Anthropic’s 2028 Electricity to Be 87% Fossil Fuel: An Analysis Reveals the Energy Truth of AI Infrastructure

A quantitative analysis using Anthropic as a case study estimates that by 2028, about 87% of the electricity for its contracted U.S. data centers will come

On November 16, 2026, the nonprofit research organization Better Data Center Project released a quantitative analysis report using Anthropic as a case study. Core conclusion: by 2028, about 87% of the electricity consumed by Anthropic's contracted data centers in the United States will come from natural gas, 11% from coal, and the share of renewable energy will be close to zero.

These figures were derived by a third-party organization through site-by-site tracing, based on Anthropic SEC filings, the Epoch AI compute resource tracker, and filings by Amazon and SpaceX partners.

Scale: One Company's Emissions Match a State's

Anthropic currently has 15 GW of contracted compute capacity, of which 5.9 to 9.6 GW has been assigned to specific sites. Under the report's median scenario, annual carbon emissions in 2028 will be about 32 million metric tons of CO2 equivalent, with a range of 17 million to 38 million tons.

California's total emissions from all power generation in 2024 were 31.1 million tons (excluding imported electricity), and the combined Scope 2 emissions of Meta, Microsoft, and Google in 2025 were about 32.9 million tons. Anthropic's median emissions in 2028 will exceed both benchmarks at the same time.

In terms of power supply scale, 9.6 GW can meet the electricity demand of about 7 million U.S. households, or cover 18% of California's peak electricity demand.

Why Fossil Fuels, Not Renewables

When a company adds new electricity load to the grid without signing new renewable energy procurement contracts, that additional load will be met by the grid's "marginal generating units." In most parts of the United States, this means natural gas or coal. The output of wind and solar is constrained by time and location and cannot be tied to a specific load at the contract level unless locked in through an explicit power purchase agreement (PPA).

As of November 2026, Anthropic has no publicly disclosed renewable energy procurement commitments, contracts, or investments in the United States. Its disclosed green energy partnerships are limited to Norway and Canada.

Anthropic spent $3.4 billion to pre-purchase 1.7 GW of heavy-duty gas turbine capacity from OnePWR Solutions.

BYOP: How Self-Built Power Supply Bypasses Regulation

Traditional data centers rely on the public grid for power, and electricity procurement is subject to energy regulators and utility regulations. Under the BYOP model, data centers bypass the grid and build or contract dedicated generation facilities directly on campus, so that this generation neither goes through utility approval nor enters the unified dispatch system.

Nationwide, 177 GW of BYOP gas-fired installed capacity is currently in the planning stage, enough to meet the electricity demand of 65% of U.S. households. Across all BYOP projects, renewable energy is expected to remain below 20% by 2032. In addition, about another 200 GW of new fossil fuel power plants on the grid side are being advanced, driven mainly by incremental data center demand.

Data from the International Energy Agency (IEA) and McKinsey shows that data centers contribute 50% to 75% of the increase in global electricity demand.

What Anthropic Says Itself

Anthropic's public statement is that the company "is working to expand its energy portfolio to include geothermal, wind, solar, and emerging technologies." This sentence describes a direction, not signed contracts.

Anthropic's language in its SEC prospectus acknowledges that clean energy alternatives "may not be available in the quantities and on the timelines required, or may be more expensive to obtain."

Training and Inference: Two Different Energy Consumption Logics

AI compute load has one key difference from other industrial electricity use in terms of energy characteristics: it is extremely non-dispatchable. Large-scale model training requires stable, high-density power output continuously for weeks or even months and cannot be interrupted because wind speed drops or clouds block the sun. The load curve of inference clusters is relatively smooth, but the scale is enormous and they run around the clock. For both scenarios, the core demand for electricity is stability and reliability, not greenness.

The intermittency of wind and solar needs to be offset by large-scale energy storage or interregional transmission, and the construction timelines for such infrastructure are not on the same timeline as the expansion speed of AI compute.

Natural gas can generate power immediately, be dispatched on demand, and has large unit capacity; under existing technical conditions, it is the only solution that can support several gigawatts of new load within two to three years.

Independent Judgment: This Is a Visible Structural Bill

Better Data Center Project's analytical methodology is transparent, its data sources are verifiable, and its conclusions are cross-corroborated in reports by multiple independent media outlets. The value of this report lies in the measurement framework it provides: translating AI compute contracts into actual power sources, and then into quantifiable carbon emissions.

When an AI company says "we are working hard to use green energy" while spending $3.4 billion in the same year to buy gas turbine capacity, both things are true, but their weight is completely different. Contracts are a future already locked in; intentions are possibilities yet to be realized.

The IEA forecasts that global data center electricity consumption will double by 2030, with three-quarters of the increase coming from AI. If marginal electricity demand continues to be met by fossil fuels, this growth is handing the carbon emissions bill to the future.

The energy transition needs green power contracts, energy storage deployment, and grid upgrades that expand in step with compute scale. Before these infrastructure elements are in place, behind every large-scale AI training run, there is a row of gas turbines running in parallel.