The company moved to an AI platform to validate key strategic decisions over the past three years! Goldman Sachs is bullish on Microsoft (MSFT.US) to $640

The Zhitong Finance App learned that Goldman Sachs recently released a research report saying that the bank has held a series of investor meetings with Microsoft (MSFT.US) executives in the past two weeks. Goldman Sachs said it feels that Microsoft has made a number of key AI decisions in the past three years, and as companies evaluate the importance of AI platform products, these decisions are paying off.

Specifically, they include: 1) upfront long-term capital expenditure, giving the company greater flexibility in short-term capex decisions; 2) balancing capital expenses between self-developed applications and third-party customers to improve the output quality of CoPilot and MAi (Microsoft's self-developed AI model); 3) balancing capital expenses between cutting-edge model customers and enterprise customers, so that the increase in RPO (remaining performance obligations) of 51 billion US dollars in the fourth quarter came entirely from enterprises (not frontier model laboratories), and the company can reduce any single model during the evolution of the model ecosystem Dependency. Finally, Microsoft is using its decades of presence to drive the implementation of AI in enterprises, including its reputation for security and governance (such as Agent 365 bundled with E7), the ability to combine models with enterprise-specific contexts to achieve continuous improvement in enterprise environments, and the rapid expansion of front-line deployment engineers (FDEs) to support a huge backlog of commitments.

Goldman Sachs maintained its “buy” rating on Microsoft in the research report, and believes that future value discovery areas worth focusing on will focus on MAIA (Microsoft's self-developed AI chip) and MAI improvements, and the momentum of enterprise platforms entering the market; the 12-month target price is $640, which has nearly 28% upside compared to the stock's closing price of $501.61 on Monday.

Key topics

1) Regarding capital expenses

Goldman Sachs said it feels like Microsoft is sending a signal that it has stronger control over the supply chain and production capacity climbing compared to a year ago. This change is mainly reflected in:

The share of long-term capital expenditure has declined. Over the past three quarters, long-term capital expenditure as a share of total capital expenditure has dropped from about 50% to about 33%. This is partly because Microsoft has previously invested in long-term capital expenses such as land, construction, and “cold shell” (meaning that the building structure/shell is basically complete, but there are few or very few systems installed indoors).

In short-term capital expenditure, CPUs have now replaced GPUs as the largest capital expenditure component. At the same time, relevant decisions will still be postponed until the final stage of the supply constraint process as much as possible (in retrospect, the time for GPUs from loading and unloading to operation has now been reduced by 50%).

Microsoft has never indicated that CPU and GPU supplies are limited. Instead, the more general limitation comes from the data center space required to install these chips. Therefore, if the constrained portion of capital expenditure becomes smaller and smaller, the bank deduces that Microsoft as a whole has greater capacity and can flexibly adjust the size of unrestricted capital expenditure to meet demand.

Microsoft believes that the entire industry needs to clearly explain the costs and benefits of building new data center communities. In Quincy (home to one of Washington State's earliest data centers), the proportion of the local population living below the poverty line has dropped from more than 30% to less than 10%. Microsoft can observe the expansion of its most mature AI customers over the past 3 to 5 years, so it can make a relatively reasonable benchmark forecast for the expansion of enterprise customer demand in the next 3 to 5 years based on this.

2) About ROIC

Goldman Sachs points out that the economic benefits of the unit are superior to the same phase of the initial cloud computing cycle, which will support the formation of larger profit pools as each batch of capital expenditure matures.

An important experience in the cloud computing era is the adoption of a unified technology stack between Azure and self-developed applications (rather than different technology stacks, which require more detailed matching, which ultimately results in lower utilization rates). Furthermore, Microsoft was able to acquire AI-native customers from the first year, rather than being a late-entrant like the cloud computing cycle.

In response to investor feedback, Goldman Sachs believes Microsoft may provide more details about ROIC in the next few months (in review, Google revealed earlier this month that its AI server payback period was less than 2 years; CoreWeave disclosed 2.5 years at the time of the IPO).

At the same time, Goldman Sachs deduced the following: if every batch of capital expenditure in this cycle is ahead of the previous cycle, then unless rental income is unevenly distributed to semiconductors and token (token) companies, etc., the return on investment (ROI) of this cycle will also be ahead of the previous cycle.

In terms of semiconductors, Microsoft believes that the chip layer will become more diversified, just like previous developments in the CPU market. In terms of token rent, Microsoft believes that the company can provide sufficient value on top of third-party tokens while coordinating the relationship between first-party and third-party tokens, so the impact on profit margins will be very limited over time. Taken together, Microsoft believes that there are no structural reasons why the gross margin of the AI business cannot be close to the gross profit margin of the cloud computing business.

As the industry matures, the growth rate of cloud business revenue will instead be faster than the growth rate of capital expenditure, which is in line with the development rules of the cloud computing cycle. The challenge of predicting this cross curve is that demand signals are still increasing, causing this time point to be continuously delayed.

In terms of pricing, Microsoft has not adopted a comprehensive price increase method that is too general and too aggressive. Instead, it points out several pricing methods that can drive long-term customer lifetime value (LTV) optimization: a) the discount rate currently used for contract renewals is lower than normal; b) the launch of new SKUs, such as new products in the CPU field, which have a higher price point than existing products (the price reduction trend has been common in the industry over the past 10 years); c) dynamically allocating production capacity on a weekly basis to adjust between first-party applications, first-party model development and product roadmaps, and a broad range of customer groups (Q4 RPO (A month-on-month increase of 51 billion US dollars, all driven by orders from non-frontier laboratory customers). Many customers have made extensive commitments and partnered with Frontier Co (FDE) to enable Microsoft to collaborate with customers around their specific AI goals and longer-term development paths.

3) About the chip

Goldman Sachs notes that Microsoft's goal is to provide the lowest possible cost of tokens. Microsoft owns Jalapeno's intellectual property, so it has two cards in the chip race, the other being MAIA. The MAIA 200 has made progress recently, and its benchmark performance is already well comparable to Trainium. Similar to Microsoft's overall vertical integration strategy (such as the model layer), using multiple chips can avoid excessive reliance on a particular solution or generation of architectures while the new chip technology is still rapidly evolving. This also allows customers to select previous-generation chips when appropriate.

4) On the pace of development of cutting-edge models

Goldman Sachs notes that Microsoft believes any new technology must create social and economic value. Microsoft has a long history of encouraging regulatory measures to support safety and security.

The current bottleneck is not what level the original performance of cutting-edge models has reached, but how to apply this performance to enterprise usage scenarios. So even after an extreme ideological experiment — assuming that the performance of cutting-edge models doesn't improve any further — there is still huge room for growth.

Regarding model diversity, Microsoft has reiterated its view that the model ecosystem is evolving in real time into a collection of models that are highly heterogeneous in terms of performance and price ranges. Currently, the number of models hosted by Foundry has reached 11,000, and the company intends to adopt a multi-model strategy.

Other than reaffirming that it will not pay for OpenAI or MAI's tokens, Microsoft has provided no directional information on the economic benefits between the different models. Microsoft said that even if some models require payment, the company can accept it because of its ability to cross-sell platform services with higher margins. Microsoft's Scout is an autonomous agent, built on OpenClaw, but has undergone an enterprise-level transformation.

Supporting a lower-cost open source model can in turn free up budget to increase usage or drive more business transformation. Internally, Microsoft has successfully experimented: limiting engineering teams to using cutting-edge models only when they can actually benefit from cutting-edge models, thereby optimizing resource allocation and avoiding the use of cutting-edge models by engineers who don't see significant benefits.

Will changes in the pace of development of cutting-edge models affect capital expenditure? It should be recalled that Microsoft's capital expenditure decisions depend on current demand signals, and changes in the pace of development of cutting-edge models have not affected demand signals at present.

Microsoft can of course choose to allocate additional capital expenses to the cutting edge model, but the current approach is to prioritize production capacity that can serve a broad range of customer groups and cover a variety of potential workloads. The reduction in Stargate's production capacity is a good example; the fact that all RPO growth in the fourth quarter came from non-frontier customers is another example.

5) About your own model strategy

Goldman Sachs said Microsoft is reaffirming that it has benefited from OpenAI's experience in cutting-edge fields. This experience can be further used by Microsoft and help it build its own model independence after 2032. As a result, Microsoft has a model technology stack with a clear development context. These technology stacks absorb OpenAI's experience but do not rely on OpenAI.

For MAI, Microsoft's current focus is on combining cutting-edge intellectual property with business areas where Microsoft has significant field experience, including knowledge workers (with 17EB data in the M365 system), programming, and security. Benchmarks have shown significant advancements in cost/performance. A recent example is the launch of Project Perception in the security sector to continuously conduct penetration testing.

6) About M365

Microsoft emphasized that FY27 was the first year in recent years to guide the acceleration of revenue growth. Judging from the penetration rate of E5 and E7, Microsoft points out that the newly added users are mainly concentrated in the low-end market, including frontline employees and small and medium-sized enterprises. The E7 launch performed well, in part because it bundled Agent 365 features to provide customers with visibility and value. AI will be adopted in different forms across the customer base, but penetration disclosure may become less useful.

The pricing of application software will also change: although more value will be embedded in seat-based (seat-based) SKUs, consumer billing elements will gradually be added in the future as the software addresses the “labor unit/outcome” issue more comprehensively and not just around the added value of seats. In the long run, the dollar-scale opportunities brought by consumer businesses will be greater than the dollar-scale opportunities brought by each user.

7) About the risks of software abstraction layers compared to cutting-edge models

Microsoft pointed out that the competitive landscape in the field of cutting-edge models is changing dynamically, so companies do not want to be locked into a specific cutting-edge model. At the same time, intelligent experiences will create more artifacts (artifacts) in existing software ecosystems, which in turn will enhance the value of software ecosystems such as M365.

Microsoft's core idea is that there is still plenty of value at the “harness level” (harness level), including ongoing security, context, and cost optimization, and that model companies may not be able to replicate these capabilities in a model-agnostic way.