Goldman Sachs put a number on it this week: $7.6 trillion in cumulative AI infrastructure CapEx between now and 2031, growing from $765 billion annually in 2026 to $1.6 trillion by 2031. That is the clearest single data point yet on the scale of what’s being built. The environmental cost side of the ledger is also coming into focus — a Carnegie Mellon study released May 1st found U.S. data centers generated $25 billion in pollution and health-related costs in 2025 alone, with Virginia and Texas accounting for 30% of the national total. Both numbers are going to drive policy and community conversations for years.
On the build side, the BlackRock-led $40 billion acquisition of Aligned Data Centers is on track to close in the first half of this year — still the largest data center M&A deal in history. Meta’s dual-track strategy of owning AI campuses while simultaneously leasing from cloud competitors (including a $10B Google Cloud deal) is a useful template for how even the most capital-rich operators are managing capacity constraints right now. Nvidia’s Q2 earnings on May 20th are the next major signal to watch — guidance of $78 billion represents roughly 77% year-over-year growth, and the market has already priced in a beat. Reach out if anything this week connects to what you’re working on.
The most consequential infrastructure shift happening inside data center campuses right now isn’t compute or cooling — it’s power generation. Faced with grid interconnect timelines stretching 3–7 years in major markets, the largest hyperscalers are moving aggressively to build dedicated on-site generation that bypasses the public grid entirely. Meta’s Prometheus campus in Ohio is the clearest current example: two 200MW natural gas plants, operated by a Williams Companies subsidiary, targeting Q3 2026 completion and providing 400MW of total behind-the-meter capacity. PJM has approved $1.7 billion in AEP transmission upgrades for Central Ohio — but AEP has said those grid improvements are 7–10 years away from full completion. The on-site generation is the bridge.
The technology stack for behind-the-meter power is evolving rapidly across three primary approaches. Natural gas reciprocating engines and turbines are the fastest path to energized capacity — deployable in 12–24 months versus 5+ years for new grid interconnect. Battery storage systems (Tesla Megapacks, BYD, CATL) provide peak shaving, frequency regulation, and islanding capability to ride through grid disruptions. And a longer-horizon category — small modular reactors, advanced geothermal, and dedicated solar-plus-storage — is being locked in via PPAs today for delivery in the 2028–2032 window.
The economics are becoming compelling even before accounting for interconnect delay savings. Data centers that control their own generation source avoid demand charges, capacity charges, and transmission and distribution markups that can add 30–60% to the effective cost of grid power in constrained markets. At gigawatt scale, that cost differential is hundreds of millions of dollars annually. The regulatory picture is evolving: Wisconsin regulators just approved a landmark framework requiring data centers to fund 100% of grid infrastructure serving them, and 77 utility tariffs across 36 states now apply special rules to large data center loads. Behind-the-meter generation is increasingly attractive not just for speed and cost, but as a regulatory hedge.
The constraint on this approach is permitting for on-site generation — particularly natural gas. Environmental reviews, air quality permits, and local opposition to combustion facilities near residential areas are real friction points. Operators are increasingly pairing natural gas with battery storage and renewable energy credits to reduce the regulatory and ESG exposure of on-site fossil generation. The direction of travel is clear: the data center of 2030 will generate a meaningful portion of its own power, and the developers who secure those generation rights today are building a durable competitive advantage.