Anthropic Could Emit 32 Million Tonnes of CO2 in 2028, 87% of Power From Gas
Better Data Centre Project finds Anthropic's AI operations may rely on 87% gas and 11% coal by 2028, with BYOP gas capacity of 177 GW planned across the US.

By Clara Tan, Artificial Intelligence
7 min read
What changed
The Better Data Centre Project published a case study on 16 November 2026 breaking down how Anthropic's AI operations are powered. The research organisation estimates that by 2028, Anthropic's total electricity consumption could reach 14.5 terawatt-hours annually, producing 32 million metric tonnes of carbon dioxide, roughly the same as 7 million petrol-powered cars driven for a year.
The breakdown shows 87% of that power coming from fossil gas and 11% from coal, with only 2% from renewables. The figures come from the group's "Bring Your Own Power" (BYOP) analysis, which tracks data centres securing dedicated generation rather than relying on utility grids.
What the BYOP model means
"Bring Your Own Power" refers to data centre operators contracting or building their own primary generation, typically gas turbines, to guarantee supply and bypass utility interconnection queues. The Better Data Centre Project identified 177 GW of planned BYOP gas capacity across the United States, enough to power 65% of all US homes.
Anthropic is not alone. The report finds that across current BYOP proposals, less than 20% of energy generation will come from renewable sources by 2032. Gas dominates because it can be permitted and built faster than utility-scale solar or wind with storage, and because data centres need 24/7 firm power.
Worked example: a single large language model training run
According to the report's methodology, training a frontier model like Claude 3.5 Opus requires approximately 50 gigawatt-hours of electricity over several months. At the projected 2028 grid mix for Anthropic's contracted capacity:
Gas portion: 43.5 GWh → ~24,000 tonnes CO2
Coal portion: 5.5 GWh → ~5,000 tonnes CO2
Renewables portion: 1 GWh → near-zero emissions
Total: ~29,000 tonnes CO2 for one training run. The report notes this excludes embodied emissions from chip manufacturing and data centre construction.
Worked example: inference at scale
Anthropic serves billions of tokens daily. The case study models 2028 inference load at 9.5 TWh annually. At the same 87/11/2 split:
Gas: 8.3 TWh → 4.6 million tonnes CO2/year
Coal: 1.0 TWh → 0.9 million tonnes CO2/year
Renewables: 0.2 TWh
Combined with training and overhead, the 32 million tonne figure emerges.
Where the numbers come from
The Better Data Centre Project cross-references:
EpochAI's compute forecasts for frontier model scaling trajectories
Public utility filings and interconnection queue data for 177 GW of gas projects
Anthropic's disclosed partnerships, including a 20-year lease with TeraWulf for a Pennsylvania data centre campus backed by gas generation
EPA emissions factors for US gas (490 kg CO2/MWh) and coal (950 kg CO2/MWh) generation
The group publishes its calculations and a comparison with EpochAI's independent estimates in an open spreadsheet linked from the case study.
Limits and uncertainties
Anthropic has not publicly confirmed its 2028 power mix. The 87/11/2 split is the Better Data Centre Project's projection based on contracted and planned BYOP capacity, not a company disclosure.
The 32 million tonne figure assumes all planned gas capacity comes online on schedule and runs at projected capacity factors. Permitting delays, gas price spikes, or policy changes could alter the mix.
The analysis covers Scope 2 emissions (purchased electricity) only. Scope 3 (supply chain, employee travel, model deployment by customers) would be larger.
Renewables share could rise if utility-scale solar-plus-storage costs continue falling or if federal tax credits drive more clean BYOP projects.
Why it matters
AI's energy demand is no longer theoretical. The BYOP trend means new fossil infrastructure is being built specifically for data centres, not just drawn from the existing grid. This locks in emissions for 20-30 year plant lifetimes.
For context, 32 million tonnes is:
More than the total 2023 emissions of Ireland (31.4 Mt CO2)
Roughly 40% of Singapore's 2023 emissions (77 Mt CO2)
Equivalent to adding 7 million cars to US roads
The report argues that without a shift in BYOP economics, carbon pricing, faster renewable interconnection, or binding corporate clean-energy procurement, the AI build-out will cement a gas-heavy power system through the 2040s.
Background timeline
Date | Event |
|---|---|
Nov 2024 | Better Data Centre Project publishes "The AI Gas Rush" identifying 177 GW planned BYOP gas |
Mar 2025 | Anthropic announces 20-year TeraWulf data centre lease in Pennsylvania |
Sep 2025 | US DOE reports data centre load could triple by 2028 |
Nov 2026 | This case study: Anthropic-specific breakdown using BYOP data |
What to watch
Anthropic's next sustainability report, whether it discloses actual vs projected power mix
FERC Order 1920 implementation, new transmission planning rules that could accelerate renewable interconnection
State clean energy standards, California, New York and Washington have 2030/2035 targets that affect data centre siting
Gas turbine supply chain, GE Vernova and Siemens Energy report 3+ year backlogs for large frame units
EpochAI's 2027 compute forecast update, due Q1 2027, will revise training/inference energy projections
This story was first published by TechWhoKnows on 22 November 2026. Original: https://techwhoknows.com/ai/anthropic-could-emit-32-million-tonnes-of-co2-in-2028-87-of-power-from-gas
Sources
- 1Better Data Center Project case studybetterdatacenterproject.com
- 2Hacker News discussionnews.ycombinator.com