The US power grid is approaching a critical constraint point as AI data centers demand more electricity than utilities can deliver, according to a new analysis from SemiAnalysis. By 2028, more than 50% of new data centers may need to operate behind-the-meter—bypassing the grid entirely—to secure the power required for AI workloads.
Key Takeaways
- SemiAnalysis projects over 40GW of behind-the-meter data center capacity by 2028
- Grid expansion cannot keep pace with AI-driven power demand from hyperscalers and AI labs
- The shift reflects infrastructure constraints, not technology preferences
What Happened
On June 25, 2026, SemiAnalysis researchers Jeremie Eliahou Ontiveros, Sebastian Orejas, Ellie Holbrook, and Dylan Patel published an analysis warning that US grid constraints will drive a structural shift in how large-scale data centers are powered. The report argues that AI labs and hyperscalers are accelerating power consumption faster than grid operators can add transmission and generation capacity.
According to the analysis, the US grid currently serves most data center load, but that model is reaching a tipping point. The source states directly: "As the insatiable demand for power of AI Labs and hyperscalers keeps accelerating, the grid simply can't add capacity fast enough."
What Is Confirmed
The SemiAnalysis report projects that behind-the-meter data centers—facilities that generate their own power or connect directly to generation sources without using the public grid—could exceed 40 gigawatts of total capacity by 2028. The analysis further suggests that more than 50% of new data center builds per year will adopt this model by that date.
Behind-the-meter configurations allow data center operators to bypass grid interconnection queues, transmission constraints, and regulatory approval processes that can delay power access by years. The authors frame this shift as driven by necessity rather than preference—grid infrastructure cannot expand quickly enough to meet the scale of AI training and inference workloads planned by major technology companies.
The report does not specify which companies are leading behind-the-meter deployments, nor does it detail the generation technologies these facilities will use. The analysis focuses on capacity projections and grid limitations rather than individual projects or vendor strategies.
Why It Matters
This projection signals a fundamental change in US data center infrastructure strategy. For decades, large computing facilities relied on grid power with utility partnerships. If SemiAnalysis is correct, that model is ending not because of cost or efficiency advantages, but because data center grid constraints are forcing operators to self-provision power at scale.
The shift has implications for utilities, regulators, and energy developers. Grid operators lose visibility into load growth, making long-term planning harder. Behind-the-meter facilities also bypass transmission infrastructure entirely, which could accelerate the build-out of distributed generation but reduce revenue for transmission-dependent utilities.
For AI companies, this represents a new infrastructure dependency. Securing power becomes as critical as securing chips—and potentially more difficult, given permitting timelines for natural gas plants, nuclear reactors, or renewable arrays. The analysis suggests AI labs and hyperscalers are already factoring these constraints into expansion plans.
What Remains Unclear
The SemiAnalysis report does not disclose the methodology behind the 40GW projection or identify which data center operators are driving the shift. The analysis does not specify whether behind-the-meter capacity will come from natural gas generation, nuclear reactors, battery storage, or renewable sources. It also does not address permitting timelines, capital costs, or regulatory barriers that could slow self-generation deployments.
Details on regional differences are absent. Some US markets face tighter grid constraints than others, but the report does not break down capacity projections by geography or utility territory. The analysis also does not address how behind-the-meter facilities will handle grid interconnection for backup power or whether they will operate entirely islanded from the grid.
What To Watch Next
Monitor major hyperscalers and AI labs for announcements of power purchase agreements tied to dedicated generation assets rather than grid connections. Permitting filings for natural gas plants, small modular reactors, or large-scale renewable projects co-located with data centers will signal which generation pathways operators are choosing.
Utility earnings calls and regulatory filings should reveal whether grid operators are adjusting load forecasts to account for behind-the-meter capacity. If SemiAnalysis is correct, utilities will begin reporting slower-than-expected data center load growth on the grid even as total data center capacity continues expanding. That divergence would confirm the structural shift is underway.
