip liquid cooling becomes mandatory, and the company that controls both power delivery and thermal management owns the most defensible position in the supply chain. If Eaton can actually integrate Boyd's cooling technology with its existing power distribution products and sell them as a single system, it compresses procurement timelines for hyperscalers and squeezes out smaller cooling specialists. If integration stumbles — different engineering cultures, incompatible product architectures — the $9.5 billion becomes a very expensive adjacency play. The signal: watch whether Eaton's next product catalog includes co-engineered power-and-cooling rack solutions, or whether Boyd operates as a standalone division with a new logo.

SEIA released its first sector-dedicated Energy Storage Market Outlook this week, projecting U.S. battery energy storage system (BESS) deployments will reach 70 GWh and 35 GW in 2026 — roughly $25.2 billion in capital investment and nearly double 2025's record. The report, produced with Benchmark Mineral Intelligence, forecasts annual installations potentially exceeding 110 GWh by 2030. The structural shift: BESS is transitioning from a solar complement to a standalone grid product competing in capacity markets, frequency services, and energy arbitrage. But there's a tension in the data — global BESS deployments actually fell about 25% in January 2026 compared to January 2025, according to Energy Storage News. If Q1 delivery data from ERCOT and CAISO confirms the dip, the $25.2 billion annual figure starts looking aspirational rather than inevitable. Watch interconnection queue throughput and tariff-related import delays — those are the leading indicators of whether the annual number lands.

On March 20, 2026, FERC proposed rules to prioritize interconnection for large loads like data centers while mandating more accurate demand forecasting from utilities. The draft order, open for comment until May 15, 2026, targets the 2.6 terawatt backlog in interconnection queues — a bottleneck that affects renewables, storage, and large industrial loads alike. This complements FERC's March 19, 2026 update to Critical Infrastructure Protection cybersecurity standards and its ongoing PJM co-location proceeding, where the April 17 reply deadline will reveal whether hyperscalers fight or concede on minimum transmission charges. If the proposed rules survive the comment period intact, they could materially accelerate grid access for both AI campuses and the clean energy projects needed to power them. If smaller developers get squeezed out by priority rules favoring large loads, the political dynamics shift fast. The observable signal: read the comment letters filed by independent power producers and community solar developers by May 15, 2026.

• Trane A2L-compatible chillers launched March 20, 2026, for commercial cooling applications, accelerating the transition to low-GWP (global warming potential) refrigerants — the mildly flammable A2L class that's replacing legacy HFCs under updated safety standards. These are the machines that will cool the buildings around data centers, even as the racks themselves go liquid.
• ASUS liquid-cooling ecosystem, showcased at NVIDIA GTC 2026 (March 16–19, 2026), demonstrated a direct liquid-cooling system achieving a PUE of 1.18 — meaning only 18 watts of overhead per 100 watts of compute. Per ASUS, the system is engineered for rack densities approaching 140 kW, the range NVIDIA's Blackwell architecture demands.
• BioLargo/GC Environmental AEC system completed its first commercial municipal installation for PFAS treatment, using Aqueous Electrostatic Concentration technology that BioLargo says produces lower waste volumes and energy use than conventional granular activated carbon or ion exchange. The system is positioned as an EPA validation site.

A $4.2 billion private transmission line to a former nuclear site in Appalachian Ohio. A Swedish city heating 30,000 homes with server exhaust. A mobile PFAS-destruction unit on a flatbed truck headed to Minnesota. The physical world is weird, expensive, and increasingly non-negotiable — which is exactly why the people who said "just build more data centers" are now writing checks to water districts. Forward this to someone who still thinks infrastructure is boring.

—The Lyceum

Know someone who'd want this? Send it their way.

The White House released an AI legislative blueprint on March 19, 2026, that seeks to preempt state-level regulation — directly relevant to the California and PJM cost-allocation fights above. Read → White House AI legislative blueprint and state preemption