HWNEWS
The Grid Can't Keep Up. AI Datacenters Are Building Their Own Power.
When utility interconnection queues stretch seven years, hyperscalers stop waiting and build their own gas turbines and fuel cells on site.
When utility interconnection queues stretch seven years, hyperscalers stop waiting and build their own gas turbines and fuel cells on site.
When the power grid tells a hyperscaler to wait seven years for an interconnection slot, the rational response is to stop waiting. That is the structural story behind a new forecast from SemiAnalysis, a paid research firm focused on semiconductors and AI infrastructure, which now projects US datacenter capacity will more than triple from 21 gigawatts added in 2026 to roughly 84 gigawatts by 2030, with the share that arrives as behind-the-meter generation (power produced on the datacenter campus itself rather than drawn from the utility grid) climbing past 50% by 2028. forecast
The shift reframes a debate that has dominated AI coverage for the past year. Headlines about "datacenter delays" have usually framed the bottleneck as a demand problem: too much AI appetite, not enough capital or will. SemiAnalysis's buildout model, which works bottom-up from named chip accelerators, building permits, and grid headroom maps that mirror the reserve-margin conventions used by US grid operators (RTOs and ISOs, the regional utilities that actually balance supply and demand in real time), argues the opposite. The constraint is moving, not disappearing. Demand is being rerouted around the grid rather than cancelled.
That distinction matters because it changes who wins. Behind-the-meter generation has been a niche category for decades, used mostly by hospitals, military bases, and industrial sites that needed a hedge against outages. The AI buildout is now turning it into a structural answer. A single ~1 GW campus that would otherwise sit at the back of a multi-year utility queue can instead install on-site gas turbines, reciprocating engines, or fuel cells and run independently. The trade-off is a different cost stack, a different emissions profile, and a different set of vendors.
The new entrants are not the ones most people associate with power generation. SemiAnalysis flags Bloom Energy, the fuel-cell maker; Bergen Engines, a Norway-based maker of medium-speed reciprocating engines; and Wärtsilä, the Finnish marine-engine company that has spent the last decade pivoting toward land-based industrial power. forecast and vendor list The foil is the legacy turbine makers, GE Vernova and Siemens Energy, whose order books were widely expected to be the binding constraint. So far, those bottlenecks have eased faster than feared: SemiAnalysis notes that the early warnings about turbine capacity are giving way to a more nuanced picture in which reciprocating engines and fuel cells pick up the slack.
The equipment category is large enough to move the global supply chain. SemiAnalysis's model points to roughly 50 gigawatts per year of behind-the-meter datacenter equipment demand by 2029, a market that did not meaningfully exist two years ago. forecast
What regulators do next will determine whether the reroute holds. The Federal Energy Regulatory Commission (FERC), the agency that sets interstate electricity rules, opened Docket RM26-4-000 to rework how grid operators handle large new loads, and in mid-June it followed up with a fast-track mandate intended to accelerate grid access for data centers while shielding retail ratepayers from the cost of upgrades built mainly to serve hyperscalers. docket fast-track mandate The agency is also expected to act on its broader large-load interconnection docket by the end of June, a decision that will reshape how hyperscalers negotiate with the utilities they have been trying to bypass. docket
The first major grid operator to react was PJM Interconnection. PJM proposed behind-the-meter reforms in February as part of a data center colocation effort, then published "Powering Reliability through Market Design" in May, framing the buildout as a market-design problem rather than a generation shortage. PJM May report PJM BTM reforms PJM announcement The reforms try to thread two needles: giving hyperscalers a faster path to power while making sure the rest of the grid's residential and industrial customers do not subsidize the upgrades.
This is where the bubble question gets its sharper edge. The standard worry about an AI infrastructure bubble is that the demand is phantom: too many chips chasing too few useful applications. SemiAnalysis's framing points to a different risk. The constraint is on the supply side, specifically the speed at which the grid and equipment supply chains can scale. forecast If that is correct, the buildout may continue to grow faster than the headlines suggest, not because the demand is unreasonable, but because it is being satisfied in a way the public numbers do not yet capture. The bottleneck is not whether AI is real; it is who gets to plug it in.
Three things to watch. First, whether FERC's June large-load docket decision actually forces grid operators to price hyperscaler upgrades separately from residential rates, the mechanism that would make the cost-allocation question real. Second, whether the equipment supply chain, particularly Bloom Energy's order book and Bergen Engines' delivery slots, confirms the 50-gigawatt trajectory or whether the model overstates demand that has not yet been ordered. Third, whether PJM's market reform survives contact with its member utilities, several of whom have publicly questioned whether behind-the-meter generation is being used to skip interconnection queues that exist for reliability reasons. FERC docket PJM May report industry pushback
For now, the structural fact is simple. The bottleneck has moved from the chip fab to the gas turbine, from the foundry to the fuel cell. The grid has not stopped the AI buildout. It has rerouted it.