Microsoft flags tokens per joule

Microsoft is starting to describe the AI race with a utility-style metric: output per unit of electricity. On its April 29 fiscal third-quarter earnings call, the company said Azure and other cloud services revenue grew 40%, or 39% in constant currency, and forecast Azure growth of 39% to 40% in constant currency for the next quarter. At the same time, outside reporting in May showed Microsoft executives talking more directly about “tokens per joule,” a measure of how much model output a system can produce for the energy it consumes. The shift comes as Microsoft continues to spend heavily on data centers and AI talent while demand for electricity becomes a more visible constraint on AI expansion. ### Why is Microsoft talking about “tokens per joule” now? DigiTimes reported on May 22 that a Microsoft executive had highlighted “tokens per joule” as a metric for judging whether AI systems are improving in ways that matter operationally. The phrase points to a simple question for cloud providers: how much usable model output can they deliver for each unit of energy. The company has not made that phrase a centerpiece of its formal investor materials, but the timing is notable because Microsoft is adding AI capacity while trying to serve fast-rising demand. (microsoft.com) Microsoft Research added to that discussion in an April paper on AI inference energy use. The paper said public estimates often miss how efficiently large-scale inference systems operate and estimated median energy use of 0.31 watt-hours per query under what it called realistic assumptions for optimized frontier-scale deployments. That research does not use the same investor framing as “tokens per joule,” but it shows Microsoft is trying to document efficiency gains as scrutiny of AI power use grows. (microsoft.com) ### How does that fit with Microsoft’s cloud growth numbers? Microsoft told investors on April 29 that Azure and other cloud services revenue increased 40%, up 39% in constant currency, in the quarter ended March 31. Chief Executive Satya Nadella said the company was focused on delivering cloud and AI infrastructure and solutions for what he called the “agentic computing era.” Finance Chief Amy Hood said on the earnings call that Azure growth in constant currency should be 39% to 40% in the following quarter. (microsoft.com) Those figures matter because Microsoft has said results were helped by bringing capacity online earlier in the quarter, which enabled more consumption across AI and non-AI services. That links the growth outlook directly to infrastructure buildout rather than only to software demand. ### Where does responsible AI fit into a story about electricity? CNBC reported on May 23 that Jenny Lay-Flurrie, who became head of Microsoft’s Trusted Technology Group in February, described responsible technology in two parts: how to “build it right” and how to “keep it that way.” Microsoft launched the group in early 2025 and consolidated responsible-tech initiatives under it, according to CNBC. (microsoft.com) That language sits alongside a period of heavy infrastructure spending. Microsoft’s public filings and earnings materials have repeatedly tied capital expenditure to cloud and AI capacity, and the company has continued to add technical talent around AI products and platforms. Lay-Flurrie’s comments did not focus on power markets, but they showed that Microsoft is trying to pair rapid deployment with governance language as the scale of its buildout increases. (cnbc.com) ### Is power supply becoming a separate bottleneck for AI data centers? Microsoft already has one high-profile answer on the supply side. Constellation Energy said in September 2024 that it would restart Three Mile Island Unit 1 and sell about 835 megawatts of power to Microsoft under a long-term agreement tied to AI-related data-center demand. Politico reported at the time that the contract ran for 20 years. (microsoft.com) That deal has become part of a wider debate over whether AI expansion is running into grid and generation limits. Commentary and industry coverage have increasingly pointed to nuclear power as one possible source of firm electricity for large data-center campuses, especially where intermittent renewables and transmission constraints complicate planning. Microsoft’s own recent emphasis on efficiency metrics suggests the company is not treating power as an abstract issue. (constellationr.com) ### What should readers watch next? July 28, 2026, is the estimated date listed by CNBC for Microsoft’s next earnings report. Investors will be looking for whether Azure growth remains near the 39% to 40% range, whether Microsoft says capacity came online as planned, and whether management gives more detail on capital spending, infrastructure utilization or efficiency measures tied to AI workloads. (cnbc.com) (constellationr.com)