Microsoft Experiences Robust Growth in Cloud Services, While Copilot Performance Remains Uncertain
Microsoft Azure’s Growth and Investment in AI
Microsoft’s Azure cloud platform has experienced an impressive growth of 35% compared to the same period last year, according to financial results for the third quarter of fiscal year 2025. While advancements in artificial intelligence (AI) have played a significant role in this growth, much of it is attributed to traditional cloud use.
Financial Highlights
- Cloud Revenue: Microsoft Cloud revenue reached $42.4 billion, which is an increase of 22% when adjusted for constant currency rates.
- Overall Company Revenue: Total revenue for the company climbed to $70.1 billion, with earnings per share reported at $3.46 — both figures exceeding analysts’ expectations.
CFO Amy Hood noted that AI services contributed 16 percentage points to Azure’s overall growth. However, she emphasized that much of the surprising acceleration could be traced back to Microsoft’s core cloud business. CEO Satya Nadella pointed to the increased demand for migration and data services, specifically mentioning products like PostgreSQL, Cosmos DB, and Microsoft Fabric.
Custom AI Applications and Token Use
Over 70,000 companies have developed custom AI applications using Microsoft’s Foundry platform, which grants access to various models from notable providers, such as OpenAI and Meta.
- Token Processing: Microsoft processed over 100 trillion tokens during this quarter, with 50 trillion of those processed in the final month alone. This increase in token volume is partly attributed to the rising use of reasoning models that generate more internal text before arriving at a final output.
The adoption of Microsoft’s Copilot-branded tools has also surged. GitHub Copilot has over 15 million users, which is more than four times the reported number from last year. Microsoft reported a threefold increase in Microsoft 365 Copilot users compared to the previous year.
Infrastructure Adjustments
Despite advancing AI capabilities, Microsoft is adapting its infrastructure strategy. Executives have indicated that resource constraints are anticipated starting in June. Hood pointed out capacity constraints, while Nadella discussed the necessity for "power in specific places."
Reports indicate a shift in Microsoft’s approach toward infrastructure, with the company canceling letters of intent for leasing 2 gigawatts of capacity and halting plans for 1.5 gigawatts of in-house expansion for 2025 and 2026. Major data center projects have seen delays or pauses in construction and contracting.
It has been suggested that Microsoft’s exclusive partnership with OpenAI has ended, leading the startup to transition some of its infrastructure to other providers. Demand for Microsoft’s enterprise AI products has apparently not met expectations.
Focus on Efficiency and Software Optimization
In the third quarter, Microsoft invested $21.4 billion in cloud and AI infrastructure, though this figure was slightly below expectations. Hood attributed this variance to "normal fluctuations in the provision of rented data centers." Future spending is projected to increase, aiming for components that generate revenue in the short term.
During recent earnings calls, efficiency emerged as a major theme. Nadella shared that the cost per token has seen a reduction of more than half, while the performance per watt has improved by 30%. These advancements are largely the result of software optimizations rather than hardware enhancements. Improvements span various levels—from system software to model architecture. Nadella indicated that software could yield roughly a tenfold improvement.
Nadella also noted that the performance of AI models is doubling approximately every six months, thanks to multiple scaling laws, including advances in chip design, software enhancements, changes in architectural models, and improved efficiency from application servers. Additionally, Microsoft’s newer Phi 4 reasoning models align with this strategy, focusing on creating smaller, more efficient models instead of larger ones that require extensive resources.
