- Microsoft’s carbon emissions rose 25% to 20 million metric tons in fiscal year 2025, with AI datacenter expansion cited as the primary driver in its 2026 Annual Sustainability Report.
- Microsoft paused purchases of unbundled renewable energy certificates in favour of direct grid investment, a deliberate accounting shift that pushed reported emissions higher in the near term.
- Microsoft’s own report states sustainability solutions are not scaling fast enough to meet AI infrastructure demand, a pattern visible in Google’s 25% rise in supply chain emissions and Amazon’s 16% increase over the same period.
Microsoft’s carbon footprint grew 25% in a single fiscal year, reaching 20 million metric tons of CO2 equivalent, and the company is saying out loud what most hyperscalers will not: AI infrastructure is the reason. The admission, published in Microsoft‘s 2026 Annual Sustainability Report, puts a concrete number on a tension the industry has largely talked around.
AI Infrastructure Powers Emissions Surge
Microsoft is reportedly investing $3.3 billion in a new AI datacenter in Wisconsin alone, with similar buildouts underway globally. The scale of that physical expansion is what makes the emissions number hard to argue with: more racks, more power draw, more reported CO2.
Shifting Renewable Energy Procurement Strategy
Part of the emissions increase is self-inflicted, in the most deliberate sense. Microsoft paused purchases of unbundled renewable energy certificates (RECs), the credits that allowed companies to offset electricity-related emissions by buying into renewable generation elsewhere on the grid. RECs are a recognised instrument, but they do not guarantee that carbon-free power is actually reaching a specific facility. Microsoft is now prioritising direct investment in adding carbon-free electricity to the grids where its operations sit, which it expects to produce more durable decarbonisation over time.
The near-term cost of that decision is higher reported emissions, because the short-term offsets are gone. Microsoft President Brad Smith and Chief Sustainability Officer Melanie Nakagawa have said the shift, though painful short-term, is expected to accelerate genuinely new carbon-free electricity sources rather than recycling credits from existing generation.
Global Context of AI’s Energy Toll
Microsoft is not alone. Google reported a 25% increase in supply chain emissions in its own latest sustainability report, while Amazon recorded a 16% rise over the same period, both companies pointing to datacenter growth and AI demand as the driver.”
The physics are straightforward. Large language model training runs are among the most computationally intensive tasks ever industrialised. Each new model generation is larger than the last, and inference, running those models at scale for users, adds a continuous, 24-hour power draw that training alone does not. The workload does not peak and trough; it compounds. That is why hyperscalers are building custom silicon and liquid cooling into their datacenter designs rather than trying to solve the problem with software optimisation alone.
Navigating Sustainability Goals Amidst AI Growth
Smith and Nakagawa have not walked the 2030 carbon target back, but Microsoft’s sustainability report is candid: “AI infrastructure is driving demand for energy, water, land, and materials,” and “sustainability solutions are not scaling fast enough to meet demand.” Microsoft’s Climate Innovation Fund has committed $761 million toward technologies including direct air capture and sustainable aviation fuel.
Hardware Demands and Future Efficiency
The physical layer is where the problem is most concrete. AI datacenters are purpose-built for dense GPU arrays, specialised accelerators and high-bandwidth memory, hardware that runs hot and draws heavily from the grid. Embodied carbon adds to the operational footprint: manufacturing semiconductors, servers and racks carries its own emissions cost before a single inference job runs. As model size grows, both training and inference escalate in compute demand, which means electricity consumption and heat output rise in parallel.
Engineers are making real progress on efficiency: advanced liquid cooling, rack-level power management and purpose-built AI chips that deliver more FLOPS per watt than general-purpose hardware, but efficiency gains are running against a demand curve that is growing faster. A 10% improvement in power usage effectiveness matters less when the number of racks being deployed doubles. The path to Microsoft’s 2030 target requires hardware design, energy sourcing and operational discipline to move together simultaneously, and the current numbers suggest that coordination has not yet been reached. For more coverage of AI chips and infrastructure, visit our AI Hardware section.
