Microsoft's Q3 Earnings Confirm the Numbers Behind Azure's AI Hangover — and Why the Market Is Still Buying the Thesis
There is a version of this story where Microsoft's Q3 earnings are simply good news: revenue of $82.89 billion and EPS of $4.27 both beat estimates, Azure grew 40% in constant currency, and the AI business crossed $37 billion annualized. Nadella called it. The Street bought it. That version is true, and it is also the boring part.
The more honest version involves the margins. Microsoft Cloud gross margin came in at 67.6% — the narrowest since 2022. Capital expenditures hit $31.9 billion in a single quarter, up 49% year-over-year. CFO Amy Hood then walked analysts through calendar 2026 capex guidance of $190 billion, up 61% from 2025. She also disclosed that $25 billion of that number is purely a function of component price inflation, driven by a global memory crunch. The market had modeled something closer to $154 billion. The delta is a supply-chain problem nobody controls.
None of this is news to anyone watching the hyperscaler build-out. The interesting question is what comes after the build-out. And that is where Microsoft's earnings tell a more nuanced story than the headline beats suggest.
The $37 Billion Number Is the Real Headline
Microsoft's annualized AI revenue run rate of $37 billion — up 123% year-over-year — deserves more attention than it is getting. A year ago, that number was roughly $16.6 billion annualized. At this compound trajectory, AI services on Azure alone are on a path to represent a material fraction of Microsoft's total revenue within two years. This is not a hobby. It is a business unit.
The earnings call gave a few details that illuminate what that business actually looks like in production. Nadella mentioned that Microsoft's Maya 200 AI accelerator is now live in Iowa and Arizona data centers, delivering over 30% better tokens-per-dollar versus the latest silicon in the fleet. Cobalt, the custom server CPU, is deployed in nearly half of all data center regions and running real workloads for Databricks, Siemens, and Snowflake. These are not lab experiments. They are production infrastructure carrying paying customer workloads.
That matters for a specific reason. The custom silicon story is Microsoft's best argument that its AI margin will improve over time, not just its AI revenue. If Maya continues closing the cost-per-token gap at the rate implied by "over 30% better," Azure's AI infrastructure gets structurally cheaper to operate as scale increases — the same dynamic that normalized cloud margins for AWS and Azure after their own infrastructure build-outs. The catch is that $190 billion in capex has to keep flowing to get there, and investors are currently being asked to fund that thesis on faith as much as evidence.
What the Margin Compression Actually Means for Builders
Nadella said on the call that Azure supply will remain constrained at least through 2026. That is not a platitude — it is a capacity planning signal for enterprise teams building AI-dependent applications on Azure. If you are running production AI workloads today, the realistic option right now is not "add more Azure capacity at will." It is "compete for a finite supply of AI compute alongside every other Azure customer." That has real implications for architecture decisions, fallback planning, and how aggressively you want to pursue multi-cloud or hybrid strategies.
The memory crunch Hood cited is worth understanding in a bit more detail. The $25 billion component price impact she disclosed is not about GPUs being expensive because demand is high — it is about memory pricing specifically compressing across the industry. HBM3, HBM3e, and the broader DDR5 market are all being squeezed by AI demand that is drawing capacity away from commodity memory production. That is a supply-chain headwind that hits every cloud provider simultaneously, which means it is not a Microsoft-specific problem. But Microsoft's guidance explicitly calling it out suggests the company's procurement team sees this as a 2026 issue, not a 2027 issue.
For teams building cost models for AI infrastructure, the takeaway is that memory-adjusted compute pricing is not going to normalize in the next couple of quarters. Budget accordingly, and do not assume cloud credits or pricing concessions will materialize because demand is high.
The Q4 Guidance Gap
Hood guided Q4 revenue of $86.7 to $87.8 billion. The midpoint — $87.25 billion — came in below the $87.53 billion consensus. Combined with the margin compression picture, this produced the reaction that matters more than the beat: Microsoft guiding operating margin down to 44% from 46.3%, below even the reduced StreetAccount consensus of 44.6%.
The market's read is that AI capex is running ahead of AI monetization on the income statement, even if the top-line growth is real. Microsoft stock is down 12% year-to-date. That is not panic. It is a specific verdict: investors believe in the AI thesis but are not yet convinced about the timing of when margin recovery arrives.
What changes that verdict is not more revenue beats. It is margin direction. If the $190 billion capex year produces visible capacity expansion and cost-per-token improvements in 2027, the narrative flips. If it produces another year of margin compression with vague "supply remains constrained" guidance, the stock stays under pressure. Nadella has a clear path to redemption here. He also has a very specific timeline to deliver it.
The OpenAI "Royalty-Free" Comment Worth Reading
One earnings call detail that flew past most coverage: Nadella said Microsoft now has "a frontier model, royalty-free, with all the IP rights that we will have access to all the way to '32, and we fully plan to exploit it."
This is the first time Microsoft has described its post-deal OpenAI access in quite those terms. "Royalty-free" and "non-exclusive" are not the same thing, and the distinction matters. Microsoft is saying it pays no ongoing licensing fee for its access — which is consistent with no revenue-share payments going forward. It is also saying the IP license through 2032 covers the full range of OpenAI capabilities. That is a meaningful statement of continued strategic value from the partnership even after exclusivity ended.
For Azure customers and developers building on Azure OpenAI Service, the message is straightforward: the API is not going away, and the contractual foundation for Microsoft's ability to offer it remains solid through the decade. What changed is that it is no longer the only game in town.
Sources: CNBC, The Motley Fool, GeekWire