Microsoft disclosed in its annual sustainability report on July 9, 2026 that its emissions for the latest fiscal year (FY25) reached 20 million metric tons of CO2 equivalent, up 25% from 16 million the prior year. The increase was driven by new data center construction for AI and cloud services, plus a previously announced pause in purchasing some renewable energy credits (RECs). In the same period Microsoft matched 100% of its global electricity use with clean energy and, for the first time, replenished more water than it withdrew (over 14 million cubic meters) — yet total emissions still rose, colliding with its pledge to remove more carbon than it emits by 2030. ASAP reads the numbers and their limits against Microsoft's primary report.

What Pushed Emissions to 20 Million Tons

Microsoft's FY25 emissions hit 20 million metric tons of CO2 equivalent, a 25% jump from 16 million the year before. The report, signed by Vice Chair and President Brad Smith and Chief Sustainability Officer Melanie Nakagawa, named two drivers: new data center construction and a pause in purchasing some renewable energy credits. Notably, Scope 2 emissions from purchased electricity rose to 13% of the total, up from roughly 2% a year earlier — the direct result of stepping back from REC purchases that had previously offset those emissions on paper.

The two executives wrote that "while AI infrastructure is driving demand for energy, water, land and materials, sustainability solutions are not scaling fast enough to meet demand." Microsoft has also signed a deal with Chevron for a natural-gas-fired power plant in West Texas to supply a new data center complex, accepting fossil-fuel electricity as part of securing its position in the AI age.

The Paradox: 100% Clean Power, Yet Emissions Up

The most striking feature of this report is that 100% clean-energy matching and a 25% emissions increase held true in the same year. The two figures look contradictory only because "matching" is an accounting method. Annual matching counts as 100% when a company buys or offsets, somewhere over the year, clean energy equal to what it consumed. But if the grid feeding a data center is running on coal or gas at the actual hour and location of consumption, the physical emissions still occur. A paper 100% and a real increase can coexist by design.

The jump in Scope 2 from 2% to 13% honestly exposes this paradox. When Microsoft reduced its REC purchases, the electricity emissions previously masked by credits reappeared in the statistics. In other words, the emissions did not suddenly materialize — the portion hidden by offset accounting became visible. The headline "100% clean electricity matched" reads as progress, but how real emissions actually moved beneath it is only visible under a 24/7 carbon-free energy (CFE) standard that accounts for power sources hour by hour, not annual matching.

The Real Weight Is in Building the Data Centers

The largest share of total emissions is not the fuel Microsoft burns directly (Scope 1) or its electricity (Scope 2), but Scope 3 across the entire supply chain. The report states Scope 3 remains the largest portion of the overall footprint. Read alongside the finding that data center construction drove the increase, this means the core of AI's carbon cost lies less in the electricity that runs servers and more in the steel, concrete, chips, and network gear needed to build the buildings themselves.

Metric	FY25 actual	Meaning
Total emissions	20M tons (prior year 16M)	AI data center expansion drove the 25% rise
Scope 2	13% of total (~2% prior year)	REC purchase pause made it visible
Scope 3	Largest share	Construction, materials, supply chain are the real bulk
Clean power matching	100% annual	Separate metric from hourly real emissions

This structure suggests AI's carbon cost arrives far upstream of the electricity bill — in the materials and construction phase of building data centers. A discussion that only counts the power of a single inference or training run sees barely half of AI's true carbon footprint.

Implications for Korean Data Centers and RE100 Firms

As Korean firms such as Naver, Kakao, Samsung, and SK aggressively expand AI data centers, Microsoft's numbers are not someone else's problem. The message for domestic RE100 participants and data center operators is twofold: annual renewable matching alone does not control real emissions, and the decisive battleground for emissions management is shifting from power procurement to construction and supply chains.

Korea's grid is regionally concentrated with a low renewable share, making hourly carbon-free power structurally hard to secure in the Seoul metropolitan area where data centers cluster. The fact that even Microsoft saw Scope 2 rise more than sixfold when it paused credit purchases reveals how fragile carbon accounting built on offset credits can be. Korean practitioners scaling AI infrastructure need to weigh not just the renewable share of power contracts but the embodied carbon of construction materials and the hourly source of electricity.

Can the 2030 Carbon-Negative Goal Survive? Limits and Open Questions

Microsoft pledged in 2020 to remove more carbon from the atmosphere than it emits by 2030, and only four years remain. With emissions rising 25% in a single year rather than falling toward the target, an honest question mark hangs over the pledge. The report shows progress on circular-economy metrics — first-ever water replenishment, 92% server component reuse for a second consecutive year — but the single most important number, total emissions, moved the wrong way.

Three open questions follow. First, as AI demand keeps growing, how will Scope 3 emissions from data center construction be reversed within four years? Second, can hourly carbon-free power targets actually be met while signing more fossil-fuel generation deals? Third, if the remaining gap is filled with carbon-removal credits, are those removals verified to actually take carbon out of the atmosphere? This report shows that sustainability in the AI age is not a matter of declarations but of how accounting methods, physical power sources, and construction supply chains are handled. ASAP will keep verifying Microsoft's next steps against measured numbers.

---

Sources: Microsoft On the Issues, "Responsibly building the AI future" (July 9, 2026; Brad Smith, Melanie Nakagawa) · Fortune (2026-07-09) · GeekWire (2026-07-09)