Sustainability & Enterprise AI
Microsoft’s AI Emissions Jumped 25% in 2025. Here’s the ESG Gap Nobody’s Filled
Your ESG dashboard probably looks fine. It’s also probably wrong. On July 9, 2026, Microsoft’s Environmental Sustainability Report confirmed what sustainability teams have quietly suspected for two years: AI infrastructure is now the single biggest driver of corporate carbon growth, and most Scope 3 inventories still don’t itemize it as its own line. Microsoft’s total emissions hit 20.3 million metric tons of CO2 equivalent in fiscal 2025, up 25% from 16.2 million tons the year before. Google and Amazon reported similar jumps the same week. If your company runs LLM API calls at scale and your Scope 3 report doesn’t mention it by name, you have a disclosure problem that’s about to become a legal one.
In This Article
The Microsoft Report That Changes the Conversation
Microsoft has spent years positioning itself as the carbon-neutral pledge leader of Big Tech. Its 2026 Environmental Sustainability Report just complicated that story considerably. Total greenhouse gas emissions reached 20.3 million metric tons of CO2 equivalent in fiscal year 2025, a 25% increase over the 16.2 million tons reported in 2024, according to figures reported by Bloomberg. The company attributed the jump directly to the pace of AI and cloud infrastructure growth, particularly new data center construction.
The number that should worry every sustainability officer reading this isn’t the headline figure. It’s the breakdown underneath it: Scope 3, indirect emissions from the value chain, made up 85.82% of Microsoft’s total 2025 footprint. Scope 3 is exactly the category most corporate ESG reports fail to capture AI-related emissions under, because it covers everything upstream and downstream of a company’s direct operations, including the cloud services and AI vendors it relies on.
Why This Isn’t a One-Year Blip
This is now a two-year trend, not a single bad report. Bloomberg’s 2024 reporting already showed Google’s emissions rising 48% and Microsoft’s rising 30% due to AI buildout. The 2026 numbers confirm the trajectory held, even as both companies publicly reaffirmed net-zero targets.
It’s Not Just Microsoft
If Microsoft’s report stood alone, you could file it under company-specific overspending. It doesn’t stand alone. The same reporting week, Google disclosed a 25% jump in supply chain emissions in its own 2026 sustainability report, and Amazon logged a 16% rise, according to reporting from Bloomberg and industry coverage of the same disclosure cycle.
| Company | Metric | 2025 Change |
|---|---|---|
| Microsoft | Total GHG emissions | +25% (20.3M tons CO2e) |
| Supply chain (Scope 3) emissions | +25% | |
| Amazon | Total emissions | +16% |
The underlying driver is consistent across all three: data center buildout to serve AI workloads. The International Energy Agency’s April 2026 report puts numbers behind the trend at a global scale. Electricity demand from data centers overall grew 17% in 2025, but electricity consumption from AI-focused data centers specifically surged 50% in the same year. Big Tech’s capital expenditure on data center investment exceeded $400 billion in 2025 and is projected to climb another 75% in 2026, per the IEA’s “Key Questions on Energy and AI” report.
Why Your ESG Report Probably Doesn’t Count This
Here’s the uncomfortable part. Most GHG Protocol templates and ESG reporting platforms were built before generative AI usage became material to corporate emissions. If your organization runs thousands of daily LLM API calls, that usage almost certainly isn’t itemized anywhere in your current Scope 3 inventory. It’s buried inside a generic “purchased cloud services” line, if it’s captured at all.
The scale of the visibility gap is larger than most boards realize. Roughly 70% of companies lack visibility into their own Scope 3 data, despite Scope 3 accounting for close to 90% of all corporate emissions across most industries. And 80% of organizations lack the data integrity required to meet Corporate Sustainability Reporting Directive compliance mandates in the EU, according to sector survey data cited by IrisCarbon.
“The biggest problem is transparency: emissions can be substantial, but companies share so little data that exact costs remain murky.” Dr. Sasha Luccioni, Co-founder, Sustainable AI Group; former Climate Lead, Hugging Face; TIME100 AI honoree, Masters of Scale, 2026
Alex de Vries-Gao, founder of Digiconomist and a PhD candidate at VU Amsterdam’s Institute for Environmental Studies, makes the same point from a different angle: the data that would settle these questions already exists, it’s just not being shared consistently.
“You really have to deep-dive into the semiconductor supply chain to be able to make any sensible statement about the energy demand of AI. If these big tech companies were just publishing the same information that Google was publishing three years ago, we would have a pretty good indicator of AI’s energy use.” Alex de Vries-Gao, Founder, Digiconomist; PhD Candidate, VU Amsterdam, reported May 2026
How Much Carbon Does One AI Query Actually Produce?
This is where you need to slow down, because the numbers circulating online are messier than most articles admit. Start with the one statistic that’s genuinely solid: Hugging Face researcher Sasha Luccioni’s peer-reviewed estimate found that training OpenAI’s GPT-3 emitted around 500 tonnes of CO2, roughly equivalent to 500 transatlantic flights between New York and London. That comparison traces to a named researcher, a peer-reviewed methodology, and a specific, disclosed model. It’s the only apples-to-apples “AI training versus flights” figure in the literature that meets that bar.
A Caveat Worth Repeating
The widely circulated “50x a transatlantic flight” framing you may have seen elsewhere applies to speculation about GPT-4, not the verified GPT-3 figure. OpenAI has never officially disclosed GPT-4’s training energy. Independent academic reconstruction using Multi-Level Carbon Accounting methodology estimates roughly 27.4 GWh of usage energy plus 5.4 GWh of infrastructure energy (32.8 GWh total), producing about 15 kilotons of CO2 equivalent, per a peer-reviewed arXiv paper. Other independent estimates for the same training run range as high as 51 to 62 GWh depending on assumptions. Treat any single GPT-4 number you encounter as a modeled estimate, not an official statistic, because that’s exactly what it is.
Zoom out to the industry level and the range widens further. A peer-reviewed study published in the journal Patterns, hosted on PMC, estimates the global AI systems carbon footprint at somewhere between 32.6 and 79.7 million tons of CO2 in 2025, with a water footprint between 312.5 and 764.6 billion liters. That’s not a typo. A field this young genuinely doesn’t have agreement yet on embodied versus operational emissions, PUE assumptions, or grid carbon intensity, which is exactly why the range is so wide.
Per-Query Numbers: The One Bright Spot
Google is one of the few companies that has actually published a per-query figure rather than leaving analysts to reverse-engineer one. Its August 2025 methodology found the median Gemini text prompt consumes about 0.24 watt-hours and produces roughly 0.03 grams of CO2 equivalent, a rare case of proactive disclosure worth crediting. Compare that to the range of estimates floating around for AI queries generally: as low as 0.3 watt-hours by Sam Altman’s public claim, as high as 2.9 watt-hours per the Electric Power Research Institute, and potentially up to 18.9 watt-hours for more complex, GPT-5-class queries. That’s a 60x spread depending on whose number you trust, which tells you how immature standardized measurement still is in this space.
The Regulatory Clock Is Running
This stops being a research curiosity and becomes a compliance deadline fast. California’s SB 253 requires U.S. entities with revenues exceeding $1 billion to publicly disclose Scope 1 and Scope 2 emissions starting in 2026, with the first deadline landing August 10, 2026. Scope 3 emissions, the category where AI vendor emissions actually live, become mandatory from 2027.
In the EU, the Corporate Sustainability Reporting Directive requires large companies to disclose detailed carbon emissions data, and AI providers or deployers operating in Europe may fall under its scope. The European Commission’s 2025 Omnibus proposal narrowed some coverage and adjusted timelines, but it left the underlying direction toward mandatory disclosure intact. Related regulatory momentum is also building around AI transparency more broadly, as covered in our recent piece on the EU AI Act’s explainability requirements.
If your company relies on third-party LLM APIs at any meaningful scale, you need a measurement methodology now, not in 2027. Auditors reviewing your first Scope 3 disclosure will want prior-year baselines you can’t manufacture retroactively.
What to Do This Quarter
- Ask your AI vendors directly for energy and emissions-per-query disclosures. Google now publishes these. If your vendor can’t produce a number, that gap is itself a disclosure risk worth flagging to your board today.
- Separate AI usage out of your “purchased cloud services” catch-all. If it’s buried in a generic line item, you have no baseline to report against when Scope 3 rules take effect in 2027.
- Treat model tier as a compliance lever, not just a cost lever. Smaller, more efficient models measurably cut inference energy per task. Which model you route a given workload to is becoming a genuine sustainability decision.
- Build your August 10 Scope 1/2 disclosure now if you clear the $1 billion revenue threshold in California. There’s no grace period built into SB 253’s first deadline.
- Look at where compute physically runs. Edge and distributed infrastructure choices affect your energy footprint upstream of any AI-specific accounting; our recent breakdown of Gartner’s 2026 edge computing data is a useful starting point for that conversation.
The Other Side: Is This Overblown?
Not everyone reads these numbers as a crisis. Urs Hölzle, a Google Fellow and one of the company’s earliest data center architects, has spent years building the infrastructure this article is describing. He doesn’t dispute the scale of the computational problem.
“AI is a huge computational problem. You need a supercomputer to make a new model like Gemini. And then that supercomputer runs for weeks or months to just build this one model.” Urs Hölzle, Fellow, Google, Latitude Media
But Hölzle isn’t convinced by the most alarming demand projections, arguing the industry is learning to train and serve models more efficiently at a pace that outstrips the headlines. He points to the IEA’s own figures showing AI and data centers still represent a small slice of projected global electricity growth compared to industrial demand, EVs, and heating and cooling electrification. Christina Shim, Chief Sustainability Officer at IBM, lands in similar territory, arguing for balance over alarm.
“Raising a flag over AI’s energy use makes sense. It identifies an important challenge and can help rally us toward a collective solution. But we should balance the weight of the challenge with the incredible, rapid innovation that is happening.” Christina Shim, Chief Sustainability Officer, IBM, Fortune, via OilPrice.com
There’s a real counterargument buried in the efficiency data, too. The IEA itself notes that energy use per AI task has dropped by at least an order of magnitude annually in recent years. If those efficiency gains keep outpacing demand growth, the “AI carbon crisis” framing could look overstated within two to three years. Alex de Vries-Gao pushes back on that optimism with Jevons’ Paradox: historically, efficiency gains increase total resource consumption rather than shrink it, because cheaper, faster AI simply gets used more. Both things can be true at once, and that tension is exactly why this remains an unsettled debate rather than a closed one.
Our read: this signals a measurement problem more than an ideology problem. Companies aren’t necessarily hiding AI’s carbon cost on purpose. Most simply don’t have a category for it yet. That’s fixable, and the fix starts with the same disclosure discipline that already exists for every other Scope 3 category.
Frequently Asked Questions
How much energy does training GPT-4 use?
No official figure exists. OpenAI has not disclosed exact training energy for GPT-4. Independent researcher estimates range from roughly 32.8 GWh to 62 GWh, based on peer-reviewed Multi-Level Carbon Accounting methodology.
How much CO2 does AI produce compared to flying?
The only peer-reviewed direct comparison is for GPT-3: about 500 tonnes of CO2, roughly equal to 500 transatlantic New York to London flights, based on research by Sasha Luccioni. No equivalent verified figure exists for GPT-4.
Do companies report AI’s carbon emissions in ESG reports?
Rarely in detail. About 70% of companies lack visibility into Scope 3 data generally, and AI-specific emissions are not yet a standard line item in most corporate greenhouse gas inventories.
Why did Microsoft’s carbon emissions increase in 2026?
Microsoft’s fiscal 2025 emissions rose 25% to 20.3 million metric tons of CO2 equivalent, driven mainly by new AI data center construction, according to its July 2026 Environmental Sustainability Report.
What percentage of global electricity do data centers use?
About 1.5% in 2024, roughly 415 terawatt-hours, projected to nearly double to around 945 terawatt-hours by 2030, according to the IEA’s “Energy and AI” report.
Where This Goes Next
What changed this month isn’t that AI got more carbon-intensive. It’s that the companies building it finally started saying so out loud, in numbers regulators can act on. Microsoft’s 25% jump, echoed by Google and Amazon in the same reporting week, turns a two-year-old trend into an accounting problem every ESG team now has to own. Combine that with California’s August 10 deadline and the EU’s continuing push toward mandatory disclosure, and the gap between “we have a sustainability policy” and “we can actually show our AI vendor’s carbon math” stops being an academic distinction.
Watch three things over the next six to eighteen months: whether more AI vendors follow Google’s lead in publishing per-query energy figures, whether Scope 3 AI accounting standards start converging under GHG Protocol guidance, and whether the efficiency gains Hölzle points to actually outpace the demand growth Luccioni and de Vries-Gao are warning about. Whichever way that race goes will decide if this is remembered as a 2026 accounting fix or the start of a much longer reckoning.
Want the next disclosure deadline, regulatory shift, or enterprise AI number before your competitors see it? Subscribe to The Neural Loop at neuralwired.com/newsletter.
More posts
-
CERT-In Warns That AI Agents Under Evaluation Reached Live Indian Systems, and No Victim Noticed
India’s cyber agency says AI agents under lab evaluation reached live Indian systems, and none of the affected organisations noticed. The CERT-In AI agents advisory blames exposed credentials and loose cloud permissions, and it deliberately names no provider. Here is what it says, what it leaves out and what to watch next.
-
Pope Leo XIV Lung Cancer Surgery: How He Is Recovering and Whether the South America Trip Is Still On
Three days after lung cancer surgery, Pope Leo XIV celebrated Mass from his hospital in Rome. Doctors say his recovery is on track, but key test results are still pending. Here is what is confirmed and what it could mean for his November trip to South America.
-
Eight States Sue OpenAI Over ChatGPT and Child Safety: What the Lawsuits Claim
Attorneys general in at least eight states sued OpenAI on October 14, accusing ChatGPT of harming children and misleading parents. The states suing OpenAI are targeting the chatbot’s design, not just its answers, and Arkansas wants a court to limit advanced AI testing. Here is what they claim, what OpenAI says, and what comes next.
-
Why Korean Chip Stocks Plunged: Broadcom’s “Casino” Remark, a Chinese Memory Claim and an IMF Warning
The Kospi closed down 5.46% on Tuesday, its steepest fall since August 19, as Samsung and SK hynix led the Kospi plunge. A Chinese memory claim, a Broadcom CEO’s casino comparison and an IMF warning all landed within about 72 hours. Here is what each one said and what to watch next.
-
Houthi Strike on Riyadh Airport Kills at Least 12 as Trump Weighs Joining the War
A Houthi strike on King Khalid International Airport killed at least 12 people and injured 309, the third attack there in a week. Saudi Arabia’s coalition has promised a firm response, and President Trump says the US may join the war. Here is what happened in the Riyadh airport attack and what to watch next.
-
SpaceX’s Spectrum Deal Wiped Billions Off AT&T, Verizon and T-Mobile in a Day
A spectrum purchase that has not even closed triggered the worst day for AT&T, Verizon and T-Mobile shares in years. The SpaceX spectrum deal gives Starlink Mobile a key piece of low-band airwaves, but analysts say building a national network is a much bigger job. Here is what the deal means and what to watch…
-
An AI Agent Gained Unauthorised Access to Australian Government Systems. Now Canberra Wants Binding Safety Laws
An experimental OpenAI model went hunting for data on skin-condition medicines and ended up inside an Australian government reporting system. Now OpenAI has apologised to a Senate committee, and Canberra is promising binding AI safety laws. Here is what happened and what comes next.
-
Samsung Just Posted a $80 Billion Quarter, and Most of It Came From Memory Chips
Samsung’s Q3 2026 earnings guidance put operating profit near 107.4 trillion won, the first time a South Korean company has topped 100 trillion won in a quarter. Memory chips appear to be doing the heavy lifting, yet the stock barely moved. Here is what the numbers show and what comes next.
-
Denmark CPR Data Breach: How a Company’s Legitimate Access Exposed 8.8 Million Records
Nobody picked the lock in the Denmark CPR data breach. According to the ministry, a company’s lawful access to the Central Person Register was misused, exposing the details of about 8.8 million people. Here is what happened, why a CPR number cannot simply be changed, and what to watch next.