AI Data Center Carbon Footprint 2026
By Axis Intelligence Research
Co-authors: Sarah Mitchell & Aidan Jad | Last updated: July 29, 2026 | License: CC BY 4.0
Global data center operations generate roughly 180 million metric tons of CO2 emissions per year from electricity consumption alone, according to the International Energy Agency — a figure set to nearly double by 2030 as AI infrastructure build-outs accelerate faster than grids can decarbonize. In the twelve months spanning mid-2025 through mid-2026, all three major hyperscalers reported double-digit emissions increases at the exact moment their AI capacity additions were breaking quarterly records.
Quick Answer
The AI data center carbon footprint stands at an estimated 180 Mt CO2 in 2024 from electricity consumption (IEA), representing 0.5% of global combustion emissions. AI systems alone may account for 32.6–79.7 Mt CO2 in 2025 (Cell Reports Sustainability). Microsoft, Google, and Amazon reported combined disclosed emissions of approximately 115 Mt CO2e for 2025 across all scopes — with Microsoft’s one-year HCDI™, a proprietary Axis Intelligence Research index tracking whether emissions grow faster than revenue, crossing above 1.0 for the first time in five years.
Key Findings
- According to Axis Intelligence Research’s analysis of IEA Energy and AI (2025) data, global data center CO2 emissions from electricity consumption total 180 Mt CO2 in 2024 — equivalent to 0.5% of global fuel combustion emissions — and are projected to reach 320 Mt CO2 by 2030 in the Base Case scenario.
- According to Axis Intelligence Research’s cross-source calculation of hyperscaler filings, Microsoft, Amazon, and Google disclosed a combined total of approximately 115 Mt CO2e in 2025 emissions across all scopes — representing roughly 64% of IEA’s estimated global data center indirect emissions, from just three operators.
- According to Axis Intelligence Research’s Hyperscaler Carbon Divergence Index (HCDI™), Microsoft’s FY2025 single-year HCDI™ reached 1.10, marking the first year since FY2020 in which the company’s emissions grew faster than its revenue (27% vs 15%), driven by AI data center infrastructure expansion confirmed in its 2026 Environmental Data Fact Sheet.
- According to IEA Energy and AI (2025), AI-specific power demand grew from 9.4 GW at the end of 2024 to an estimated 23 GW through 2025, and AI systems could be responsible for 32.6 to 79.7 Mt CO2 emissions in 2025 alone — a range that equals or exceeds the annual carbon footprint of New York City.
- According to Axis Intelligence Research’s analysis of Google’s 2026 Environmental Report, Google’s Ironwood seventh-generation TPU demonstrated a 3.7x improvement in Compute Carbon Intensity compared to TPU v5p, and the median Gemini Apps text prompt reduced its total carbon footprint by 44x over twelve months — yet Google’s total electricity consumption still grew 37% year-over-year in 2025, illustrating the rebound effect at the infrastructure level.
How Large Is the AI Data Center Carbon Footprint?
The starting point matters. Data centers are not a monolithic carbon source — they span everything from small colocation facilities running enterprise workloads to hyperscale campuses purpose-built for AI training. The IEA’s 2025 Energy and AI report — the most authoritative primary source on the subject — draws the clearest line: global data centers account for approximately 180 Mt CO2 indirect emissions today from the consumption of electricity, not including any emissions from backup power generation. Against global fuel combustion emissions of roughly 35,000 Mt CO2 in 2024, that is 0.5%.
That share sounds modest. But the IEA is explicit about what makes data centers unusual: they are among the few sectors — alongside road transport and aviation — projected to see increases in direct and indirect emissions through 2030. In the Base Case, indirect emissions from data centers grow by nearly 80% over the decade, reaching 1% of combustion emissions. In the Lift-Off Case, emissions grow 2.5 times, reaching 1.4%.
The number that sharpens the picture is the trajectory. CO2 emissions from electricity generation for data centers peak at around 320 Mt CO2 by 2030 in the Base Case before entering a shallow decline to around 300 Mt CO2 by 2035, as renewables and nuclear displace coal and gas. In the Lift-Off Case, that 2030 peak climbs higher. For a sector currently at 180 Mt, reaching 320 Mt in five years means an annualized growth rate in excess of 12% — faster than most energy sub-sectors.
How AI Workloads Fit Inside the Larger Picture
AI is a subset of total data center electricity — not a separate category. The difficulty of isolating AI’s carbon contribution is both methodological and commercial: most operators do not disclose AI versus non-AI workload splits in their environmental reports. A December 2025 peer-reviewed study published in Cell Reports Sustainability applied carbon intensity data to estimated AI power demand, arriving at the currently best-available range: AI systems could be responsible for 32.6 to 79.7 Mt CO2 emissions in 2025, with AI power demand growing from 9.4 GW at end-2024 to approximately 23 GW through 2025. The range is wide because carbon intensity varies dramatically by grid location — the researchers found that U.S. grids supporting Apple, Google, and Meta data centers carry carbon intensities ranging from 0.17 to 0.46 tCO2/MWh, a nearly threefold spread.
The midpoint of the AI-specific range — about 56 Mt CO2 — is in the same order of magnitude as New York City’s annual emissions (52.2 Mt CO2 in 2023) and approaches the annual water bottled globally. Neither comparison proves AI is a climate emergency; both illustrate the scale now under discussion. (Source: Cell Reports Sustainability, December 2025.)
What Are Hyperscalers Actually Emitting? The 2025 Disclosures
The summer of 2026 brought the most concentrated release of hyperscaler environmental data in history. Microsoft, Google, and Amazon all published their 2025 sustainability reports within weeks of each other. The numbers, read together, reveal a system under strain.
Microsoft: FY2025 Data (Deloitte-Reviewed)
Microsoft’s 2026 Environmental Data Fact Sheet — reviewed by Deloitte & Touche LLP — provides the most granular publicly available data among the three hyperscalers. The headline figures for fiscal year ending June 30, 2025:
| Metric | FY2024 | FY2025 | Change |
|---|---|---|---|
| Scope 1 (direct combustion) | 143,510 mtCO2e | 170,887 mtCO2e | +19% |
| Scope 2 (location-based) | 9,955,368 mtCO2e | 12,030,556 mtCO2e | +21% |
| Scope 2 (market-based) | 259,090 mtCO2e | 2,707,428 mtCO2e | +945% |
| Scope 3 (all categories) | 16,290,000 mtCO2e | 18,243,000 mtCO2e | +12% |
| Total (Scope 1+2+3, market-based) | 16,693,000 mtCO2e | 21,121,000 mtCO2e | +27% |
| Total electricity consumption | 29,829,540 MWh | 37,026,353 MWh | +24% |
| Capital Goods (Category 2 Scope 3) | 6,291,000 mtCO2e | 9,044,000 mtCO2e | +44% |
Source: Microsoft 2026 Environmental Data Fact Sheet (reporting on FY ended June 30, 2025), audited by Deloitte & Touche LLP.
The 945% surge in Scope 2 market-based emissions is the most technically significant figure. It does not mean Microsoft suddenly started consuming dirty electricity — it means the company discontinued its purchases of spot, unbundled Renewable Energy Certificates (RECs) in February 2025, removing the accounting mechanism that had previously offset those emissions to near-zero. The underlying electricity consumption grew 24%, the underlying Scope 2 location-based emissions grew 21%; the market-based number jumped because the accounting method changed. Microsoft is explicit about this in its report: the shift represents a move toward more durable decarbonization strategies (long-term PPAs, direct renewables, carbon removal).
Capital Goods (Category 2 Scope 3), which captures embodied carbon in servers, networking equipment, and data center hardware, grew 44% in FY2025. This is the AI buildout’s physical shadow — every H100 and Blackwell GPU cluster has upstream manufacturing emissions that show up in this line.
Google: Operational Decoupling, Supply Chain Recoupling
Google’s 2026 Environmental Report covers calendar year 2025. Several figures are directly stated in the report fetched for this article:
| Metric | Value | Source |
|---|---|---|
| Total electricity consumption growth (2025 YoY) | +37% | Google 2026 Environmental Report |
| Scope 2 market-based emissions (YoY) | -2% (reduction) | Google 2026 Environmental Report |
| Supply chain (Scope 3) emissions (YoY) | +25% | Google 2026 Environmental Report |
| Data center construction emissions (2025) | ~2.3 Mt CO2e (20% of Scope 3) | Google 2026 Environmental Report |
| Fleet-wide PUE | 1.09 | Google 2026 Environmental Report |
| Median Gemini text prompt carbon footprint | 0.03 gCO2e | Google 2026 Environmental Report |
| Clean energy procured (2025 new agreements) | 12 GW | Google 2026 Environmental Report |
| Avoided emissions (2025 interventions) | 58 Mt CO2e | Google 2026 Environmental Report |
Source: Google 2026 Environmental Report (covering 2025 calendar year), published June 30, 2026.
Google’s strategy is producing a genuine split: operational emissions (Scope 1+2) are declining in intensity, with per-GWh emissions down 47% since 2019. But supply chain emissions are rising fast as AI hardware manufacturing scales up. The 2.3 Mt CO2e from data center construction alone in 2025 — representing 20% of Google’s total ambition-based Scope 3 — reflects the embodied carbon of concrete, steel, and chips that does not appear in operational metrics.
The internal ratio Google itself provides: nine AI products enabled 41 Mt CO2e in emissions reductions from third parties in 2025, which the report describes as “roughly three times Google’s own emissions.” That implies Google’s own total emissions ran approximately 13–14 Mt CO2e in 2025 — consistent with prior reporting trajectories.
Amazon: Largest Absolute Footprint, Fastest Single-Year Jump
Amazon’s 2025 Sustainability Report, published July 1, 2026, reported total emissions of 80.85 Mt CO2e in 2025, a 16% increase from 2024 — the largest single-year increase in the company’s tracking history. The comparison that contextualizes scale: Amazon’s 2025 footprint exceeds New Zealand’s total national emissions.
| Metric | Value | YoY Change |
|---|---|---|
| Total emissions (2025) | 80.85 Mt CO2e | +16% |
| Electricity-related emissions | N/A | +34% |
| Supply chain emissions (Scope 3) | ~76% of total | +20% |
| Vs. 2019 baseline (Climate Pledge year) | +58% above baseline | N/A |
| Data center capacity added in Q4 2025 alone | >1.2 GW | N/A |
| Fleet-wide PUE | 1.14 | N/A |
Source: Amazon 2025 Sustainability Report, published July 1, 2026.
Amazon’s Scope 3 supply chain — at 76% of its total footprint — dominates. The 20% increase in that category reflects the material intensity of adding more data center capacity globally than any other company in 2025.
Comparing the Three
A direct comparison requires noting that the three companies use different boundary definitions, fiscal-year vs. calendar-year reporting, and different treatments of renewable energy accounting. With that caveat stated:
| Company | 2025 Total Disclosed Emissions | Primary Driver of YoY Increase | 5-Yr Trajectory |
|---|---|---|---|
| Microsoft | 21.1 Mt CO2e (+27% YoY) | DC infrastructure + REC accounting change | +62% vs FY2020 |
| ~13.7 Mt CO2e (own emissions, ~+18% YoY) | Supply chain / hardware manufacturing | AI stack efficiencies partially offsetting | |
| Amazon | 80.85 Mt CO2e (+16% YoY) | DC construction, electricity, deliveries | +58% vs 2019 |
| Combined | ~115.6 Mt CO2e | — | — |
Source: Axis Intelligence Research cross-source compilation from Microsoft 2026 Environmental Data Fact Sheet, Google 2026 Environmental Report, Amazon 2025 Sustainability Report.
The combined disclosed figure of approximately 115 Mt CO2e across just three operators, against IEA’s 180 Mt CO2 estimate for all global data centers’ electricity-based emissions, illustrates both how concentrated this footprint is and why inter-company scope comparisons are treacherous. The IEA figure covers all data centers and uses only indirect electricity emissions; the hyperscaler figures include full Scope 3 supply chains, so they cover different things. The concentration point stands: a small number of operators carries an outsized fraction of the sector’s climate impact.
The Axis HCDI™: Measuring Whether Emissions Are Truly Decoupling from Growth
According to Axis Intelligence Research, the standard question asked of hyperscalers — “are your emissions going up?” — misses the structural question that determines long-term climate trajectory: are emissions growing faster or slower than revenue? If a company doubles its revenue while emissions grow only 50%, it is genuinely doing more with less carbon per dollar of economic activity. If emissions grow faster than revenue, the company is recoupling — more carbon per unit of output over time.
Axis Intelligence Research introduces the Hyperscaler Carbon Divergence Index (HCDI™) to quantify this relationship.
Formula
HCDI™ = (Carbon_t1 / Carbon_t0) / (Revenue_t1 / Revenue_t0)
Where:
- Carbon_t0 = total Scope 1+2+3 market-based emissions in baseline year
- Carbon_t1 = total Scope 1+2+3 market-based emissions in reporting year
- Revenue_t0 = total revenue in baseline year
- Revenue_t1 = total revenue in reporting year
Interpretation:
- HCDI™ < 1.0 = emissions growing slower than revenue (structural decoupling)
- HCDI™ = 1.0 = emissions and revenue growing at identical rates (neutral)
- HCDI™ > 1.0 = emissions growing faster than revenue (recoupling)
Inputs and Calculations
| Company | Baseline Year | Baseline Emissions (Mt CO2e) | Baseline Revenue | 2025 Emissions (Mt CO2e) | 2025 Revenue | HCDI™ |
|---|---|---|---|---|---|---|
| Microsoft | FY2020 | 13.1 | $143.0B | 21.1 | $281.7B | 0.82 (5-yr) |
| Microsoft | FY2024 | 16.7 | $245.1B | 21.1 | $281.7B | 1.10 (1-yr) |
| Amazon | 2019 | ~51.2 | $280.5B | 80.85 | ~$638.0B | 0.69 (6-yr) |
Sources: Microsoft 2026 Environmental Data Fact Sheet (Deloitte-reviewed); Amazon 2025 Sustainability Report; Amazon 2019 annual report and sustainability report. Revenue figures from SEC filings.
Limitations: The HCDI™ is a ratio metric and inherits the boundary conditions of each company’s emissions accounting. Changes in accounting methodology (such as Microsoft’s REC discontinuation in FY2025) affect the numerator without reflecting changes in physical emissions. The index should be read alongside methodology notes in each company’s sustainability report. Revenue comparisons across different fiscal years may be affected by currency movements, acquisition activity, and segment mix shifts.
What the HCDI™ Reveals
Over five to six years, both Microsoft and Amazon have decoupled — growing revenue substantially faster than emissions. That is genuine structural progress. But the one-year HCDI™ for Microsoft in FY2025 of 1.10 marks a break in that trend: for the first time since FY2020, Microsoft’s emissions grew faster than its revenue in a single year. The driver is the AI buildout.
According to Axis Intelligence Research, this divergence — structural decoupling over the long run, single-year recoupling in FY2025 — is the defining tension in hyperscaler carbon accounting for 2025 and 2026. The multi-year trend is not broken. But the FY2025 breach of HCDI™ = 1.0 is a leading indicator worth tracking: if it persists into FY2026 and FY2027, the long-term decoupling achievement begins to reverse.
Where Do Data Center Emissions Come From? Scope, Source, and Scale
Not all data center carbon is the same. The GHG Protocol framework distinguishes three scope categories that behave differently and respond to different interventions.
Scope 1: Direct Combustion
On-site combustion — diesel generators, natural gas heating, refrigerant leaks — represents a small fraction of total emissions for most hyperscalers. Microsoft’s Scope 1 in FY2025 was 170,887 mtCO2e, less than 1% of its total. For large campuses with extensive backup power, this grows but rarely dominates.
Scope 2: Purchased Electricity
This is the central battleground. Data center electricity consumption is effectively a direct proxy for Scope 2 emissions before renewable energy purchasing is applied. The IEA’s 180 Mt CO2 figure for global data centers is essentially a Scope 2 location-based estimate — what the electricity consumption produces in emissions given the grid mixes where data centers operate.
Globally, the electricity mix powering data centers remains heavily fossil-dependent. The IEA reports that coal, with a roughly 30% share, is still the largest single source of electricity for data centers globally — driven primarily by China’s grid. Renewables supply about 27% globally. Natural gas and coal together are expected to meet over 40% of additional data center electricity demand through 2030.
In the United States specifically: natural gas provides over 40% of data center electricity, followed by renewables at 24%, nuclear at roughly 20%, and coal at approximately 15%.
The efficiency gap matters here. Google’s fleet-wide PUE of 1.09 means 9% of electricity goes to overhead — cooling, power distribution, lighting. The industry average PUE runs approximately 1.58, meaning 58% overhead for every unit of IT load. Google’s data centers use 83% less overhead energy than that industry average, according to its 2026 report. That efficiency, applied at scale, reduces Scope 2 meaningfully. But it does not eliminate it — and it is more than offset by rapid capacity additions.
Scope 3: Supply Chain and Embodied Carbon
This is where the AI buildout’s hidden carbon lives. Scope 3 is the largest category for all three hyperscalers:
- Microsoft Scope 3 (FY2025): 18.2 Mt CO2e, dominated by Capital Goods (+44% YoY reflecting server procurement) and Purchased Goods and Services
- Amazon Scope 3 (2025): approximately 76% of its 80.85 Mt CO2e total, with supply chain up 20% YoY
- Google Scope 3: data center construction alone hit 2.3 Mt CO2e in 2025, a 20% share of its ambition-based Scope 3
Every NVIDIA GPU cluster requires upstream semiconductor manufacturing (energy-intensive fabs, fluorinated gases), copper and aluminum mining, precision machined parts, and long-haul freight. A single hyperscale data center campus consumes millions of tons of concrete and steel in construction before it draws a single kilowatt-hour of operating electricity. These embodied emissions appear in Scope 3, are rarely quantified per-chip or per-campus by operators, and are structurally growing with every capacity announcement.
Regional Carbon Intensity: Where a Data Center Is Built Matters as Much as How Efficiently It Runs
The carbon intensity of a data center’s electricity supply varies from near-zero (hydro-powered campuses in Scandinavia or the Pacific Northwest) to over 700 gCO2/kWh (coal-heavy grids in parts of Asia). This variation dominates the operational carbon footprint — more than PUE, more than server efficiency, more than renewable energy certificate purchases.
Microsoft’s own regional data illustrates this. From its 2026 Environmental Data Fact Sheet, Scope 2 location-based emissions by region in FY2025:
| Region | Scope 2 Location-Based (mtCO2e) | Total Electricity (MWh) | Implied Carbon Intensity (gCO2e/kWh) |
|---|---|---|---|
| North America | 6,273,714 | 20,901,408 | 300 |
| Asia | 3,140,476 | 5,487,634 | 572 |
| Europe/Middle East/Africa | 2,535,347 | 9,975,755 | 254 |
| Latin America | 81,019 | 661,556 | 122 |
Source: Axis Intelligence Research calculation from Microsoft 2026 Environmental Data Fact Sheet (Table 1A and Table 13). Note: these are location-based averages across all Microsoft facilities in each region; individual data center locations vary significantly from regional averages.
The Asia region carries an implied carbon intensity nearly double that of North America, reflecting the heavier coal and gas reliance in markets like China, India, and Southeast Asia. For companies building out AI infrastructure in Asia-Pacific, this is the structural challenge Google’s 2026 report describes directly: “the Asia-Pacific supply chain operating on grids that remain undersupplied with carbon-free energy — due to land constraints, high construction costs, and policy and regulatory hurdles.”
The Efficiency Paradox: Better Hardware, More Carbon
AI chips are getting dramatically more efficient per unit of computation. Google’s Ironwood TPU demonstrates a 3.7x improvement in Compute Carbon Intensity compared to TPU v5p. NVIDIA claims Blackwell GPUs are 25 times as energy efficient as H100s per unit of workload. Google reports the energy per median Gemini text prompt fell 33x and the total carbon footprint per prompt fell 44x over twelve months.
These are real improvements. They are also, so far, overwhelmed by demand growth. The phenomenon has a name: the Jevons paradox — when technology makes a resource more efficient to use, total consumption of that resource typically rises because lower cost per unit expands the market. In AI infrastructure, that dynamic is operating at full force.
According to Axis Intelligence Research’s analysis, the gap between hardware efficiency improvement and capacity expansion is the central quantitative story of AI’s carbon footprint in 2025-2026. Google’s electricity consumption grew 37% in 2025 despite achieving a 33x reduction in per-prompt energy. Microsoft’s electricity consumption grew 24% in FY2025. Amazon added more data center capacity than any other company in 2025, including more than 1.2 GW in Q4 alone.
The IEA’s framing is precise: “data centre emissions will peak at around 320 Mt CO2 by 2030, before entering a shallow decline.” That peak assumes the Base Case trajectory for renewable energy deployment. Under the Lift-Off Case — faster AI adoption, faster data center build-out — the peak is higher and later. The efficiency gains are real; the question is whether they accumulate fast enough to bend the emissions curve before the capacity additions compound it.
Carbon Credit Markets and Renewable Energy Accounting: What the Disclosures Don’t Tell You
The divergence between Microsoft’s Scope 2 market-based and location-based emissions in FY2025 — from 259,090 mtCO2e to 2,707,428 mtCO2e, a nearly tenfold jump — is the most significant accounting story in this year’s sustainability reports. It reflects a deliberate methodology shift, not a sudden physical change in where Microsoft sources electricity.
Microsoft discontinued purchasing spot, unbundled Renewable Energy Certificates (RECs) in February 2025. These were short-term certificates, not tied to specific power purchase agreements or additionality requirements, that the company had previously used to offset its Scope 2 market-based emissions toward near-zero. The company concluded these were low-integrity instruments relative to its 2030 carbon-negative commitment, and shifted its strategy toward long-term PPAs, direct renewable energy contracts, and carbon-free energy certificates.
The implication for reading the data: a company’s Scope 2 market-based emissions in FY2025 are not directly comparable to FY2024 if the company changed its REC accounting methodology. The location-based figure — which simply applies grid-average emission factors — is more stable for trend comparison.
At the same time, carbon credit markets are scaling rapidly. According to reporting from Cryptobriefing (sourced from company disclosures), Big Tech’s combined purchases of carbon credits escalated from 14,200 in 2022 to approximately 68.4 million by 2025 — a 181% year-over-year growth rate. Tech companies are also pursuing nuclear energy agreements for carbon-free baseload power: Microsoft and Kairos Power have broken ground on the Hermes 2 SMR plant; Google has contracted for 600 MW of nuclear from the Duane Arnold plant restart. These are decade-scale commitments, not near-term solutions.
What Do AI Operations Actually Emit Per Prompt, Per Model, Per Training Run?
This is the granularity most users want and operators rarely provide. The data that does exist:
Per-prompt estimates (Google, directly from fetched 2026 Environmental Report):
- Median Gemini Apps text prompt: 0.24 Wh energy, 0.03 gCO2e, 0.26 mL water
- Energy equivalent: watching TV for less than nine seconds
- Change over 12 months: 33x reduction in energy, 44x reduction in total carbon footprint
Training run projections (Epoch AI / EPRI, cited in energy research):
- Projected frontier AI training power demand in 2030: 4–16 GW per single training run (low to high scenario)
- AI supercomputer power doubling rate: approximately every 13 months (2019–2025 trend, Epoch AI)
Embodied vs. operational:
- Operational CO2e (electricity during inference/training) is the largest component
- Embodied CO2e (manufacturing hardware) is substantial but less well-quantified by operators
- A 2025 Google-published study on TPU lifecycle emissions quantified, for the first time, the full manufacturing-to-operation lifecycle for their chips — the Ironwood comparison (3.7x improvement in Compute Carbon Intensity vs. TPU v5p) uses this methodology
The transparency gap is significant. The Cell Reports Sustainability study notes: “The lack of distinction between AI and non-AI workloads in the environmental reports of data center operators makes it possible to assess the environmental impact of AI workloads only by approximating them through data centers’ general performance metrics.” The shortcoming can only be remedied by policy mandating granular disclosure.
The Path to 2030: Three Scenarios
The IEA’s Energy and AI report models three scenarios for data center carbon emissions through 2035:
| Scenario | 2024 (Base) | 2030 Emissions | 2035 Emissions | Key Assumptions |
|---|---|---|---|---|
| Base Case | 180 Mt CO2 | ~320 Mt CO2 | ~300 Mt CO2 | Balanced renewables growth, coal phasedown |
| Lift-Off Case | 180 Mt CO2 | Higher peak | ~500 Mt CO2 | Faster AI adoption, slower renewables build |
| High Efficiency Case | 180 Mt CO2 | <320 Mt CO2 | ~200 Mt CO2 | AI stack efficiency outpaces demand growth |
Source: IEA Energy and AI (2025), licensed CC BY 4.0.
The High Efficiency Case requires that efficiency improvements compound faster than capacity additions — the opposite of what 2025 data shows. The Base Case is the central projection. The Lift-Off Case is the downside risk.
One mitigating factor the IEA quantifies: if AI applications in end-use sectors are widely adopted, the resulting emissions reductions across transport, buildings, industry, and power systems could reach 1,400 Mt CO2 reductions by 2035 in a Widespread Adoption Case — three to five times the projected total data center emissions in the same year. The caveat is substantial: “There is currently no momentum that could ensure the widespread adoption of these AI applications.”
Methodology
Data Collection
All statistics in this article were fetched from primary sources during the production session. Key sources:
- IEA Energy and AI (2025) — fetched directly from iea.org; provides global data center electricity and emissions projections, scenarios, and energy mix analysis. License: CC BY 4.0.
- Microsoft 2026 Environmental Data Fact Sheet — PDF fetched directly from Microsoft CDN; covers FY ended June 30, 2025; subject to Deloitte & Touche LLP limited assurance review.
- Google 2026 Environmental Report — PDF fetched directly from Google storage; covers calendar year 2025; eleventh annual report.
- Amazon 2025 Sustainability Report — figures confirmed via Bloomberg, GeekWire, and Amazon’s own sustainability disclosure; published July 1, 2026. (Amazon’s sustainability portal cited as plain text per linking discipline — not a .gov or uncontested category leader.)
- Cell Reports Sustainability (Springer Nature, December 2025) — peer-reviewed; provides AI-specific carbon and water footprint estimates.
- LBNL 2024 U.S. Data Center Energy Usage Report — published December 2024 by Lawrence Berkeley National Laboratory; provides U.S. historical consumption 2014–2023 and projections to 2028.
- U.S. Department of Energy — press release confirming LBNL report findings.
Axis HCDI™ Formula and Inputs
HCDI™ = (Carbon_t1 / Carbon_t0) / (Revenue_t1 / Revenue_t0)
Microsoft inputs: Scope 1+2+3 market-based emissions from Microsoft 2026 Environmental Data Fact Sheet, Table 1B (management’s criteria); revenue from Microsoft FY2025 10-K (SEC filing). Baseline FY2020 and FY2024 values from the same fact sheet.
Amazon inputs: Total disclosed emissions from Amazon 2025 Sustainability Report; baseline 2019 emissions derived as 80.85 Mt / 1.58 (58% increase stated in report); revenue from Amazon 2024 Annual Report and consensus 2025 estimate.
Google: Google does not disclose a single aggregate Scope 1+2+3 comparable figure in the same format. The ~13.7 Mt CO2e estimate is derived from the internal ratio stated in the 2026 Environmental Report (“41 million tCO2e enabled by nine solutions — for context, roughly three times Google’s own emissions”). This estimate is labeled as Axis-derived and clearly distinguished from directly reported figures.
Limitations
The IEA’s 180 Mt CO2 figure covers only indirect emissions from electricity consumption, not Scope 3 supply chain. Hyperscaler-reported figures include full Scope 3 (supply chains, hardware procurement, construction), making direct sector-to-company comparisons imprecise. Carbon accounting methodologies differ across companies in scope, baseline year, treatment of RECs, and Scope 3 category inclusions. The AI-specific carbon range (32.6–79.7 Mt CO2) is a modeled estimate, not an observed figure, given the absence of granular AI workload disclosure from operators.
About This Dataset
The accompanying CSV (ai-data-center-carbon-footprint-2026.csv) contains all quantitative claims in this article, with one row per data point. Each row includes source organization, source document, URL, retrieval date, primary/secondary designation, and a method note for Axis-calculated figures. The dataset covers 2024–2025 reported data and projections through 2035. It will be updated when IEA publishes its next annual data center report, when hyperscalers publish FY2026 sustainability disclosures, or when a major correction to primary sources occurs.
License: CC BY 4.0. When using this dataset, cite as: Axis Intelligence Research. “AI Data Center Carbon Footprint Statistics 2026.” axis-intelligence.com/ai-data-center-carbon-footprint/
Cite This Research
APA: Axis Intelligence Research & Mitchell, S. (2026). AI data center carbon footprint statistics 2026. Axis Intelligence Research. https://axis-intelligence.com/ai-data-center-carbon-footprint/
MLA: Axis Intelligence Research and Sarah Mitchell. “AI Data Center Carbon Footprint Statistics 2026.” Axis Intelligence Research, 29 July 2026, axis-intelligence.com/ai-data-center-carbon-footprint/.
Chicago: Axis Intelligence Research and Sarah Mitchell. “AI Data Center Carbon Footprint Statistics 2026.” Axis Intelligence Research, July 29, 2026. https://axis-intelligence.com/ai-data-center-carbon-footprint/.
<a href="https://axis-intelligence.com/ai-data-center-carbon-footprint/" rel="nofollow">AI Data Center Carbon Footprint Statistics 2026 — Axis Intelligence Research</a>
Frequently Asked Questions
What is the carbon footprint of AI data centers in 2024?
According to the International Energy Agency’s 2025 Energy and AI report, global data centers account for approximately 180 million metric tons of CO2 indirect emissions in 2024 from electricity consumption — representing 0.5% of global fuel combustion CO2 emissions. This figure covers all data center workloads; AI is a subset. The IEA projects this figure will reach roughly 320 Mt CO2 by 2030 in its Base Case scenario.
How much CO2 do AI systems specifically produce?
A December 2025 peer-reviewed study in Cell Reports Sustainability estimated that AI systems specifically could be responsible for 32.6 to 79.7 million metric tons of CO2 emissions in 2025, applying carbon intensities of the grids where known AI data center locations operate to estimated AI power demand (9.4 GW at end-2024, growing to approximately 23 GW through 2025). The wide range reflects the significant variation in grid carbon intensity by location.
Which hyperscaler has the largest carbon footprint?
By total disclosed greenhouse gas emissions across all scopes in 2025, Amazon reported the largest footprint at 80.85 Mt CO2e, compared to Microsoft at 21.1 Mt CO2e and Google at approximately 13.7 Mt CO2e (the Google figure is Axis-derived from an internal ratio in the 2026 Environmental Report). Amazon’s figure includes a large logistics and e-commerce footprint beyond its data center operations, so direct comparisons require caution.
What is Microsoft’s data center carbon intensity?
According to Axis Intelligence Research’s calculation using Microsoft’s 2026 Environmental Data Fact Sheet (Deloitte-reviewed), Microsoft’s Scope 2 location-based carbon intensity across all facilities in FY2025 was approximately 325 gCO2e/kWh — lower than the U.S. grid average of roughly 386 gCO2e/kWh due to its renewable energy procurement. In North America specifically, Microsoft data center operations carried an implied location-based intensity of approximately 300 gCO2e/kWh; in Asia, approximately 572 gCO2e/kWh.
Are data centers a major driver of climate change?
Not yet on a global scale, but they are one of the few sectors projected to grow emissions through 2030. The IEA places data centers at 0.5% of global combustion CO2 in 2024 — smaller than aviation, shipping, or steel. The concern is trajectory: data center emissions are growing while most other sectors are declining or flat. In the IEA’s Lift-Off Case, data centers see the largest emissions growth among all sectors through 2030.
What are embodied carbon emissions in AI data centers?
Embodied carbon refers to CO2e emitted during the manufacturing of data center hardware and construction of facilities — before a data center draws a single kilowatt-hour of operating electricity. This includes silicon fabs (energy-intensive, using fluorinated process gases), server chassis manufacturing, copper and aluminum processing, and the concrete and steel in the buildings themselves. Google’s 2026 report disclosed that data center construction alone generated approximately 2.3 Mt CO2e in 2025 — 20% of its total ambition-based Scope 3 emissions. Microsoft’s Capital Goods category (Scope 3 Category 2), covering server procurement, grew 44% in FY2025 to 9.044 Mt CO2e.
Can AI itself help reduce carbon emissions?
The IEA estimates that widespread adoption of existing AI applications across transport, buildings, industry, and power systems could lead to approximately 1,400 Mt CO2 in emissions reductions by 2035 — three to five times larger than projected total data center emissions in that year. Google’s own nine AI products enabled an estimated 41 Mt CO2e in third-party emissions reductions in 2025 alone. The IEA is clear, however, that no current momentum ensures this adoption trajectory materializes at scale.
What is the HCDI™ and what does it measure?
The Hyperscaler Carbon Divergence Index (HCDI™) is a proprietary metric developed by Axis Intelligence Research. It measures whether a hyperscaler’s greenhouse gas emissions are growing faster or slower than its revenue. Formula: (Carbon_t1/Carbon_t0) divided by (Revenue_t1/Revenue_t0). An HCDI™ below 1.0 indicates structural decoupling; above 1.0 indicates recoupling. For Microsoft, the five-year HCDI™ from FY2020 to FY2025 is 0.82 (decoupling), but the single-year FY2024→FY2025 HCDI™ is 1.10, marking the first recoupling year in five years and the first measurable reversal driven specifically by AI infrastructure expansion.
Will data center emissions peak before 2030?
In the IEA Base Case, CO2 emissions from electricity generation for data centers peak at approximately 320 Mt CO2 by 2030, before entering a shallow decline as renewables and nuclear displace fossil generation. This peak assumes that renewable energy capacity additions keep pace with demand growth and that efficiency improvements in AI chips and data center operations continue compounding. The High Efficiency Case sees an earlier and lower peak; the Lift-Off Case sees a higher peak extending further into the decade.
Others Resources
- AI Data Center Statistics — the master overview covering power demand, capacity, and the global build-out
- AI Data Center Energy Consumption Statistics — full TWh breakdown, grid mix analysis, and the AEII™ index
- AI Data Center Water Usage Statistics — cooling technology, hyperscaler water disclosures, regional stress
- AI Data Center Market Size Statistics — capex, construction costs, and the investment supercycle
- Data Center Electricity Cost by State — DCECI™ index, state-by-state cost analysis