The thesis: own the stack, not the winning model
Analyst Adam Wood reiterated an Overweight rating and a $600 price target in the Oct. 7 report, titled "Battle of the AI Stack: Built for a Multi-Model World." The target sits about 13% above Microsoft's $529.76 close on Oct. 7.
The report's core idea is that cheaper open-weight models, once seen as a threat, could feed Microsoft's growth. Open-weight models can be downloaded and run by anyone, and the worry was that they would shrink what companies spend on AI and leave Microsoft with costly data centers it can't fill. Morgan Stanley's counter-argument, in Wood's words: Microsoft doesn't necessarily need to own the single winning model if it owns the infrastructure that runs the models, the enterprise context that makes them useful and the control plane through which they are consumed.
The mechanism has four layers:
- Azure supplies the compute.
- Fabric (data management and analytics) supplies the enterprise data.
- Foundry lets customers pick and build with different models.
- Copilot delivers the result inside the apps employees already use.
Every layer can bill separately, so Microsoft could be paid several times for one AI task. Wood calls this loop an "Enterprise AI Flywheel," with the resulting usage driving more Azure consumption.
That is a model of how the pieces could fit together. It does not show that every customer buys every layer.
What Microsoft has actually reported
The company-reported numbers are the solid part of the story. In the quarter ended June 30, 2026, Microsoft reported:
- Revenue of $90.0 billion, up 18%.
- Microsoft Cloud revenue of $59.3 billion, up 27%.
- Azure and other cloud services growth of 43%.
- Full-year revenue of $331.8 billion.
On the earnings call, executives also said:
- Microsoft 365 Copilot has passed 30 million paid seats, with net seat additions more than doubling from the previous quarter. That is a seat count, not a count of active users.
- Foundry and Fabric adoption figures were cited in the call, including a figure for customers using both products. These are different measures, so they shouldn't be conflated.
- Full-year cloud revenue passed $214 billion, with nearly 90% coming from customers outside frontier model companies. That matters for anyone worried the business leans too heavily on a few AI labs.
Microsoft's own model-choice pitch
The earnings call echoes the multi-model theme. Satya Nadella said the platform should keep the "harness" (memory and context) separate from the model, so any given model can be swapped out. Customers can use frontier models, low-cost models, or train their own.
Microsoft is also pushing its own models. It said MAI-Code-1-Flash on GitHub Copilot delivered higher code acceptance rates and 10% lower median token usage. In Excel, it said the model delivers quality comparable to GPT-5.6 on common tasks at significantly lower cost.
This helps explain the margin angle. CFO Amy Hood described model diversification as a margin-improvement opportunity, alongside silicon and efficiency work.
The open-weight and survey numbers need caveats
The Rolling Out article, citing Morgan Stanley, says:
- Open-weight models made up about 65% of token volume among OpenRouter's 50 most popular models, versus roughly 38% a year earlier.
- 63% of technology leaders in a McKinsey survey use open-weight models, often alongside proprietary ones.
- The cost of reaching frontier-level intelligence comparable to 2023 capabilities fell about 1,000-fold in two years, citing MIT and Boston University research.
- In Morgan Stanley's AlphaWise survey, 47% of CIOs named Microsoft their preferred vendor for custom AI apps, against 10% each for Amazon and Google.
- 88% of CIOs expect to use Microsoft 365 Copilot in the next year, up from 72%.
I could not locate the underlying Morgan Stanley report, the survey methodology, or the MIT/BU paper, so treat these as attributed claims.
The OpenRouter figure deserves particular care. OpenRouter ranks models by tokens processed through its own API, which measures one platform's traffic. It is not a count of enterprise deployments, users or spending. A survey of CIOs' intentions is likewise not the same as actual adoption.
Where the thesis gets tested: spending and margins
The big open question is cost. The Rolling Out article cites a roughly $175 billion capex figure for fiscal 2027. Microsoft's call gives that number a different framing. Management said a shift from finance leases to operating leases adjusted its calendar-year 2026 capex expectation to approximately $175 billion. Underlying investment expectations were unchanged. So that figure is a calendar-2026 number and not a confirmed FY27 budget.
Microsoft also said:
- Company gross margin was 67% for the quarter, down year over year, because of a sales mix shift toward Azure and continued AI infrastructure investment.
- Microsoft Cloud gross margin was 65%.
- It expects FY27 capex to grow year over year, with first-quarter capex over $50 billion.
- It expects FY27 operating margins to be down less than a point.
- It plans to extend the useful lives of datacenters and office buildings from 15 to 25 years.
Morgan Stanley's reported projection of gross margin falling to 65.7% from 67.9% is an analyst estimate, not company guidance.
Rival offerings and a wrinkle for Windows readers
Microsoft has competition in the model-routing layer: Amazon Bedrock, Google's Gemini Enterprise Agent Platform, and services such as Cloudflare and OpenRouter. Morgan Stanley's view that Microsoft's blend of cloud, data, security and workplace software gives it an edge is an assessment, not a measured outcome.
Another wrinkle is that the same earnings call painted a weaker picture of the consumer PC side. Microsoft expects Windows OEM and Devices revenue to decline in the high teens for fiscal 2027. It cited higher component costs raising device prices, a prior-year boost from Windows 10 end-of-support, and elevated inventory. Windows OEM and Devices revenue already fell 7% in the June quarter. The AI growth story is firmly a cloud and enterprise story.
Reading the analyst call
Morgan Stanley's $600 target is not new. Its base case has applied a 25x multiple to fiscal 2028 EPS of $24.06. The Oct. 7 note reiterates that target and adds the multi-model argument. Analyst price targets are opinions tied to assumptions, and the bank's bull-case scenario in earlier coverage was far higher.
For IT admins and architects, the practical takeaway is less about the stock and more about platform design. If you are planning AI workloads on Azure, the vendor's own pitch is model portability. Keep your data, context and governance layers independent of any single model, and revisit your model mix as costs fall.
The central question stays unanswered: will cheaper tokens produce enough extra usage to pay back a very large infrastructure build? The thesis says yes. Microsoft's strong cloud growth supports it so far, but its margins show what that growth costs.