SemiAnalysis: US Grid Heading Toward 40GW+ Behind-The-Meter Datacenter Capacity by 2028 as Grid Constraints Bite
Research published June 25, 2026
The US power grid is approaching a structural breaking point that will force more than half of new datacenter builds into behind-the-meter configurations by 2028, according to new analysis from SemiAnalysis. The research shop's newly developed Energy Model projects that BTM solutions will power over 50% of new US datacenters starting in 2028, with the total addressable market for datacenter BTM equipment crossing 50GW per year by 2029. The shift reflects a harsh reality: available grid headroom turns negative by 2027, even as datacenter demand accelerates from 21GW in 2026 to 84GW by 2030.
Grid Headroom Vanishes While Datacenter Demand Surges
SemiAnalysis tracked 40,000 generation assets across the US and modeled quarter-by-quarter commercial operation dates to reach a sobering conclusion about new grid supply. The firm estimates barely 15GW of net-new Effective Load Carrying Capability being added annually, trending toward 20GW by decade-end. That's the firm capacity grid operators can actually recognize to serve datacenter load after accounting for forced outages and reliability requirements. When netted against peak demand and required reserve margins, the available headroom already approaches zero and turns decisively negative by 2027.
The constraint isn't temporary. Near-term gas turbine additions remain anemic, with less than 10GW per year projected for 2026 and 2027, only picking up in 2028 and beyond. SemiAnalysis points to a cascade of bottlenecks: institutional queue friction, permitting denials that killed 24GW of fully contracted PJM projects since 2020, and supply chains stretched to breaking. Gas turbine and main power transformer lead times have extended to three to four years versus a historical 18 months, pushing combined-cycle gas turbine projects to four to six year development timelines from initial planning to energization.
Renewables and storage offer little relief when measured correctly. While solar and battery storage each add over 20GW of nameplate capacity per year, their contribution on an ELCC basis is minimal. SemiAnalysis notes that as solar penetration rises, the marginal value of incremental capacity declines sharply because all plants generate at roughly the same hours. In some ERCOT congested zones, planners now credit incremental solar at essentially zero when sizing firm capacity needs. Storage faces the same declining marginal ELCC problem: as 4-hour battery systems saturate the grid's exposure to sub-4-hour reliability events, the system's residual risk shifts to longer-duration events that those batteries cannot address.
BTM Wins on Speed and Timeline Certainty
For AI labs and hyperscalers, the calculus has shifted decisively toward behind-the-meter. Speed and certainty dominate the decision. Requested BTM in-service dates cluster around 2027 to 2028, against grid timelines that routinely slip toward 2030. More critically, BTM timelines sit in the buyer's hands rather than with utilities whose promised delivery schedules have proven notoriously unreliable. SemiAnalysis notes that utilities increasingly push back or revise down committed load with little to no contractual penalties, a dynamic that simply doesn't work for AI labs whose access to large-scale compute is the lifeblood of their business.
The economic and operational context reinforces the shift. Power as a percentage of total cost of ownership is mostly insignificant according to SemiAnalysis's AI Cloud TCO Model, meaning any amount of power secured by an AI lab translates to billions in value given the revenue potential per gigawatt. Meanwhile, uptime requirements have relaxed. Many of Meta's self-built AI datacenters now target just two nines of uptime and forgo backup generators entirely, removing historical cost barriers to BTM adoption. SemiAnalysis observes that the main challenge and cost driver for BTM has always been redundancy and reliability, but now that customers accept lower redundancy levels, the economics of grid versus BTM are much more balanced.
The practical result is a surge in BTM activity, particularly in Texas where permitting onsite gas is easier. SemiAnalysis reports that many top-tier developers are planning 5GW-plus behind-the-meter facilities in ERCOT, with the firm's Datacenter Model separating the few credible 5GW-plus campuses from those showing no sign of serious development.
ERCOT's Batch Zero Process Codifies Hybrid Structures
ERCOT is racing to formalize the hybrid structures that blend onsite generation with continued grid access. The Batch Zero process, governed by framework rules approved June 1, 2026 and effective July 11, introduced two new co-location constructs alongside the established Private Use Network. The framework centers on a single metric for every large load: a maximum withdrawal limit specifying how much the site may draw from the grid independent of onsite generation.
Two sourcing models have emerged. Net-Metering Arrangements allow already-operating generators to co-locate with new loads and net that consumption against onsite output behind a single meter, with only the residual settling against the grid. Because the generator was operating before September 1, 2025, the arrangement undergoes a 120-day transmission security assessment, with regulators able to approve, deny, or condition the setup. The bulk of announced ERCOT co-location activity sits in this bucket, including AWS's 1,200MW arrangement at Vistra's Comanche Peak nuclear plant, CyrusOne's 400MW total at Calpine's Thad Hill plant, and Crusoe's roughly 1GW Goodnight Campus.
Bring Your Own Generation structures allow loads to build or contract new co-located generation rather than wait on grid upgrades. ERCOT evaluates these on three parallel tracks: a Batch Study sets the grid withdrawal limit, Generation Interconnection sets the export limit, and Transmission Planning identifies required network upgrades. Sites draw up to their withdrawal limit from day one and ramp toward full capacity as each generation unit comes online. The distinction from NMA is vintage: BYOG generation is newly built and sits outside regulatory net-metering review.
Two new metering constructs formalize the arrangements. Withdrawal-Limited Private Use Networks allow loads bringing their own new generation to connect more megawatts than transmission alone could support, in exchange for an enforced cap on grid withdrawal. Provisional Controllable Load Resources represent dispatchable, flexible loads needing no onsite generation that connect at full requested size but can be dispatched down in real time during transmission constraints, with the cap relaxing toward a defined exit date as transmission is built out.
Equipment OEMs and IPPs Face Reshuffled Winners and Losers
The BTM surge creates clear divergence among equipment manufacturers and independent power producers. SemiAnalysis flagged GE Vernova, Siemens Energy, and Mitsubishi Heavy Industries as key losers of the trend in a report published a month earlier. The explanation is portfolio positioning: all three remain highly exposed to the grid-connected buildout that's now structurally constrained. While their BTM exposure is also high, it's not as high as vendors like Bloom Energy, which SemiAnalysis first called out in December 2024 as the biggest beneficiary.
With BTM becoming more favorable for buyers and expected timelines centering on 2028, SemiAnalysis does not expect utility turbine orders for 2030-plus capacity to increase. The firm sees 2026 as a potential peak for turbine orders for the big three OEMs, with most buyers focused on 2028 delivery flowing to Bloom, INNIO, Wärtsilä, Bergen, and similar vendors. The surge in contracted load drove massive orders, but SemiAnalysis now detects growing skepticism about utilities' ability to serve promised capacity on time. Combined with the financing challenges facing grid-connected projects, the setup points to peak turbine orders this year.
IPPs exposed to grid constraints and rising power prices face headwinds as demand for grid capacity eases on a relative basis while BTM surges. Constellation, Vistra, and Talen are negatively exposed to the trend. NRG Energy represents a potential exception through ERCOT-specific plays. The company appears well positioned to benefit from ERCOT's new BYOG and Withdrawal-Limited Private Use Network framework, given its available gas turbines to pair with co-located load.
NRG's 5.4GW Opportunity and Front-of-the-Meter Focus
On its fourth-quarter fiscal 2025 call, NRG management pointed to a contracted large-load opportunity implying roughly $2.5 billion of incremental EBITDA, built on blocks exceeding 1GW under 10 to 20 year contracts with investment-grade counterparties, with first power potentially online by late 2029. CEO Larry Coben stated on the February 24, 2026 call that "we're looking at blocks in excess of a gigawatt. I think we're looking at contracts of minimum 10 and frequently 20 years, with investment-rated entities that can actually support the kind of credit required to make this happen." He added that first power "could be on by the end of late 2029, and then ratably, probably 1 GW a year, maybe more, for each year after that."
Against the backdrop of the recent 20-year Microsoft-Chevron agreement in West Texas for the roughly 2.67GW Project Kilby, SemiAnalysis sees no reason NRG could not land a comparable long-dated, hyperscaler-anchored gas deal. Coben indicated on the same call that "our focus in PJM, at least initially, will be the 1 GW of uprates. It's just faster and quicker to market, the demand is there for Texas."
Management nonetheless continues to frame front-of-the-meter generation as its primary near-term focus, which may be reinforced by recent ERCOT PUN and BYOG rulings. President and CEO Rob Gaudette stated on the May 6, 2026 first-quarter fiscal 2026 call that "our primary focus is front-of-the-meter generation, front-of-the-meter data center, because we believe that's the right thing for the market." He acknowledged looking at behind-the-meter solutions but emphasized that current conversations are front-of-the-meter and "progressing as well as they have been over the last 12 months."
SemiAnalysis views potential solutions outside the immediate BTM buildout, including load flexibility, interconnection queue and market incentive reforms, and renewed transmission buildout. The firm notes that if grid-connected datacenters could curtail a determined number of hours per year, tens of gigawatts could be unlocked. PJM characterizes this flexibility as workload shifting, onsite backup generation dispatch, and battery discharge. However, SemiAnalysis remains cautious in the near term given commercial and regulatory constraints slowing broad-based adoption. On transmission, the firm expects the 2030s will likely witness major buildout as the largest power offtakers become creditworthy enough to post parent guarantees and underwrite whole transmission projects, but for now building new bulk transmission is simply too slow for the pace of AI datacenter construction. In practice, only a handful of US transmission corridors could physically support load growth at this scale, and meeting NERC reliability requirements on the necessary timeline is itself a binding constraint.