Q3 revenue guidance slightly missed estimates. PC revenue guidance was 8% below consensus and the other two segments were 2% better than expected. Guidance includes a 3-point currency tailwind that was 2 points larger than expected. That currency help was true across all three segments, with PBP getting a 4-point boost and the other two segments 2 points of help. PBP would have missed by about 2% without currency help, IC would have been in line and the 8% PC miss would have been closer to 10%. EBIT guidance missed estimates by 1%.

Other guidance notes:

CapEx, Capacity and Backlog Risks:

One of the big areas for concern from analysts was that 60%+ Y/Y CapEx growth greatly surpassed expectations while Azure was roughly in line with expectations. Shouldn’t faster CapEx growth mean faster revenue growth for the constrained Azure that can sell all of the capacity it can find? Maybe, but not necessarily. There are several other compute needs for Microsoft that it must balance. Copilot and Foundry need a ton of chips too and serving customers great experiences for those important products requires that a portion of CapEx doesn’t go to Azure. On the call, it sounded like there were some more pressing shortages popping up in the PBP segment that potentially caused a slight shift in prioritization away from Azure. 

On the backlog, the current portion rose 25% Y/Y, while the non-current portion rose by 156% Y/Y. Furthermore, 45% of its total backlog is now from OpenAI, although the non-OpenAI portion still grew by a strong 28% Y/Y. They were quick to highlight that and push back against analysts asking about concentration risks from a partner that remains deeply in cash burn mode. I think those concerns are valid, but the 28% disclosure was still great to hear.

And finally, leadership was asked how they’re confident in depreciating these assets over 6 years if the average contract duration is about 2.5 years. Per Amy Hood, confidence stems from the reality that most of these GPUs are “already contracted for most of their useful life.” That was nice to hear and supports elongated depreciation schedules from Microsoft and its fellow mega-caps.

AI Infrastructure Layer:

Microsoft is hard at work on optimizing tokens (unit of AI model input and output processing) per watt per dollar. Through various optimizations, they delivered a 50% token throughput boost to copilot workloads that run on OpenAI inference this quarter. It also successfully connected compute networks for two of its data center sites (scaling across) to enhance density and training throughput. Microsoft, Meta, Google and Amazon are all in a big race to lay as much data center capacity as they can right now. All of them view more compute per talented AI engineer as a large future edge that will set the stage for sustainably profitable compounding. Getting more efficient unlocks more asset creation from the same amount of CapEx dollars and a more competitive Microsoft in the age of AI. Whether it’s from chips like the new Maia 200 custom accelerator, updated models, the software-level or other areas of the value chain, it continues to make great strides in efficiency boosts and has a lot more progress left to enjoy.

Going back to the new Maia semiconductor debut for a moment, analysts praised the launch on the call for its large leap forward in inference capabilities. It comes with a 30% total cost of ownership improvement vs. its predecessor and will be a key part of inference and synthetic data generation for its Copilot, Foundry and Superintelligence teams. On the CPU side, its Cobalt 200 model boosts performance by 50% vs. the previous generation and will be an important part of cloud workloads. Agents readily tap into existing software apps and CPU-based infrastructure for things like container services and to access the simple tasks they need to string together to complete something much more complex. That drives usage of these assets and will likely make CPUs relevant for the foreseeable future.

AI Platform Layer:

Microsoft’s slew of tools for building, storing, testing, optimization and maintaining AI assets is enjoying great traction. In a developer world focused on model choice to use different LLMs for different workloads based on best fit, MSFT thinks it’s ahead. They continue to add more 3rd-party models across all of the industry leaders, as well as advancing their 1st-party models more purpose-built for smaller subsections of workflows. Another way to drive choice and flexibility is via enabling customization. With Foundry providing the tools and Fabric providing the data, Microsoft ensures its cloud suite thoughtfully does exactly this – all under a singular product suite called Unified iQ.

These two products (Foundry + Fabric) have also become noticeably popular for context engineering. This is essentially AI input optimization, or making sure the tools and data being plugged into these models are all tight, streamlined and as cheap as possible. Microsoft offers Fabric and Foundry products purpose-built for these specific use cases, which are gaining popularity and helping Fabric reach a $2B+ revenue run rate (+60% Y/Y). Microsoft’s Fabric provides the needed information and an Azure product called Foundry automates source curation and selection based on strict, highly malleable guardrails.

In terms of agent workflow governance, monitoring and orchestration, MSFT launched something called Agent 365 during the period. This gives them another tool to add to their massive product ecosystem. The offering does compete with ServiceNow’s AI Control Tower, but the two actually deeply partner with these two products specifically. In complementary and collaborative fashion, Microsoft handles asset identity (with its directories) and permission organization and ServiceNow monitors those permissions and access hygiene to ensure compliance.

AI App Layer:

Microsoft is building Copilot to actionably complete tasks across every single major product that it provides. It’s integrating it everywhere and finding strong traction in doing so. Copilot consumer app customers rose 3x Y/Y, while Microsoft 365 Copilot enjoyed 2x usage per user Y/Y, 160% Y/Y seat growth and 10x Y/Y DAU growth. Upgrades to Copilot algorithms yielded the “biggest Q/Q improvement in response quality to date,” and great momentum directly followed. GitHub Copilot delivered 77% Q/Q subscriber growth for its Pro Plus subscription and overall paid subscribers rose by 75% Y/Y to 4.7M. And now, with GitHub Copilot SDK, it’s giving developers the ability to embed Copilot’s agentic capabilities into their own apps, custom agents and services.

This was a solid quarter. The guidance was a bit weak when stripping out currency help, but the misses were quite modest and the raise to Y/Y EBIT margin guidance was solid. I don’t think Azure missing a growth number by a point comes close to qualifying as a red flag or a reason for panic for shareholders. And I don’t think CapEx growth leading revenue and profit growth this quarter is either. They have a ton of demand to fulfill and, again, are buying chips that are already contracted for most of their planned depreciation schedules.

While slight guidance disappointment did likely contribute, I think a lot of the continued share price weakness is simply a byproduct of how historically terrible enterprise software sentiment is today. Beyond that, hefty OpenAI reliance likely has shareholders a bit on edge. Microsoft is pretty dependent on the accuracy of OpenAI’s multi-year demand forecasts and its ability to fund these ambitious plans with willing capital markets. It’s a clear risk anytime nearly 50% of your backlog is with any single customer. It’s especially risky when that customer is years away from turning any kind of profit. I think Microsoft is a legendary company and Nadella is an excellent CEO. I just think the investment case is a bit too tied to OpenAI’s future for me to pick this over Google or Amazon for public cloud exposure.