Every efficient US 2030 data-centre portfolio used gas—but the result is modelled
A peer-reviewed optimisation ranked energy portfolios for AI growth in the United States and China. Its results depend on demand, lead-time, technology and policy assumptions; no grid plan or data centre was observed.
The short version
- What happened
- The peer-reviewed framework fully enumerated feasible energy portfolios for modelled 2030 AI data-centre demand in the United States and China, then reported Pareto-efficient choices rather than one forecast.
- Who it affects
- Residents and electricity customers may bear the cost, reliability and pollution consequences of data-centre power choices. This study models those choices but measures none of those lived outcomes.
- What remains unproven
- Peer reviewed early access
Keep in mind
Peer reviewed early access
Two-country 2030 model
Full portfolio enumeration
Scenario assumptions
No observed data-centre deployment
No household or community outcomes
ADIA affiliations disclosed
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The model searched for feasible portfolios, not a single future
A peer-reviewed study led from Khalifa University and the Abu Dhabi Investment Authority modelled how the United States and China might supply electricity for expanding AI data centres in 2030. The framework treats energy planning as a multi-criteria portfolio problem. It enumerates combinations that satisfy assumed firm-capacity coverage, lead-time and technology-readiness limits, then identifies Pareto-efficient portfolios: options where improving one objective would worsen another.
In the base case, every efficient US portfolio combined natural gas with firmed renewables, and none of the resources assumed deployable by 2030 reached an emissions intensity of 0.10 tonnes of carbon dioxide per megawatt-hour. China's efficient set ranged from coal-backed options to nuclear-plus-renewable combinations depending on how the model weighted cost, speed, resilience and environmental performance. The frontier contains several trade-off options rather than one objectively best answer. These are optimisation results under chosen constraints, not forecasts that utilities have adopted.[1]
Accounting and explainability assumptions moved the ranking
The framework distinguishes physical electricity delivery from certificate-backed contractual claims. When the authors counted physical delivery, a certificate-backed grid-purchase option in China fell from first place to below the top 1,800 portfolios. That result shows how an apparently technical accounting choice can dominate a sustainability ranking. It does not establish that one accounting rule is always correct or that the same ordering would hold under another country's grid and market structure. Nor does a rank alone quantify how large the practical difference is between neighbouring portfolios.
The paper also models the electricity needed to explain regulated AI decisions. Under a scenario with one million audited decisions a day, exact or amortised explanation methods add little, while perturbation-based methods can add tens to hundreds of gigawatt-hours a year. That range is an assumption-driven calculation, not a metered workload. Sensitivity analyses vary demand growth, technology cost, lead time, preference weights and explanation parameters, underscoring that the output is conditional.[1]
The distinction that matters
How a scenario becomes a ranked portfolio
Assumptions define the feasible set before the model can rank any option.
Inputs
Set 2030 demand, costs, lead times and emissions limits
A bounded scenario
Search
Enumerate feasible energy combinations
Pareto-efficient portfolios
Reality check
Compare with measured loads and local grid outcomes
Not performed in this study
Communities need measured loads and enforceable planning choices
People living near data centres, power plants and transmission projects may experience changes in electricity prices, reliability, water use, air pollution and land use. A transparent portfolio model can make trade-offs visible before infrastructure is locked in. Yet this study does not measure any household bill, outage, emission, water draw or community impact, and its two-country analysis cannot represent local grids within either large country.
Several authors are affiliated with the Abu Dhabi Investment Authority or ADIA Lab; the paper declares no specific external funding and no competing interests. Confidence would rise with public demand measurements, region-level grid constraints, lifecycle emissions, water and transmission costs, and comparisons against utility planning decisions. It would also help to test whether planners facing real permitting and community constraints choose similar portfolios. The work offers a reproducible decision framework and sensitivity tests—not evidence that gas is inevitable or that the modelled portfolios are socially acceptable.[1]