Microsoft: Scaling Faster Than the Grid
The Sober Analyst Report 9/28/25
Each week, The Sober Analyst tracks the companies building the AI stack. These notes cut through claims and focus on what decides the quarter: capacity, supply, and performance. Every story ends with a Decision Box — the single metric, date, and change that matter most.
Microsoft: Scaling Faster Than the Grid
Microsoft is building more digital infrastructure than any other company. Its cloud empire already spans more than 300 datacenters worldwide, and analysts expect it to spend over $100 billion on new facilities and equipment through 2027. That pace already outstrips rivals: Microsoft’s quarterly capital expenditures now run higher than both Amazon and Google, making it not just a software giant, but the largest builder of datacenters on the planet (Reuters; Synergy Research; CNBC).
The centerpiece of its current build-out is in Mount Pleasant, Wisconsin. There, a new AI campus— Fairwater, is projected to draw as much power as a small country. Phase one alone will consume about 400 megawatts, with future phases reaching close to 900 megawatts. A new substation is being constructed to deliver at least 752 megawatts, and state regulators have advanced plans for 1.2 gigawatts of new gas generation to support the site (Microsoft blog; WISN; WPR; Canary Media).
Microsoft has promised to match every fossil kilowatt-hour with carbon-free power. This self-imposed pledge creates an additional constraint: the company isn’t just racing to find power, but specifically clean power. With 1.2 gigawatts of new gas generation approved for Wisconsin, Microsoft must now source equivalent carbon-free capacity elsewhere, adding complexity, cost, and delay to an already strained expansion (Microsoft sustainability blog; WPR; Canary Media).
This makes the grid itself a practical constraint. The company can build racks and cooling, but if interconnection timelines or generation additions lag, utilization will suffer. Fairwater’s nearly 1-gigawatt demand is one instance of a much larger story: industry analysts estimate the U.S. will need 55 additional gigawatts of new data center capacity by 2030. That is more than the entire state of New York consumes, and underscores why electricity is now the scarcest input to the AI boom (McKinsey; WPR).
Custom silicon is the second stress line. Microsoft’s Maia 200 accelerator has slipped to mass production in 2026, with reporting that it may fall short of NVIDIA Blackwell performance; Microsoft has not published quantitative performance or yield metrics. In the interim, Azure continues to market Blackwell-based virtual machines (The Information; Tom’s Hardware; Azure blog). Fairwater is being promoted as “ten times the performance of today’s fastest supercomputer,” but no public benchmarks or service SKUs substantiate that claim. Its architecture mirrors Azure’s new GB200 clusters, which are themselves gated by TSMC’s advanced packaging capacity and high-bandwidth memory supply.
Together, these pressures frame the near term: Azure’s AI and cloud services revenue grew 39 percent year-over-year in the latest quarter, yet the platform is scaling into bottlenecks in power and silicon while its anchor partner diversifies capacity. OpenAI has expanded contracts with CoreWeave and reportedly signed a multi-year compute deal with Oracle. The competitive landscape is shifting (Microsoft Q4 FY2025 earnings; Reuters; TechCrunch).
What to watch for: concrete updates on grid interconnections and new clean-power additions tied to the Wisconsin campus; official performance disclosures for Maia 200; and signs of workload allocation shifts by OpenAI or other anchor customers (Microsoft infrastructure updates; The Information; Reuters).
Why it matters: Microsoft’s infrastructure choices set the pace for AI availability across much of the market. If grid timelines slip or Maia underdelivers relative to NVIDIA, the constraints will be felt first by developers and enterprises building on Azure, even as topline cloud growth remains robust (WPR; Reuters; Microsoft earnings).
Decision Box
Tension: Can Microsoft deliver Fairwater’s near 1-gigawatt campus on time while matching fossil generation with clean power and securing GPUs in a tight supply chain.
Metric that decides: First public Fairwater benchmark or SKU performance against the “10× fastest supercomputer” claim.
What changed: Wisconsin regulators approved 1.2 GW gas build and substation plan. Azure GB200 clusters live. Still no Fairwater-specific metrics.
Next date: Next TOP500 or MLPerf release and Microsoft’s Q1 FY26 earnings guide.
End of public info: No disclosed GPU delivery schedules. No campus-specific benchmarks or carbon-free PPA volumes beyond initial solar.
Disclaimer: This analysis is based on public disclosures, regulatory filings, and reported statements available as of September 2025. Figures cited reflect company announcements and may include forward-looking projections.
I use a proprietary diagnostic engine I developed to evaluate business fundamentals and systemic risks across complex supply chains. Large language models assist with research support, source cross-checking, and drafting. The analysis, conclusions, and any errors are solely my own.
This is not investment advice. Corrections are welcome.