explainx.ai350K+ learn from us
TrendingAI News TodayExplainx for Business
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcpestacktoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR: what the study measures
  • Why Anthropic became the case study
  • Finding 1: no disclosed U.S. renewable procurement
  • Finding 2: fossil-heavy 2028 power share
  • Finding 3: emissions range and comparisons
  • Finding 4: carbon intensity above grid average
  • Finding 5: health damages from NOx and PM2.5
  • Colossus 2, Monarch, and the permitting gap
  • The $3.4 billion gas generator purchase
  • How the model works (without trusting headline percentages blindly)
  • What developers are saying on Hacker News
  • What Anthropic has said publicly
  • How to stress-test the 87% headline
  • Policy and investor surfaces to watch
  • What you can do this week
  • Related reading
← Back to blog

explainx / blog

Anthropic Data Center Power: 87% Gas and 11% Coal by 2028 in New Study

Anthropic, Data Centers, Climate, AI Infrastructure, Energy

Part of Anthropic and Claude

Better Data Center Project models Anthropic U.S. sites at 87% gas and 11% coal power in 2028, up to 32M tCO2e yearly, with BYOP turbines and no U.S. renewable contracts disclosed.

Nov 22, 2026·9 min read·Yash Thakker
add explainx.ai
go deep
Anthropic Data Center Power: 87% Gas and 11% Coal by 2028 in New Study

A November 16, 2026 case study from the Better Data Center Project estimates Anthropic contracted U.S. data centers could draw about 87 percent of electricity from gas and 11 percent from coal in 2028, with 17 to 38 million metric tons of CO2 equivalent per year from power use at the mid-case. The nonprofit ties the numbers to 15 gigawatts of contracted compute, on-site gas turbines, and marginal fossil plants on the grid — not to a company carbon report. Anthropic's responsible data center development page promises air permitting and emissions controls; the study asks how that lines up with BYOP gas, Colossus 2 temporary turbines, and no disclosed U.S. renewable contracts as of November 2026.

TL;DR: what the study measures

table · 2 cols
Input (per study)2028 modeled outcome
15 GW contracted compute; 5.9–9.6 GW tied to named sites87% gas / 11% coal electricity share
2.5 GW on-site behind-the-meter gas; 80% simple-cycleHigher lb CO2e/MWh than U.S. grid average
Marginal grid attribution via EPA AVERTShort-run fossil meets new load without new renewables
Utilization 60–80%, PUE 1.1–1.217–38 MtCO2e/yr; mid-case 32 MtCO2e

The primary write-up sits inside a broader BYOP gas rush report that counted 177 GW of planned U.S. BYOP gas and another 200 GW of new fossil grid plants driven largely by data-center demand.

Why Anthropic became the case study

Anthropic is scaling compute faster than almost any lab, and it leaves a paper trail other private AI shops lack. The authors pulled site lists from Epoch AI's tracker, Anthropic SEC filings and news posts, partner filings from Amazon and SpaceX, and reporting outlets including The Information and Semafor.

That matters because power contracts are not marketing slides. When a lab signs for hundreds of thousands of GPUs on SpaceX Colossus or AWS training clusters Anthropic names in filings, somebody still buys electrons. explainx.ai already covered Texas permit fights and Harvard's hyperscale fossil snapshot. This post narrows the lens to one frontier lab's attributed mix.

Finding 1: no disclosed U.S. renewable procurement

Renewable symbols beside a data center — contrast with undisclosed U.S. clean power contracts in the studyRenewable symbols beside a data center — contrast with undisclosed U.S. clean power contracts in the study

The study's first headline is contractual, not atmospheric. It says Anthropic publicly disclosed renewable power commitments or investments only in Norway and Canada as of November 2026, while quoting Anthropic's statement that it is "working to broaden our energy mix with geothermal, wind, solar, and emerging technologies."

For builders leasing capacity, that gap is a diligence question. A Norwegian wind deal does not change AVERT marginal emissions for a West Virginia reciprocating-engine campus unless linked by contract to that meter.

Finding 2: fossil-heavy 2028 power share

The 87% gas / 11% coal split comes from stacking:

  1. On-site gas at named projects (Monarch, Nexus, Colossus 2-style sites).
  2. New grid gas tied to indirectly contracted builds (Schahfer expansion for AWS New Carlisle, Entergy Mississippi plants for Madison and Ridgeland sites in the study table).
  3. Marginal fossil on existing grid connections — the study uses EPA AVERT to ask which plants historically meet incremental load when renewables are not contracted.

Mid-case math assumes 9.6 GW of identified compute online in 2028, 70% utilization, and 1.15 PUE. Scale check from the study: 9.6 GW could power on the order of 7 million homes or a large fraction of California peak demand.

Finding 3: emissions range and comparisons

Annual CO2 from power lands at 17–38 MtCO2e with 32 MtCO2e mid-case. The authors compare that to:

  • California power sector emissions (~31.1 MtCO2e in 2024 per EIA, excluding imports).
  • Meta, Microsoft, Google combined 2025 location-based Scope 2 before offsets (~5.9M, 12M, 15M MtCO2e respectively in their reports).
  • Entire 2024 U.S. data center sector (~105 MtCO2e) — Anthropic mid-case would be a 16–36% uplift on that baseline.

Treat comparisons as illustrative. Anthropic may book some emissions as Scope 3 when it rents GPUs; lessors may book Scope 2. The study flags both.

Finding 4: carbon intensity above grid average

Modeled 1,033–1,076 lb CO2e/MWh average intensity beats the cited U.S. grid ~767 lb/MWh by about 35%. Drivers include simple-cycle on-site gas (study says 80% of behind-the-meter capacity) and coal-heavy subregions where sites sit.

That connects to community opposition threads and Ohio municipal bans — voters hear local air permits, not global model FLOPs.

Finding 5: health damages from NOx and PM2.5

Using EPA benefit-per-ton estimates, the study attributes $2–3 billion in annual health damages from 13,000+ tpy NOx and 2,000 tpy PM2.5 tied to Anthropic-attributed plants. That is separate from climate CO2 and matters for SCR requirements Anthropic says it expects partners to meet.

Colossus 2, Monarch, and the permitting gap

Two examples carry compliance weight:

Colossus 2. Anthropic has used SpaceX capacity since May 2026, per the study. It says temporary unpermitted turbines still supply much of the power, often without selective catalytic reduction, while Anthropic could reach ~600 MW there by end of 2026.

Monarch Compute Campus (Mason, West Virginia). 460 MW contracted, talks for up to 920 MW, powered by reciprocating engines the study says emit more formaldehyde and hazardous air pollutants than conventional gas plants.

Anthropic's October responsible-development post requires partners to follow air permitting and use SCR. The study is essentially a gap analysis: which public filings show permits and controls versus unpermitted turbines or engine fleets?

For SpaceX financing context, see Colossus 2 GPU targets and Anthropic S-1 compute spend — dollars signed before every turbine has a permit.

The $3.4 billion gas generator purchase

The study cites reporting that Anthropic bought 1.7 GW of heavy-duty gas generator commitments for $3.4 billion. That is BYOP hardware, not a renewable certificate. It aligns with the parent report's claim that under 20% of energy in current BYOP proposals will come from renewables by 2032.

If you run inference economics, pair this with TSMC upfront cash terms — capex and opex pressure show up in chips and electrons, not just API list prices.

How the model works (without trusting headline percentages blindly)

Method in plain language:

  1. Map each data center to a power source: on-site gas, dedicated new grid gas, or generic grid hookup.
  2. Use permit emissions factors when available; otherwise comparable plants.
  3. Run AVERT for incremental grid load — short-run marginal plants, not annual average grid mix.
  4. Apply PUE 1.1–1.2 and 60–80% utilization bands.

The authors also model Anthropic's evolving SpaceX lease from 300 MW (May 2026 announcement) through S-1 scale-up (~$3.1B/month at full delivery per October filing) with ±25% GPU power uncertainty.

Builders should copy the structure, not only the 87% headline: ask lessors for meter-level fuel mix, permit IDs, and marginal vs average grid claims.

What developers are saying on Hacker News

A November 22 Hacker News thread on the study surfaced recurring builder reactions (paraphrased, not verified facts):

  • vrganj asked why societies allow a small set of actors to expand fossil generation for AI.
  • dv_dt argued new data centers should require 110% new renewable capacity.
  • dan_grec said cheap solar and batteries should make new gas plants uneconomic without heavy taxes.
  • Joeri contrasted Mistral's French low-carbon mix with U.S. hyperscaler siting.
  • xorcist asked whether load-following data centers could chase cheap power geographically.

None of those comments replace the study's dataset, but they show where practitioner anger concentrates: permitting, marginal grid math, and renewable additionality.

What Anthropic has said publicly

Anthropic's responsible data center development post (October 2026) stresses air permitting, emissions controls including SCR, and requiring the same of power plants serving its sites.

The Better Data Center Project does not quote a new Anthropic rebuttal in the November 16 article. Readers should watch Anthropic newsroom posts, SEC filings, and state air board dockets for responses — not aggregators.

Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

How to stress-test the 87% headline

Modeling studies are only as honest as their inputs and boundaries. Before you paste the 87% gas / 11% coal line into a deck, pressure-test four questions the authors partially answer in their methodology appendix:

  1. Which gigawatts are actually online in 2028? The study brackets 5.9–9.6 GW of named-site capacity against 15 GW total contracts. A slower ramp lowers absolute emissions even if the mix percentage stays similar.
  2. Marginal vs average grid. AVERT marginal attribution is defensible for near-term load but pessimistic if Anthropic later signs dedicated renewables at specific meters. The study's renewables gap is contractual, not meteorological.
  3. Scope boundaries. Emissions from chip manufacture, water, construction, and employee commutes sit outside this power-only tally. Compare with Harvard's facility-year lens when you need industry roll-ups.
  4. Counterfactual workload. The same GPUs in a lower-carbon region with a longer interconnect queue might delay training but change tons CO2 per FLOP. Carbon-aware scheduling remains immature in most MLOps stacks.

None of those caveats erase the BYOP gas turbine line items; they stop you from treating one nonprofit model as Anthropic's official footprint.

Nuclear and grid power feeding servers — context for Anthropic BYOP gas attributionNuclear and grid power feeding servers — context for Anthropic BYOP gas attribution

Image: grid and generation tie-ins matter when data centers bring their own gas — context for Anthropic BYOP attribution in the Better Data Center Project case study.

Policy and investor surfaces to watch

Expect three venues to cite this case study even if Anthropic never responds:

  • State air boards reviewing SCR-equipped vs unpermitted turbines (Colossus 2 narrative in the study).
  • SEC climate risk questions in 10-Q filings after IPO — investors already parse compute commitments in the S-1.
  • Municipal ballot measures echoing Ohio data center bans where NOx and noise dominate hearings more than API uptime.

If you sell AI services to enterprises with Scope 3 targets, ask hyperscalers for hourly meter data, not only annual REC retirements. The study's core punch is additionality: new AI load met by marginal fossil without matched new clean capacity.

What you can do this week

table · 2 cols
RoleAction
ML engineerWhen choosing cloud region, ask for marginal emissions data and contractual renewable linkage, not only PUE
Corporate sustainabilitySeparate global REC purchases from site-specific fuel for Anthropic-class vendors
Policy readerCompare BYOP gas rush totals with local ban votes
FounderModel power as a cap table line alongside GPU rent

Related reading

  • U.S. hyperscale fossil power — Harvard Nature Sustainability
  • Anthropic and SpaceX Colossus partnership
  • Ohio voters pass data center bans
  • Texas data center permit politics
  • Neocloud remote-access export rules
  • Fed financial stability report on AI and utilities

Sources: Better Data Center Project Anthropic case study (November 16, 2026); BYOP gas rush report; Anthropic responsible data centers.

Spotted something out of date? Let us know.
Daily briefFrom Yash · 350K+ students taught

Catch up on AI

One AI story explained in plain English, plus one thing to try. Every day, free. Unsubscribe anytime.

  • ✓ Free
  • ✓ One email a day, no spam
  • ✓ Unsubscribe anytime
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Nov 18, 2026

US Hyperscale Data Centers: Harvard Study Flags Fossil Power Share

Researchers at Harvard publishing in Nature Sustainability estimate that 403 US hyperscale data centers emitted roughly 25.7 million metric tons of CO2 between May 2024 and April 2025, with more than half of their electricity coming from fossil fuels. They warn that thousands of additional facilities could worsen emissions without siting and grid changes.

Nov 8, 2026

Anthropic OnePWR Gas Generators: $3.4B, 1.7 GW — What FT and Piper Sandler Confirm

On November 8, 2026, the Financial Times reported that Anthropic committed roughly $3.4 billion in late September for 1.7 gigawatts of heavy-duty natural-gas generators from San Antonio-based OnePWR Solutions. Piper Sandler’s deal page confirms a $3.4 billion sale of a power platform to a frontier AI lab, without naming Anthropic on that page.

Oct 31, 2026

Sierra Club: AI Data Centers Could Add 42% to US Power Emissions by 2030

A Sierra Club analysis, first reported by Bloomberg, says data-center demand could push US power-sector emissions to 2.1 billion metric tons by 2030, erasing about 15 years of cuts. The numbers are reported second-hand and describe a high-impact scenario, not a forecast.