When Shawn and I started this project, we didn’t want our portfolio to become a Mag7 proxy. And I think we did a good job at that. Granted, Google is our biggest position, and Amazon has quickly become one of our biggest positions as well, after dipping into the $190s.

But Amazon was only our second Mag7 position, and we added it after looking at more than 60 companies. And in the end, our job is not to have the most differentiated portfolio, but one that we feel incorporates the highest-quality companies at attractive prices.

In that spirit, we will look at Microsoft today. It’s trading well below its 10-year averages right now, and I want to find out whether that’s an opportunity or whether Microsoft is truly in deeper trouble than most investors think.

Let’s dive in!

— Daniel

I could easily spend half the newsletter on Microsoft’s history and how it changed the tech landscape forever. But this is a deep dive for investors, and Microsoft isn’t the least complex business out there, to say the least. So I fear we need that space to cover the business itself.

Besides that, there are plenty of sources for anyone who wants a little history session on Microsoft. I’ll link to one in the Notes at the end.

So let’s talk business! Microsoft has three reporting segments, and the power of the business has always been less about any individual product and more about how they reinforce each other. Understanding that flywheel is essential to understanding both the opportunity and the risk that Microsoft currently offers.

The first segment is also what comes to my mind first when Microsoft comes up. Office applications such as Word, Excel, or PowerPoint. But there are many more applications to consider, such as Teams, Outlook, SharePoint, LinkedIn, and Dynamics.

This segment generated around $120 billion in annual revenue at an operating margin close to 60%. It’s remarkable to think about how dominant this software has been over the decades and how much profit it has generated in that time.

The durability of this segment comes from switching costs that are genuinely extraordinary. The .docx, .xlsx, and .pptx formats have become international standards embedded in contracts, legal filings, regulatory submissions, and thousands of enterprise workflows. They are also seen as fundamental skills in the enterprise world. When I studied finance and economics in college, there were Excel courses offered, and it was recommended that students include the certificates they received at the end of the courses on their CVs.

If you spent all that time getting better at Excel, you obviously won’t choose the first best alternative and switch. Even in the age of AI, Office data formats remain the primary output format. I usually work in the Google product suite now (Google Sheets, Google Docs, and Google Slides) since working at The Investor’s Podcast, but when I use AI tools, they still default to Excel, Word, and PowerPoint.

I won’t go into the details on some of the other parts of this segment, mostly because they are almost irrelevant to Microsoft’s economics, but if you still wanna know how, for example, LinkedIn makes money, check out our podcast to go deeper.

If you strip out LinkedIn and some smaller pieces, the pure software profit pool — Word, Excel, PowerPoint, Outlook, Teams — is roughly $70 billion in annual operating profit!

That single profit pool is larger than most S&P 500 companies’ total revenue.

For many, many years, this has been Microsoft's golden goose and the reason the market valued this company at 40x earnings. With AI on the horizon, though, that has changed. More on that shortly!

The second business segment is all about the cloud. Azure is the core of this segment, alongside GitHub, SQL Server, and Windows Server. The segment generates roughly $125 billion annually, growing at approximately 29% year-over-year. Azure alone, though, has a run rate north of $75 billion — growing at an astounding 40%.

To understand why Microsoft’s cloud positioning is so compelling, we need to look at the three layers of the cloud.

At the bottom is Infrastructure as a Service, or IaaS — raw computing power, storage, and networking rented by the hour. As a customer, you essentially get a virtual machine, and you're responsible for everything that runs on it. AWS pioneered this model in 2006 and still leads the global cloud market with roughly 30% market share. Azure sits at around 20-22%, Google Cloud at 13%. Together, the three of them control more than 60% of a market that crossed $100 billion in quarterly revenue in Q4 of last year, growing at nearly 30% annually.

The IaaS layer is increasingly commoditized. Virtual machines from AWS, Azure, and Google are largely interchangeable for most workloads. When I say commoditized, it doesn’t mean that there is no pricing power here, though. To compete in this market, you would have to compete with Amazon, Google, and Microsoft on capex investments. That’s obviously a huge barrier to entry.

So even though there’s not a highly differentiated product, it’s an oligopoly, and since transferring all your workloads from one cloud to another is a huge headache, you tend to stay with your cloud provider.

Beyond that, the other two layers increase customer lock-in. Platform as a Service, or PaaS, is the second layer. This is where developers get pre-built building blocks they can plug directly into their applications, rather than building everything from scratch.

So imagine you're building a new HR software tool for a company. At some point, your software needs to let employees log in. Of course, you could build that yourself — write the code to store passwords securely, handle forgotten passwords, add two-factor authentication, and manage who has access to what. But that would take months and be quite expensive. Alternatively, you can simply plug in Microsoft Entra ID, a platform service that handles it all in a fraction of the time, without any additional coding work.

The same goes for databases, analytics tools, and machine learning services. Microsoft has pre-built versions of all of them that developers can snap into their applications.

As you can imagine, this layer is a lot less commoditized than the infrastructure. If your software tools are woven into Microsoft’s own solutions, switching would mean that you have to rebuild a large part of the software you use on a daily basis.

And compared to Amazon and Google, Microsoft is very strong in the platform layer. Microsoft owns GitHub, the dominant global code repository with 100 million registered developers, and Visual Studio Code, the most widely used development environment in the world with roughly 70-80% developer market share. These are the tools developers use before they ever write a line of cloud deployment code. When you build your application on GitHub using Visual Studio and want to deploy it, Azure is the path of least resistance.

AWS has no equivalent at the developer tooling layer. Google has some developer tools, but nothing on GitHub's scale. This matters because, in most large enterprises, cloud provider decisions are driven more by engineering teams than by IT leadership, and engineering teams have a strong bias toward the platform on which their daily tools already run.

Above PaaS is the software layer. That’s where you find the actual applications that business users interact with. Microsoft 365, Teams, Dynamics. Once again, Microsoft has an obvious advantage here. It is the only major hyperscaler that’s simultaneously competitive across all three layers, with an enormous installed base at the top.

AWS has close to no enterprise software presence. Google Workspace competes at the application layer but doesn't have Microsoft's depth or enterprise penetration. That three-layer integration creates vertical lock-in that is almost impossible for a (mostly) single-layer competitor to replicate, which is a complicated way of saying that it’s unlikely Microsoft’s customers will go anywhere.

Part of the truth, though, is that both the PaaS and the enterprise software layer are facing the biggest disruption risk in Microsoft’s history. And it’s not only me or the market saying that. Satya Nadella, the CEO of Microsoft, is publicly speaking up about this as well. He calls most software CRUD:

"Most software today is just a CRUD (Create, Read, Update, Delete) interface for a database. AI agents are going to replace that business logic. If your software is just a database wrapper, it’s at risk."

He also said that the new era isn’t about apps, but agents. So let’s take a closer look at the parts of Microsoft that are at risk of being disrupted by AI, and, simultaneously, at Microsoft’s initiatives to prevent that.

Just quickly, though, so I’ve mentioned it, there’s a third segment called ”More Personal Computing,” which includes Windows, Xbox, and the Activision Blizzard gaming portfolio, but it's not central to the investment case today, and we covered it in depth on the podcast, so I’d like to focus on Microsoft’s major AI pivot.

For the past 18 years, Microsoft has successfully defended its productivity empire against Google and any other competitor. Google launched Docs, Sheets, and Slides in 2006 and won the low end of the market — students, startups, and small businesses like us (Shawn and I are Google Workspace power users). But 20 years later, Microsoft Office still completely dominates enterprise. Christensen's disruptive innovation model (The Innovator’s Dilemma) suggests that Google should’ve already been able to disrupt Microsoft’s business. In fact, that should’ve happened about 8 years ago!

And yet, it hasn’t…

There are multiple reasons why Google couldn’t do it. One is that file formats like .docx and .xlsx, as we’ve discussed, are deeply embedded across the enterprise world. That problem would be easy to solve in theory, though. It takes about 10 seconds to turn a Google Sheet into an Excel file. But you then have to live with some formatting issues from time to time.

The greater switching costs stem from the fact that Excel remains much more powerful than Google Sheets. While that doesn’t matter to 95% of users, it does matter for the top 5% of users at any given enterprise who use advanced features like VBA Macros and Power Query.

And as long as the top 5% can only work with Excel, it makes sense to stick with it across the entire enterprise. I’m not blind to the fact, though, that Google could’ve integrated these advanced features as well. Ultimately, the switching-cost moat will come down to the fact that integrating new software across an entire enterprise is just a huge headache. Google knows that and decided not to double down on a war that was lost from the beginning.

So why should we worry now when even Google has backed down? Because AI isn't offering a competing product. AI is offering to reduce the need for the product. When someone uses Google Docs instead of Word, they spend the same amount of human time creating documents, just in a different environment. But when someone uses an LLM to generate the first draft in 30 seconds and spends 20 minutes editing that draft rather than working in the original Word file for hours to go from zero to a finished product, then Word becomes much less important.

And Microsoft’s per-seat subscription model is built on the assumption that you need one license for each human doing cognitive work. If AI agents perform a meaningful fraction of that cognitive work, it’s questionable how viable the per-seat model still is. This isn’t the most innovative insight, I know. We’ve talked about the pressure on the seat-based model in pretty much every SaaS episode we did.

What is different today is that Microsoft’s tools have never been special or unique. Microsoft is a master copycat, and by bundling its services and selling those bundles to enterprise customers, there was little to no incentive for customers to choose any of the outside tools. Even if they were better. Zoom and Slack are better than Teams; Gmail is better than Outlook; and Google Slides or Docs are pretty much as good as PowerPoint or Word. But it didn’t make sense to pay for Slack or Zoom when you already have similar tools through your Microsoft bundle.

So if LLMs disrupt the heart of Microsoft — the 365 suite — and the platform and software layer, then Microsoft has a serious problem. And that’s why Satya Nadella has warned that software that’s just a database wrapper is at risk.

Nadella said this in the most recent earnings call last week:

About 5 years ago, Mark Zuckerberg thought he knew where the future of tech would be – in the metaverse. He spent tens of billions of dollars developing what turned out to be a surprisingly ugly and useless virtual world. And not only did Zuckerberg invest almost a hundred billion dollars directly, but he also renamed the company to “Meta.”

Why do I mention this? Well, when I listened to Microsoft’s earnings calls and went through their product suite, I couldn’t help but think of this debacle. Microsoft now has 80(!) products that include the name “Copilot.” Copilot is everywhere, and it’s supposed to be Microsoft’s answer to the AI threat.

Through Copilot, Microsoft wants to become the company that sells AI agents. After two years of commercial availability, it has 20 million paying subscribers. That sounds like a lot until you remember that Microsoft has 450 million commercial seats. Twenty million is 4.4% penetration. However, that’s better than the 3.3% penetration before last week’s earnings update.

Alas, I’m reaching for straws here.

The first problem is pretty straightforward…Copilot isn’t working well. Especially compared to other LLMs. This sounds contradictory at first because Copilot runs on OpenAI's models. So, shouldn’t it work at least as well as ChatGPT, and perhaps better, since it has a ton of context thanks to owning the user’s workspace?

One reason for the difference is that Microsoft applies additional safety filters and content restrictions on top of the base model to make it “enterprise-appropriate.” This reduces the, let’s call it “creative flexibility” that LLMs sometimes have. The downside is that it makes ChatGPT feel less powerful in open-ended tasks.

Even the fact that Copilot can reason through all the files you have access to turned into a problem for many companies when they trialed Copilot. When I go through TIP’s Google Drive (Google’s equivalent to Microsoft’s OneDrive), I don’t generally hunt for files I shouldn’t have access to but still do. I’d like to think I have more important things to do.

Copilot would do exactly that, though. It goes through all files you technically have access to. Thus, several enterprises that ran early Copilot trials discovered Copilot was reasoning with information that certain employees shouldn’t have had access to.

On an individual basis, this problem is manageable, but when you lose track of permissions for thousands of employees, the risk becomes intolerable. To be clear, if Copilot were a huge value-add, I’m sure companies would work as quickly as possible to resolve the permission issues and then return to Copilot. But with Copilot not being a huge help anyway and the $30-per-user monthly cost, the low Copilot penetration isn’t surprising.

However, I wouldn’t be surprised if Microsoft turns around the Copilot story and makes it a success still. After all, every couple of months, another LLM provider seems to have an insurmountable lead. First ChatGPT, then Gemini, and now Claude.

Since we, as investors, want to be ready for all possible scenarios, let’s discuss the economics of Microsoft if we assume that, first, an enterprise keeps its Microsoft software — Word, Excel, Teams — but deploys a third-party AI agent. Say they use Anthropic's Claude to handle contract drafting, or OpenAI's GPT to process accounts payable.

In this scenario, Microsoft keeps the M365 subscription revenue, but they capture nothing from the agent layer, nothing from model inference, and they only cover the compute from OpenAI, since Claude’s main compute partner is Amazon’s AWS. That’s certainly not the most bullish outcome.

In the second scenario, the enterprise uses Microsoft's Copilot agent, built on Microsoft's infrastructure and eventually on Microsoft's own AI models. This is the ideal outcome for Microsoft from an economic standpoint. They capture the subscription, the agent orchestration layer, the compute, and, eventually, the full model inference margin once they shift Copilot off OpenAI's models and onto their own.

This is precisely why Microsoft is investing heavily in its own AI model program, internally called MAI-1, and why the Maia 200 custom chips are important. Every query Copilot routes to Microsoft's own model rather than OpenAI's is one fewer check written to a partner. The economics of this scenario are significantly better, but it requires both that Copilot works well enough that enterprises choose it over third-party alternatives and that Microsoft's own models are competitive enough to replace OpenAI's.

Importantly, “competitive enough” doesn’t mean better. Microsoft still has its bundling advantage. They never need to be better, just good enough.

And then last but not least, scenario three. In that case, Azure becomes more like an AI model distributor. Customers could use Claude, ChatGPT, or Llama, but access them through Microsoft’s infrastructure. The business model would then resemble an app store, where Microsoft takes a cut when a customer uses its platform and, let’s say, Claude. The reason it would still be attractive for enterprise clients to go through Microsoft rather than bypass it is that Microsoft offers the kind of secure, auditable environment that matters enormously to enterprise clients.

This might be the most realistic vision for the future. At least the near term. Azure AI Foundry is basically working on exactly this — being a model-agnostic way for clients to interact with.

We can’t talk about a Mag7 company in 2026 without talking about capex investments. For whatever reason, I was under the impression that the Mag7 have always spent enormous amounts of money on capex. And relatively, that’s certainly true. But there is no doubt that it has reached a new level in recent years.

Microsoft spent $72 billion on capital expenditures in just the first half of their current fiscal year. The full-year projection is somewhere between $120 and $146 billion. For context: fiscal 2023 capex was $28 billion. They’ll spend more in a single quarter this year than they spent in an entire fiscal year two and a half years ago.

And about two-thirds of that capex is in short-lived assets — primarily GPU chips with three-to-five-year useful lives. The depreciation charges from those investments will be in the tens of billions of dollars annually and certainly reduce earnings and margins in the next few years.

That said, my doubts about Microsoft don’t stem from its capital investments. To me, the risk of underinvesting seems greater than the risk of building too much capacity today. Losing a customer to AWS or Google right now does not only mean that you’ll lose revenue, but you also lose that customer’s data to a competitor whose AI will then get smarter. Plus, cloud customers are sticky. So once you lose them, you likely won’t win them back.

When discussing Microsoft’s AI strategy, we have to talk about OpenAI. As of today, Microsoft has committed between $13 and $14 billion to OpenAI across multiple funding rounds in exchange for roughly 49% economic interest and a partnership under which OpenAI's workloads run on Azure, and Microsoft gets the rights to embed OpenAI's models in its own products.

That partnership created Copilot, built Azure OpenAI Service into one of the fastest-growing cloud products ever, and drove Microsoft's commercial backlog, essentially a pipeline of committed future revenue, to surge 100% year-over-year to ~$630 billion.

Looking at it from this angle, it’s unimaginable how one could be critical of this investment. If anything, it should be up for debate to be one of the most successful investments of all time, considering OpenAI’s current valuation of about $850 billion.

But there’s a lot of side noise that makes it seem as if this partnership isn't working as well as it should. Not only did Copilot struggle more than expected, but OpenAI has also publicly voiced its discontent about allegedly not being able to partner with certain companies due to Microsoft’s restrictions. And yet, there seem to be loopholes that OpenAI has found to work directly with competitors instead of Microsoft and Azure.

For example, OpenAI launched a new enterprise AI platform and signed a $50 billion cloud deal with Amazon Web Services, arguing the new product fell outside the original exclusivity agreement with Microsoft. OpenAI is also converting from a nonprofit to a for-profit company, and whether Microsoft's economic interest fully converts into a comparable stake in the new entity is not yet settled.

Microsoft’s AI Foundry is definitely also a response to the weakening relationship with OpenAI. All that said, assuming Microsoft keeps its 49% economic interest after the shift to a for-profit entity, this has certainly been a successful investment.

If you check Microsoft’s valuation metrics, you’d see that it’s trading at a forward P/E of ~22x. Significantly cheaper than usual. But we shouldn’t forget the depreciation impact from all that AI capex, which will become even more severe in the coming years.

It’s also why Microsoft's free cash flow conversion used to be much higher, and it might take a while, if ever, until we get back to those levels.

With this in mind, let’s walk through the assumptions of my model.

The base case assumes Microsoft successfully navigates the AI transition on roughly the terms management has described. Revenue compounds at approximately 17% annually through fiscal 2030, driven by Azure growing in the high twenties as capacity constraints ease and Copilot adoption improving from 3% to 10 to 12% of the commercial installed base. This might seem like a stretch today, but we shouldn’t forget that bundling and upselling are Microsoft’s biggest strengths.

It hasn’t yet worked for Copilot, but once that changes, conversions could go up quickly. And the (AI-)landscape in 5 years will, without a doubt, look completely different from today.

On the margin side, the key assumption is that normalized free cash flow margins, which are currently depressed by the capex buildout, will recover modestly from around 32% of revenue today to 35% by fiscal 2030 as the spending peak passes and depreciation stabilizes. Pre-AI, Microsoft was running normalized FCF margins above 38%, so I’m not assuming a full recovery to those levels within the forecast period.

On those assumptions, normalized free cash flow per share reaches approximately $30 by fiscal 2030. At a 25x exit multiple, the present value of that cash flow implies a fair value of approximately $513 per share today. That would imply a return of about 13-14% from today’s prices. Granted, revenue growing at 17% for the next five years is not the most conservative base case I have ever modeled. And the vast majority of that growth needs to come from the Cloud.

Beyond reducing my top-line growth assumption, the main difference in the bear case is that I halved my expectations for FCF/share growth. If we think back to the three scenarios of how the AI revolution might turn out for Microsoft, the bear case resembles the scenario in which Microsoft is unable to play a major role as an AI player. Copilot fails, the Foundry does not become the way for most enterprise clients to access Claude and co., and Microsoft’s seat-based Office Suite model starts to lose pricing power.

Under these assumptions, margins begin to deteriorate, and growth slows. If that’s the case, the multiple should contract as well. I’ve gone with an exit multiple of 20x, which leads to a fair value of $285. Applying a margin of safety of 20%, you get to $230. The expected return from today’s prices, then, would be close to 1%.

What do I take from this? Microsoft is far from being a no-brainer. If you believe in its AI positioning, it can yield good returns, but if not, there’s still a lot of downside. I don’t consider it likely that Microsoft will fail the transition, but given the upside potential I see (I could very well underestimate it), it’s not a bet that Shawn, Kyle, and I would take at this time. Especially compared to the companies we currently own in our portfolio.

For more Deep Dives and Portfolio Updates, you can listen to our podcast here.

More updates on our Intrinsic Value Portfolio below 👇

“We concentrate on quality, not frequency. If a significant issue arises, you will hear from me, but it will not be through quarterly commentary, given our long-term horizon.”

— Greg Abel

📺 WATCH: The Story of Microsoft by Jack Chapple

🎧 LISTEN: Drew Cohen’s Interview with Rose Celine on finding Compounders

📖 READ: Hedge Fund Alpha Article about our Intrinsic Value Conference in Omaha

You can also read our archive of past Intrinsic Value breakdowns, in case you’ve missed any, here — we’ve covered companies ranging from Alphabet to Airbnb, AutoZone, Nintendo, John Deere, Coupang, and more!

Enjoy reading this newsletter? Forward it to a friend.

Was this newsletter forwarded to you? Sign up here.

Join the waitlist for our Intrinsic Value Community of investors

Shawn & Daniel use Fiscal.ai for every company they research — use their referral link to get started with a 15% discount!

Use the promo code STOCKS15 at checkout for 15% off our popular course “How To Get Started With Stocks.”

Follow us on Twitter.

Read our full archive of Intrinsic Value Breakdowns here

Keep an eye on your inbox for our newsletters on Sundays. If you have any feedback for us, simply respond to this email or message [email protected].