Goldman Sachs said Microsoft’s AI strategy is beginning to pay off, reiterating a Buy rating and keeping the stock on its Conviction Buy List in a Sept. 20, 2026 NDR note. The bank set a 12-month price target of $640. With Microsoft shares cited at $493.78 in the report, that implies 29.6% upside.
The note frames the call against investor concern that returns on AI capital spending may take too long to arrive. Goldman’s takeaway from Microsoft’s roadshow was that the company’s shift toward an enterprise AI platform is now starting to validate strategic choices made over the past three years. Those include early long-duration capex commitments, balancing investment between first-party applications and third-party customers, and allocating resources across frontier labs and enterprise demand.
Capex mix is shifting as Microsoft gains more flexibility
Gabriela Borges of Goldman wrote that Microsoft has more control over its supply chain and capacity ramp than it did a year ago. Long-duration capex now accounts for about 33% of total capex, down from roughly 50%. The report said part of that shift reflects Microsoft’s earlier spending on land, buildings and cold shell infrastructure.
Within short-duration capex, CPU has replaced GPU as the largest component. Microsoft is making decisions as late in the binding process as possible, according to the note, and GPU turn-up time has been cut by 50%. Microsoft also said it has never framed CPU or GPU supply as the binding constraint. The tighter limit, in its view, is the physical space available to insert chips.
Goldman said Microsoft’s stronger capex control shows up in several ways. With long-duration capex now a smaller share of the total, management has more room to adjust short-duration spending. Microsoft’s view is that if the constrained slice of capex is only a small part of the whole, it can still meet demand by reallocating the unconstrained portion. The company also said it has clear visibility into the three- to five-year expansion path of its most mature AI customers, which supports a reasonable baseline forecast for enterprise demand over the next three to five years.
Pricing and bookings point to enterprise strength
On pricing, Microsoft is using several levers to optimize long-term customer lifetime value. Goldman said renewal discounting has been more moderate than normal. New SKUs, including CPU-side offerings, are being introduced at higher price points, changing what the report described as more than a decade of price-down dynamics.
Capacity is being allocated weekly across first-party applications, first-party model development, the product roadmap and a broad customer base. Goldman highlighted a $51 billion quarter-over-quarter increase in fourth-quarter remaining performance obligations, or RPO, with all of that increase coming from enterprise customers outside frontier labs.
Goldman says AI unit economics look healthier than in the early cloud cycle
Goldman’s central argument is that Microsoft’s AI unit economics are in better shape than its cloud economics were at the same stage of the original cloud buildout. One reason is that Azure and first-party applications run on a unified technology stack. The report said separate stacks require more fine-grained matching and usually end with lower utilization.
Another point is customer timing. Microsoft began capturing AI-native customers in year one, while it was a later entrant during the earlier cloud cycle. Goldman said Microsoft may provide more detail on return on invested capital in the coming months. The bank’s logic is straightforward: if every queue in this cycle is ahead of the prior one, ROI should also be ahead unless rents are distributed unevenly, for example toward semiconductor companies and token providers.
On semiconductors, Microsoft said the chip layer is likely to become more diverse, much as CPU did. On token economics, the company said it expects to deliver enough value above third-party tokens and orchestrate across first-party and third-party tokens in a way that minimizes margin pressure over time. Microsoft also said there is no structural reason AI gross margin cannot move closer to cloud gross margin. Goldman added that cloud revenue should grow faster than capex, consistent with the cloud cycle as the market matures, though rising demand signals are pushing out the crossover point.
Silicon strategy centers on the lowest possible token cost
Microsoft said its goal is to deliver the lowest possible token cost. The company owns the intellectual property for Jalapeno, giving it a second option in the chip race alongside MAIA. The report said recent progress on MAIA 200 is comparable with Trainium benchmarks.
Goldman described Microsoft’s approach as a vertically integrated strategy that includes the model layer. By keeping a mix of silicon options, Microsoft can avoid overreliance on any single solution or architecture generation, especially while new chip technologies are still evolving. The note added that this also leaves room for customers to choose older-generation silicon when that makes economic sense.
Model strategy: learning from OpenAI while building independence
Microsoft reiterated that it has benefited from frontier learning at OpenAI and can build on that base while developing model architecture independence after 2032. The company said it now has a clear lineage across its model stack, reflecting what it has learned from OpenAI without becoming dependent on OpenAI.
The MAI family, Microsoft’s in-house model series, is focused on pairing frontier intellectual property with areas where the company believes it has outsized domain expertise: knowledge workers, coding and security. The note said Microsoft’s M365 system contains 17EB of data. Benchmark testing showed meaningful gains in cost-performance. Goldman cited Project Perception in security, used for continuous penetration testing, as a recent example.
Changes in frontier progress do not alter today’s demand signal
Microsoft told investors that any new technology has to create social and economic value, and that the company has a long history of supporting regulation that promotes safety and security. In Microsoft’s view, the bottleneck today is applying performance to enterprise use cases. Raw frontier performance is no longer the main constraint. The company said that even in an extreme thought experiment where frontier progress stopped, there would still be ample room for growth.
On model diversity, Microsoft said the ecosystem is evolving in real time into a heterogeneous set of models across performance and price points. Foundry currently hosts 11,000 models, and the company said its multi-model approach is intentional. Microsoft did not offer directional commentary on cross-model economics beyond repeating that it does not pay token fees for OpenAI or MAI and is comfortable supporting models that may require fees because it can cross-sell higher-margin platform services.
The report also mentioned Scout, an autonomous agent built on OpenClaw but designed for enterprise use. Microsoft said enabling lower-cost and open-source models can free up budgets for more usage or deeper transformation. Internally, the company has tested restricting frontier usage within engineering teams so that frontier capacity goes to engineers who clearly benefit from it rather than being spread uniformly across the organization.
Asked whether changes in frontier progress could affect capex, Microsoft said its capex decisions are a function of current demand signals and that shifts in frontier pace do not change those signals. The company can direct new capex toward frontier models, but the priority remains serving a broad customer base and a wide set of potential workloads. Goldman pointed to the shutdown of Stargate capacity as one example and to the fact that all fourth-quarter RPO growth came from non-frontier customers as another.
M365 enters what Microsoft called an acceleration year
Microsoft said FY27 is the first year in recent history with guided acceleration. In E5 and E7, incremental penetration is skewing toward lower-end segments, including frontline workers and small and medium-sized businesses. The E7 launch has gone well, according to the note, partly because it bundles the visibility and value of Agent 365.
Microsoft said AI adoption will show up in different forms across the installed base, making simple penetration disclosures less useful. Application pricing is also set to evolve. More value will be embedded inside seat-based SKUs, while consumption elements will be layered in over time as software prices increasingly map to units of labor and outcomes rather than only to seat-based add-ons. Over the long run, Microsoft said the dollar opportunity from consumption is larger than the dollar opportunity on a per-user basis.
On the relationship between software abstraction and frontier models, Microsoft said frontier competition remains fluid and enterprises do not want to be locked into a single model provider. At the same time, agents create more artifacts inside existing software ecosystems, which raises the value of platforms such as M365. Microsoft’s underlying view is that there is substantial value in the harness layer, including persistent security, context and cost optimization, and that model companies may not be able to replicate those capabilities in a model-agnostic way.
$640 target stays in place, with four downside risks listed
Goldman kept its 12-month price target at $640, based on an unchanged 28x multiple applied to Microsoft’s SNTM adjusted net income.
The bank listed four key downside risks:
- a longer-than-expected internal silicon ramp that could limit market share gains or gross margin expansion;
- project investment running above expectations;
- changes in key leadership;
- a more pronounced shift toward customized software that could weigh on the applications business.
Goldman’s conclusion was that Microsoft’s enterprise AI platform transition is validating strategic decisions made three years ago. If unit economics continue to outpace the cloud cycle and AI gross margins move closer to cloud margins, the company’s valuation premium would, in the bank’s view, have fundamental support.
The source article said the rating, price target, earnings forecasts and related judgments were derived from a Sept. 20, 2026 Goldman Sachs research report and reflect the views of the analysts at that institution, not the publishing platform, and should not be treated as investment advice.
