Across 403 US hyperscale data centers, electricity consumption ranged from 68 to 99 TWh between May 2024 and April 2025, producing 37 to 54 million metric tons of CO2 at a carbon intensity 48% above the US grid average. That is the headline finding of arXiv:2606.05420, posted June 3, 2026: the first facility-level carbon audit to derive its figures from EPA eGRID plant-level generation data rather than corporate sustainability disclosures. The implication for anyone underwriting AI infrastructure is that siting decisions, not power purchase agreements, determine actual emissions.
The Numbers: What the Hyperscale Audit Actually Found
Guidi et al. surveyed 403 US hyperscale data centers over the 12 months ending April 2025. Across facility-load scenarios, those facilities consumed 68 to 99 TWh of electricity, with the central scenario corresponding to roughly 1.8% of total US electricity consumption. Associated CO2 emissions fell between 37 and 54 million metric tons, with about 54% of attributed generation coming from fossil-fuel sources per the study’s EPA eGRID-based methodology.
The electricity-weighted average carbon intensity: approximately 545 gCO2/kWh, against a contemporaneous US national grid average of 370 gCO2/kWh. Hyperscale facilities run 48% dirtier than the grid as a whole, a gap the authors trace directly to geographic clustering in regions with heavier fossil-fuel generation mixes.
Why Hyperscale Data Centers Run Dirtier Than the Grid
The 48% premium is not a server efficiency problem. It is a siting problem. Hyperscale data centers cluster in a handful of states where land, tax incentives, and fiber are cheap, and where the local grid’s generation mix tends to be fossil-heavy. An earlier audit by the same research group, covering 2,132 US data centers from September 2023 to August 2024, found an identical 48% gap, confirming the pattern across both studies and time periods per arXiv:2411.09786.
That broader study found data centers accounted for more than 4% of total US electricity consumption, generating over 105 million tons of CO2e (2.18% of US emissions in 2023).
The Growth Trajectory Behind the Carbon Numbers
US data center share of national electricity consumption exceeded 4% by 2024, with AI workloads as the primary demand driver. The growth rate is what makes the carbon-intensity figure consequential. If each additional TWh comes from regions where the grid is 48% dirtier than average, the marginal emissions from AI inference scale faster than the compute itself.
The PPA Gap: Contracted Clean Energy Versus Grid Reality
Hyperscalers routinely cite Power Purchase Agreements (PPAs) and Renewable Energy Certificates (RECs) to claim carbon-neutral operations. The Guidi et al. audit uses attributional accounting based on EPA eGRID plant-level generation data, which tracks what the local grid actually delivers to a facility, not what the facility has contracted on paper.
The gap between the two methods is the entire finding. A PPA for wind in Texas does not change the generation mix in Virginia’s PJM Interconnection territory. A hyperscaler can contract entirely renewable energy on paper and still draw physical power from a grid where 54% of generation is fossil-fueled. The audit’s 545 gCO2/kWh figure reflects grid physics. The corporate sustainability report’s net-zero claim reflects accounting convention.
This is not a novel observation in energy economics. But it is now backed by a methodology-transparent, facility-level dataset covering every US hyperscale data center, which gives regulators and investors a citable number to stress-test claims against.
Grid Constraints: The Power Wall Before the Compute Wall
The growth trajectory is colliding with grid capacity. Utilities requested $31 billion in rate hikes during 2025, more than twice the near-record from 2024, according to a PowerLines report covered by Fortune. A March 2026 Brookings analysis found US electricity costs had risen 42% since 2019, significantly outpacing general inflation (CPI rose 29% over the same period).
The practical consequence: inference scaling, the thesis that more compute applied at inference time yields better model outputs, runs into a power procurement problem before it runs into a chip supply problem. An operator cannot scale inference without another 100 MW interconnection, and the queue for those interconnections is measured in years.
What the Audit Means for AI Capex Narratives
The “clean compute” story hyperscalers present to investors relies on market-based accounting: buy enough PPAs and RECs to cover consumption, report net-zero. The Guidi et al. audit provides the attributional counterpoint. At 545 gCO2/kWh, the average US hyperscale data center emits substantially more carbon per kilowatt-hour than the US grid as a whole, because operators optimized for latency, tax incentives, and land cost, not for grid carbon intensity.
For investors underwriting AI capital expenditure, the audit supplies a specific, verifiable number to apply to any hyperscaler’s published power consumption. Multiply reported TWh by 545 gCO2/kWh and compare the result to the corporate net-zero claim. The difference is the PPA gap, quantified at the facility level.
For regulators, the audit gives a grounded baseline for emissions accounting that does not depend on corporate self-reporting. For opponents of new data center construction in fossil-heavy grid regions, it provides a citable, peer-reviewable figure: 48% above grid average, consistent across two independent samples spanning 2023 to 2025.
The audit does not prescribe where hyperscalers should build. It documents what the current buildout actually costs in carbon, which is a different number than what the press releases report.
Frequently Asked Questions
Does the 545 gCO2/kWh figure apply globally or only to US data centers?
The audit covers exclusively the 403 US hyperscale facilities. Global data centers consumed approximately 415 TWh in 2024, about 1.5% of worldwide demand, but carbon intensity varies widely by region. A facility drawing from Norway’s hydro-dominated grid would report a fraction of the US figure; one on a coal-heavy Southeast Asian grid would report substantially more.
How much more power does an AI rack require versus a traditional server rack?
AI-optimized racks draw 30 to over 100 kW each, compared with 7 to 10 kW for conventional enterprise racks. Global data center power demand is projected to reach roughly 945 TWh by 2030, up from 415 TWh in 2024, with this density gap as the primary growth driver.
Has any US utility actually refused new data center connections?
AEP Ohio has paused all new data center interconnections. NuScale Power, the only NRC-certified small modular reactor design in the US, continues to face cost overruns and delays, so the natural gas plants being built to serve near-term data center demand in fossil-heavy regions represent a multi-decade commitment, not a temporary bridge.
Which states account for the largest share of US data center power consumption?
Virginia alone consumed 52.21 TWh, more than 27% of all US data center electricity. Virginia, Texas, Oregon, and California together account for over 50% of the national total, concentrating the 48% carbon-intensity premium in a small number of grid-balancing areas.