87% Gas, 11% Coal: What's Really Behind the Anthropic Climate Math
Since Anthropic's record IPO, the figure “87 percent gas, 11 percent coal” has been making the rounds, often presented as the company's own number. In fact, it comes from an environmental group's model. We took the calculation apart and redid the math. It turns out the accounting method changes the result dramatically, and Google, Microsoft and Meta look cleaner mainly because of how they keep their books.

What this is about
Ever since Anthropic debuted on the Nasdaq on November 19, 2026, one number has been doing the rounds: in 2028, the electricity for the Claude maker's data centers will come 87 percent from gas and 11 percent from coal. On Hacker News, the link racked up 151 points and 190 comments. English-language blogs are repeating the figure, some as if Anthropic had published it itself.
It didn't. The number comes from an analysis by the Better Data Center Project published on November 16. According to a summary on dev.to, the nonprofit was founded by two former US Department of Energy officials. It campaigns for data centers to run without fossil fuels. Its figure is a model, not a measurement. We break the model down into its assumptions and compare it with how the competition does its accounting. Our thesis: how dirty an AI company looks on paper depends to a large extent on which accounting method you choose.
How the number comes about
The authors pieced together where Anthropic has leased compute, drawing on Epoch AI's tracker, Anthropic's announcements and IPO filings, and media reports. By October 2026, that added up to 15 GW. Only 5.9 to 9.6 GW of that can be tied to specific sites. The remaining 5.4 to 9.1 GW don't appear in the calculation at all. DCD also notes that Anthropic doesn't disclose its sites and that some attributions are speculative.
| Assumption | Value in the model |
|---|---|
| Capacity at known sites | 9.6 GW, all operating in 2028 |
| PUE (total power divided by power for the IT equipment, i.e. the overhead for cooling and infrastructure) | 1.15 (range 1.1–1.2) |
| Average utilization | 70% (range 60–80%) |
| On-site gas plants | approx. 2.5 GW, 80% of it simple-cycle turbines and engines |
| Grid power | marginal emissions based on the EPA's AVERT tool |
| Renewables in the US | zero |
The decisive step concerns grid power. For it, the authors use marginal emissions: the output of the power plant that ramps up when a new consumer comes online. They rely on AVERT, a tool from the US Environmental Protection Agency that uses historical operating data to estimate which plants in a region cover additional load. Those are almost always gas or coal plants, because wind and solar already deliver whenever they're available. Since Anthropic hasn't disclosed any contracts of its own for new wind or solar power in the US, the authors set renewables to zero. The only disclosed agreements are in Norway and Canada.
That yields 87 percent gas and 11 percent coal, and a midpoint of 32 million tonnes of CO₂ in 2028. That's more than all of California's power plants emitted in 2024 (31.1 million t). The range runs from 17 to 38 million t. The authors put the average CO₂ intensity at 1,033 to 1,076 pounds per megawatt-hour, about 35 percent above the US grid average of 767 pounds. On top of that come estimated health damages from nitrogen oxides and particulate matter of $2 to $3 billion a year.
Don't confuse this figure with a second one that DCD cites: 60 percent of the mapped power demand is directly tied to gas, meaning on-site plants or newly built gas plants on the grid. That 60 percent describes contractual links. The 87 percent describes a modeled power mix.
Running the numbers: how much the method shifts the result
We recalculated the midpoint. 9.6 GW × 1.15 × 0.7 gives an average load of 7.7 GW, which comes to about 67.7 terawatt-hours over a year. At 1,042 pounds (0.47 t) of CO₂ per megawatt-hour, you get 32 million t. So the figure is internally consistent.
The range tells a different story. The authors attribute it to uncertainty about cooling needs and utilization. But if you keep 9.6 GW and plug in the most favorable values (PUE 1.1, 60% utilization), you land at about 26 million t, not 17. By our calculation, you only reach the lower bound if you also use the smaller site total of 5.9 GW. So the range of 17 to 38 million t mostly reflects how much load is assumed.
The accounting method doesn't show up in the range at all. Yet its impact is huge:
| Method for the same 67.7 TWh | Emission factor | CO₂ in 2028 (our calculation) |
|---|---|---|
| Marginal emissions (AVERT, the study's mid case) | approx. 1,040 lb/MWh | 32 million t |
| Average per eGRID subregion (footnote in the study) | 839 lb/MWh | approx. 26 million t |
| US grid average | 767 lb/MWh | approx. 24 million t |
| Market-based with renewable energy certificates for the entire grid share | grid ≈ 0, on-site gas remains | roughly 8 million t and up |
The last row is our own back-of-the-envelope estimate. The roughly 2.5 GW of on-site gas accounts for about a quarter of the electricity. Using the EIA's average for US gas plants (0.96 pounds per kilowatt-hour), that comes to just under 8 million t. Simple-cycle turbines and engines emit more, so this value is likely too low. Either way, depending on the method, the result ranges from a single-digit million figure to 32 million t, without a single change to the power plants.
Whether marginal emissions are the right yardstick is disputed. Electricity Maps stopped offering such data in 2024. The company argues that the “marginal plant” is a theoretical simplification that can't be verified. For a single point in time on the PJM grid (parts of the US East Coast), four estimation sources ranged from about −1,000 to 600 grams per megawatt-hour. The study's authors counter that marginal values best capture the short-term impact of new load. Their model explicitly doesn't account for long-term changes to the grid.
What's actually documented
The assumptions are open to challenge. But the core of the story rests on facts:
- Monarch, West Virginia: Anthropic has booked 460 MW with operator Nscale, with talks underway for up to 920 MW. Power comes from gas engines right on site.
- Colossus 2, Memphis: Since May 2026, Anthropic has been using capacity at the SpaceX/xAI data center. According to the study, power there comes mostly from temporary, unpermitted turbines, most of them without SCR (exhaust treatment that removes nitrogen oxides). According to The Next Web, Reuters counted 59 turbines without federal permits in July. In March, Mississippi approved 41 permanent turbines totaling 1.2 GW.
- 1.7 GW of gas generators: According to the Financial Times, as cited by Fortune, Anthropic is taking over OnePWR Solutions' supply agreements with Caterpillar for $3.4 billion.
- New Carlisle, Indiana: This Amazon data center is at the heart of Project Rainier and works exclusively for Anthropic. According to Utility Dive, the utility NIPSCO is building two gas plants of 1.3 GW each for it, next to a coal plant.
An Anthropic spokesperson told DCD the company is adding its own generation so it doesn't take power away from households and businesses, and that much of it is gas. The company's infrastructure page puts it this way:
“Much of this new power is natural gas, which can be brought online quickly.”
On the same page, Anthropic promises that all power plants must use SCR technology and says it will publish emissions for new sites. Colossus 2 doesn't square with that promise.
What's in the IPO prospectus, and what's missing
In the S-1, Anthropic made mandatory disclosures on energy risks for the first time. The company writes: “Certain of our planned data centers rely on on-site natural gas generation.” According to Crypto Briefing, Anthropic bears fuel costs “without a cap” and expects competition for gas and pipeline capacity. A site could be “delayed or abandoned.” Depending on the version of the prospectus, commitments for compute and infrastructure add up to $895 to $901 billion.
According to Latitude Media, clean energy comes up in just one sentence: solar, wind, storage, nuclear and geothermal “may not be available in the quantities or on the timelines we require.” We used the full-text search of the SEC's EDGAR database. It doesn't find the terms “renewable” or “clean energy” in any of Anthropic's documents. “Emissions” and “greenhouse gas” appear only in a supplementary marketing document (FWP) dated November 12, not in the S-1 and not in the final prospectus. Anthropic has no sustainability report. According to Fortune, the company is working with the carbon accounting service Watershed on an inventory for California's SB 253 law. It could classify leased compute as Scope 3 (supply-chain emissions) and not report it until 2027.
Why the others look greener
Google, Microsoft and Meta put their market-based figures in the headlines. These subtract purchased clean energy, including through certificates generated by a wind farm somewhere else entirely. The location-based figure, by contrast, shows what the local grid actually emits. The difference is huge:
| Company (year) | Scope 2 location-based | Scope 2 market-based | Reduction |
|---|---|---|---|
| Google (2025) | 15.1 million t | 2.8 million t | −81% |
| Microsoft (FY2025) | 12.0 million t | 2.7 million t | −77% |
| Microsoft (FY2024) | 10.0 million t | 0.26 million t | −97% |
| Amazon (2025) | 21.8 million t | 3.7 million t | −83% |
| Meta (2024) | 6.0 million t | 1,358 t | > −99.9% |
At Microsoft, the market-based figure jumped tenfold in a single year. According to reports, that was mainly because the company stopped buying certain freely tradable certificates. Its actual power consumption didn't rise anywhere near that much. At the same time, these companies are also having new gas plants built. Microsoft has a 20-year deal with Chevron for a gas plant in West Texas, and for Meta's Hyperion campus in Louisiana, the utility Entergy is financing gas plants. For the Anthropic data center in New Carlisle, Amazon can claim “100 percent renewable” on an annual, global basis, while the study counts full emissions for the very same gas plants. That's why the GHG Protocol, the accounting standard, has been debating since 2025 whether clean energy should have to be matched hourly and on the same grid. So far, the proposals have found little support.
Who counts how
| Company | Report/disclosure | Accounting method | Documented gas projects |
|---|---|---|---|
| Anthropic | no report, S-1 without emissions data | none | Monarch, Colossus 2, 1.7 GW Caterpillar |
| OpenAI (Stargate) | no report, “coordinating with partners” for California | none | Crusoe turbines in Abilene |
| xAI/SpaceX | no report, S-1 mentions water, not emissions | none | Colossus 1/2, Southaven |
| annual report | market-based in the headline, both disclosed | grid load, bought Intersect Power | |
| Microsoft | annual report | both, market-based for totals | Chevron gas plant in Texas |
| Meta | annual report, 2030 net-zero target under review | both, 100% “matched” since 2020 | Hyperion (Entergy), Ohio |
The bigger picture
Our take: the “87 percent gas” figure isn't a measurement, but it isn't scaremongering either. It's a deliberately chosen perspective that asks what new load physically triggers. That perspective suits a company with no clean power contracts of its own in the US. That Anthropic uses gas plants and buys them itself is documented, including in its own S-1. The order of magnitude, however, hinges on assumptions: that all sites run at full tilt in 2028, that marginal emissions apply, and that 5 to 9 GW of unknown capacity is left out.
The comparison with Google, Microsoft and Meta is lopsided too, just in the other direction. Measured by the same yardstick, they'd look considerably dirtier as well. And with certificates, Anthropic could look a lot greener tomorrow. The study itself claims that 32 million t exceeds the location-based Scope 2 figures of Meta, Microsoft and Google combined. Google and Microsoft alone come to about 27 million t, and we couldn't find Meta's figure for 2025. So whether the claim holds is a close call and unproven. The biggest open question is whether Anthropic will keep its promise to publish emissions per site. Until then, models like this one fill the gap.
Sources
- Better Data Center Project: How AI is powered – a breakdown of BYOP and grid power betterdatacenterproject.com
- dev.to: Anthropic's Data Centers Could Out-Pollute California's Power Plants dev.to
- DCD: Non-profit claims 60% of Anthropic's mapped US data center power demand is tied to natural gas datacenterdynamics.com
- Latitude Media: Anthropic debuts in an uneasy energy week latitudemedia.com
- Fortune: Anthropic inks $3.4 billion deal for gas-powered data centers fortune.com
- Crypto Briefing: Anthropic flags gas supply risks cryptobriefing.com
- SEC EDGAR: Anthropic S-1/A-2 sec.gov
- SEC EDGAR Volltextsuche (Anthropic, „renewable“) efts.sec.gov
- SEC EDGAR Volltextsuche (Anthropic, „emissions“) efts.sec.gov
- Anthropic: Our approach to responsible data center development anthropic.com
- Fortune: AI giants are quiet on climate fortune.com
- The Next Web: xAI is running 59 unpermitted gas turbines thenextweb.com
- Utility Dive: NIPSCO/Amazon data centers Indiana utilitydive.com
- Electricity Maps: Does the marginal power plant really exist? electricitymaps.com
- EIA: CO2 emissions per kWh eia.gov
- Trellis: Meta contemplates net-zero timeline reset trellis.net
- Fortune: Microsoft carbon emissions 2025 fortune.com
- Microsoft 2026 Environmental Data Fact Sheet cdn-dynmedia-1.microsoft.com
- Alphabet FY2025 Environmental Indicators Assurance Letter sustainability.google
- Amazon 2025 GHG Assurance Scope 1 & 2 sustainability.aboutamazon.com
- GHG Protocol: Scope 2 Market-Based Method Update (Oktober 2026) ghgprotocol.org
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.