Microsoft under pressure: OpenAI accounts for 45% of Azure’s backlog, and capacity is becoming the real bottleneck
One figure encapsulates the new power dynamics in the cloud: OpenAI is said to represent approximately 45% of Microsoft’s commercial backlog (RPO), which the company values at $625 billion, reflecting a 110% increase year-over-year.
This impressive performance appears formidable on paper, but it also lays bare a vulnerability: Azure’s growth increasingly relies on a partner client with a demand for computing that outweighs the infrastructure’s immediate capacity.
RPO at $625 billion: Demand is strong, but concentration raises concerns
The RPO (Remaining Performance Obligations) represents the remaining contractual commitments to be fulfilled, serving as an indicator of future revenue “in the queue.” Microsoft reports $625 billion in RPO, with 45% of that amount attributed to OpenAI.
This sends a dual message:
- positive: Azure secures massive multiyear commitments;
- risky: a disproportionate share of visibility is dependent on a single actor in a market where strategies can shift rapidly (models, inference costs, partnerships, sovereignty, regulation).
Reuters notes that even without OpenAI, the remaining backlog continues to grow, but the revelation of OpenAI’s weight has immediately raised concerns about dependency.
Capex up 66% to $37.5 billion: The real cost of AI lies in data centers
The second figure that shifts the narrative is investment: $37.5 billion in capex for the quarter, showing a 66% year-over-year increase—largely driven by the demands for GPUs/CPUs and data center capacity.
Despite a quarter exceeding expectations, the stock fell in after-hours trading (approximately -6% to -7%): the market isn’t penalizing demand, but rather the ROI/capital intensity equation.
Azure growth capped by capacity, not by customers
Azure is growing significantly (~39% for the quarter according to Reuters; around 38% in constant currency according to other readings), but the implicit message is stark: Microsoft isn’t delivering everything it could sell due to a lack of instantly available capacity.
This is where Amy Hood’s statement hit home: she suggests reading Azure’s forecasts as an “allocated capacity guide“—in other words, a guide reflecting what Microsoft can provide, not just what the market demands.
The cloud becomes a zero-sum game: Focus on first-party products
Beyond the narrative of “Azure = open platform,” Microsoft is increasingly explicit: the GPUs arriving must be prioritized between first-party products (Copilot, GitHub Copilot, Microsoft 365), R&D, and the capacity sold to Azure customers.
In this context, OpenAI is not just “a big client”: it is also a strategic engine that drives computing needs and influences the pace of investment—which inherently puts more pressure on the rest of the Azure ecosystem.
Nadella attempts to reframe: “Diversification,” not dependence
Satya Nadella emphasizes a more balanced interpretation: not to “over-optimize” the business around a single engine and to continue pushing multiple franchises (Microsoft 365, GitHub, Copilot, etc.).
In plain terms, Microsoft wants to avoid the narrative becoming: “Azure is OpenAI’s data center.” Yet, the 45% figure complicates this reframing: although the portfolio is diverse, the physical constraint (computing capacity) requires constant trade-offs.
This situation places Microsoft squarely in the path of three very real tensions:
- Concentration risk: A disruption in OpenAI’s trajectory (technical, commercial, regulatory) would shake part of Azure’s visibility.
- Perception risk: Investors accept the capex but demand a clear understanding of when AI will “actually pay off.”
- Product risk: If resources are directed internally and/or to OpenAI, some Azure clients may feel underserved “after.” And in the cloud, dissatisfaction translates quickly into tests of competing services.
The paradox is that Microsoft finds itself both a victim of and a beneficiary of this pressure: demand is colossal, but the promise of the cloud (infinite elasticity) is encountering industrial reality—the AI revolution is making infrastructure a scarce resource, almost geopolitical.


