In 2020, Microsoft made one of the most ambitious climate commitments in corporate history: not just carbon neutrality, but carbon negative by 2030, including removing the company's historical emissions from the atmosphere. This week, reporting confirmed what many had suspected: Microsoft is quietly stepping back from those targets as its AI infrastructure buildout drives energy demand sharply higher. The retreat is significant not just for what it says about Microsoft, but for what it reveals about the energy accounting that most organisations deploying AI are not doing.
The mechanism of the retreat is telling. Microsoft announced it had "matched" 100% of its electricity consumption with renewable energy purchases. Critics noted that "matching" through renewable energy certificates is very different from actually running on renewable power. When a data centre in Virginia needs electricity at 2am, it draws from whatever the grid is supplying at that hour. A renewable certificate purchased to offset that usage changes the accounting, not the electrons. The company was simultaneously investing in carbon-removal projects and low-carbon cement for data centres while its actual emissions climbed.
The AI infrastructure build is the direct cause. Training large models and running inference at scale requires enormous, continuous computing power. GPU clusters do not have idle states. Data centres housing them consume electricity around the clock, generate heat that must be cooled (adding further energy overhead), and often operate in regions where renewable supply is intermittent. As AI workloads have grown, so has the gap between what Microsoft's climate commitments assumed and what the business actually requires.
What makes this broader than a single company's problem is a finding that runs through the AI industry at large: most organisations deploying AI tools have no visibility into the energy consumption of their workloads. Companies track AI spending. They track model performance. They track cloud costs. What they almost universally are not tracking is how much electricity is consumed per query, per workflow, or per automated process. The energy cost is treated as someone else's abstraction, buried in a cloud provider's infrastructure somewhere downstream.
This matters for a few reasons. The first is financial. As AI workloads scale, energy becomes an increasingly material cost driver, and organisations without visibility into that consumption cannot optimise it. A model that is twice as expensive to run as an alternative, at a scale of millions of queries per day, represents a meaningful budget difference that does not show up in any AI performance dashboard.
The second reason is that the efficiency case for AI is harder to make when the energy costs are invisible. AI is frequently sold to organisations as a productivity multiplier: do more with less. That case weakens considerably if the energy overhead of the AI tools is not netted against the gains. An employee who completes a task in half the time using an AI tool has been made more productive. If the AI query that helped them took three minutes of GPU compute to answer, the efficiency calculus looks different than it did in the pitch deck.
The third reason connects directly to climate commitments that organisations across the economy have made. The AI boom arrived after most of those commitments were set. The assumptions baked into 2025 and 2030 targets did not include the energy demands of training frontier models or running inference at enterprise scale. Microsoft is the most prominent example of a company discovering that the two trajectories, AI adoption and decarbonisation, are harder to reconcile than the press releases suggested. But it is almost certainly not the only one.
None of this means that AI's energy costs are necessarily prohibitive, or that the productivity gains do not outweigh them in many cases. The case for AI-assisted efficiency is real. But the energy reckoning is coming for the broader market, the same way the data readiness reckoning arrived for organisations that discovered their infrastructure was not ready to support the AI ambitions they had announced. Sustainable AI, which means AI whose energy costs are visible, measured, and accounted for, is starting to look like the next serious operational challenge for enterprises that have moved past the pilot stage.