Review: Can Microsoft Compete? Part III
Distribution, combining layers, and conclusion (which is: yes).
Summary
What it does: enterprise software.
Elevator pitch: arguably one of the finest franchises ever built, with deep distribution advantages that let it deliver innovations cheaply to a huge customer base.
Mental model: moat (read about my mental models here).
Valuation and potential returns: 21x June 2027 EPS estimates and growing EPS 20% per year.
Exchange and ticker: Nasdaq, MSFT
Stock price and market cap: $422, $3.1t.
Do I own it? Yes.
IR website: here.
Tag for finding my other articles on this stock: MSFT
About this blog: I have been investing for 25 years, professionally and personally. I look for stocks that have a high probability of compounding at 15% for at least 5 years with limited downside. I write these stocks up on my blog. You can find more about me, my philosophy, my mental models, and my portfolio structure on my site.
Disclaimer: This post is for informational and educational purposes only. Building Arks is not licensed or regulated to provide any financial advisory service and nothing published by Building Arks should be taken as a recommendation to buy or sell securities, relied upon as financial advice, or treated as individual investment advice designed to meet your personal financial needs. You are advised to discuss your personal investment needs and options with qualified financial advisers. Building Arks uses information sources believed to be reliable, but does not guarantee the accuracy of the information in this post. The opinions expressed in this post are those of the publisher and are subject to change without notice. The publisher may or may not hold positions in the securities discussed in this post and may purchase or sell such positions without notice.
Welcome to the final instalment in my Microsoft saga. The goal of this piece is to determine whether Microsoft can compete over the next 5-10 years. Part 1 introduced the layers of AI the AI stack and looked at two, silicon and infrastructure. Part 2 dived into two more, LLMs and the agentic stack. So far we have concluded that:
In silicon, Microsoft is behind because it started late, but Maia 200 is promising and Microsoft will close the gap.
Infrastructure will commoditise slowly, and a new utility compute industry will eventually allow hyperscalers to go (partway) back to being capital light.
LLMs will commoditise behind the frontier, creating a rolling wave of commoditisation and increasingly cheap intelligence.
Microsoft is well placed to 1) help enterprises develop and manage their own agents, and 2) develop proprietary agents that can be sold into large verticals. The agentic TAM is huge.
Part 3 starts with distribution and bundling, and then considers whether layers can be combined to create competitive advantage before drawing everything together in conclusion.
Distribution and bundling
Distribution and bundling have been Microsoft’s deepest moat for years. Through Windows and Office alone, they are in virtually every enterprise IT department and on virtually every enterprise employee’s screen. When a competing product comes along, they clone it and bundle it into an existing suite of products. Suddenly it’s on every desktop, for nothing. It’s not as good as the standalone version, but it is there and it is free. Adoption skyrockets. Over time Microsoft improves the product and then monetises it by re-segmenting the bundle.
On top of that, Microsoft sells trust. They wrap the bundle in all the things IT departments love - permissioning structures, data security. But they don’t charge too much. They’re selling software, so the gross margin on the incremental sale is extremely high. They can afford to keep the price of the bundle well below the total price of all the standalone options you’d have to buy to replace it. (Years ago a colleague of mine looked at this. Everyone at the company lived in Microsoft products all day. Literally all day. And yet the company spent less with Microsoft than it did on milk for coffee and tea.)
The bundle has all the products you want, wrapped in all the compliance and security stuff you need; it requires managing one vendor relationship rather than 10; and it costs peanuts. What IT department wouldn’t use it? Code was never the moat. The cost to replicate the products has been falling for years and is tiny compared to the market opportunity. Distribution, bundling, trust, and price were the moat.
Does AI change this? I don’t think so, but there are wrinkles.
We saw in Part II that Microsoft is building 1) the cloud AI stack for enterprises to build bespoke agents, and 2) first party agents for big use cases. In the AI stack, distribution is a clear advantage but bundling isn’t. Microsoft has few peers when it comes to selling this kind of relationship to enterprise, although it is clear that Google, Amazon, Anthropic, and Open AI will compete. But the sales motion here is less about bundling and more about monolithic contracts, and the economics are about leveraging fixed costs. In proprietary agents, however, distribution and bundling may be a superpower. That is where I want to focus.
Bundling isn’t easy with frontier AI. When Slack launched, Microsoft cloned it, creating Teams. Whatever the cost of doing that was, it vanished into Microsoft’s overall R&D budget. By contrast frontier models cost tens if not hundreds of billions to train, and AI labs may only have a year or two to monetise them before their capabilities are commoditised. Microsoft has chosen not to develop frontier models itself, quite possibly because of that exact dynamic. The result is that it now distributes third party products, not its own. This creates a dependency that was not present before.
However, bundling commoditised AI might be very different. My thesis really hinges on one thing: for any given task, a given level of intelligence is required. That level of intelligence may be impossible one year, frontier the year after, and commoditised the year after that. As the frontier advances, the trailing edge commoditises, the cost of inference falls, and the pricing model switches to seat + consumption, Microsoft can bundle more and more intelligence at less and less cost. And that strikes me as an immensely (and increasingly) powerful proposition for both TAM and margins.
There are about 400m M365 Commercial seats worldwide. Of these, by March 2026, over 20m had adopted the paid M365 Copilot add-on, up 35% q/q. That is with Copilot’s capabilities today. In 5 years Copilot might still not be as good as the frontier models, but in absolute terms its capabilities will be incredible, it will be cheap, and it will be bundled into an enterprise’s existing financial and control arrangements. The chances of increased adoption (and maybe increased growth in M365 Commercial seats, too) are strong.
It’s obviously dangerous to extrapolate from one example, but this post on X captures what I am trying to describe. This user found that Excel in Agent mode can now update financial models from press releases. This might be child’s play for Claude, but that’s not the point. The point is that Copilot is improving and the user 1) I doesn’t care which LLM was used, just that it works; 2) prefers letting Microsoft choose the LLM over getting locked into one frontier lab; 3) might now use Excel more, not less, and for more productive things; and 4) would happily pay Microsoft extra for this capability. And as he says in the comments: why bother using the Claude Excel plugin if Excel Agent mode just works? I think that as commodity models advance, this thought process will play out across millions of use cases in thousands of enterprises.
If this is right it creates a couple of interesting dynamics:
Copilot may always be behind the frontier and may never have a great reputation as a result. But so long as it keeps getting better, that might not matter.
Frontier models may always dominate frontier workflows, but they will also have to charge frontier prices. As Copilot gets better, workflows developed on Claude might come back to Microsoft as commoditisation and bundling kicks in.
If this plays out as I think, Microsoft will become a critical distribution partner for commoditised LLMs. As such Microsoft can be an agent of model commoditisation, not just a beneficiary of it. It will also be an (unlikely!) ally of regulatory agencies seeking to prevent any potential winner-take-most dynamic in the LLM market.
The economics of bundling AI aren’t as clean as the economics of bundling software. The nth copy of a piece of software can be sold for virtually zero marginal cost, whereas intelligence costs money every time you use it. As discussed in Part II, Microsoft is moving towards a seat + consumption pricing model. This should both increase Microsoft’s revenues and improve Copilot’s performance. But can they make a margin?
The bear case is that there isn’t a CTO on earth who hasn’t heard of Anthropic and OpenAI. Under pressure to adopt AI and aware that Copilot’s reputation is weak, they will test all options. They will not want to get locked into a single LLM provider, but if they somehow do (despite model commoditisation) then the frontier labs might generate cash flow in excess of the cost of training the next model. In this scenario the frontier labs both disrupt Microsoft’s cloud/productivity/agentic layer and build proprietary infrastructure to put pressure on the hyperscalers.
The bull case is that with proprietary silicon, commoditised infrastructure, and commoditised trailing edge intelligence, Microsoft can deliver increasingly powerful intelligence cheaply through its unique distribution setup. Meanwhile frontier labs, desperate for cash to train their next model, will have to charge high prices for their older models, holding up a pricing umbrella for Microsoft.
There is clearly some validity to at least the first part of the bear argument. Open AI and Anthropic have bypassed Microsoft’s distribution advantage to some extent - otherwise their revenues would not be growing as fast as they are. Whether the second part of the argument holds - that these revenues are sticky - remains to be seen. It’s probably not binary - both arguments will be true to a degree. But on balance, I find the second argument far more compelling.
Conclusion on distribution and bundling
As the rolling wave of LLM commoditisation advances and inference costs fall, Microsoft will be able to bundle more and more intelligence into its product suites at lower and lower costs. This plays neatly to Microsoft’s time-honoured strategy: be the second mover, clone and bundle the product, drive adoption, make it better, and then slice and dice the bundle to monetise it. If this works Microsoft will always look like it is behind the leading edge, and Copilot may always have a poor reputation compared to frontier models, but it will be good enough for a vast and growing range of uses, it will be cheap, and it will be delivered at strong margins.
Combining layers
If anything renders Microsoft’s distribution moat obsolete it’s not the insurgency of OpenAI and Anthropic; it’s the full-stack advantage of Google.
In Part I I discussed the layers in the AI stack. Google owns all of them. It is producing its own silicon at scale. It is building infrastructure at pace. It owns its own frontier LLM, Gemini. It owns the full cloud stack needed to build and deploy agents. It owns distribution at vast scale to consumers. And it owns a huge search and advertising business that AI improves.
How does this full stack create advantage? Several ways:
Google can optimise margins across the whole stack, not individually for each layer. Said another way, it can use profits in one area to subsidise another while it builds scale and competitive advantage.
It can fully integrate and optimise its stack to deliver lower costs.
It can move fast, forcing cooperation across layers in ways hard to replicate when you don’t own every layer.
It can generate an immediate return on its LLM investments by deploying them to improve its existing businesses.
There will never be any question over whether Google has access to frontier intelligence (whereas Zuckerburg argues that OpenAI and Anthropic could claim “safety” reasons for not giving third parties access to their most advanced models).
I have not included the data flywheel here. Google has enormous amounts of proprietary data on which to train models. This data can be used to optimise models for improving Google’s existing businesses, like search. But I’m not convinced it is a huge advantage in enterprise, where Google is subject to the same restrictions as Microsoft (see Part II) and where Microsoft has more data.
The counterargument is that Microsoft also owns a lot of the stack. It doesn’t own a frontier model, but that might be a blessing in disguise given how unclear the ROI is in that layer. Instead, Microsoft owns the layers where ROI is clearer - silicon, infrastructure, cloud/agent stack, and distribution. If it can integrate these well, it should get sufficiently close to Google on cost, margin, and speed. Rather than locking clients into one LLM - which they will not want - it offers a marketplace of thousands of them. It absolutely owns enterprise distribution. Google does have enterprise distribution - over the last decade Google Cloud has become a formidable competitor to Microsoft’s Azure, and it provides an effective stack to customers wanting to build their own agents. But Google does not have an operating system and productivity apps already installed on the desktop of every user. That makes it more difficult to create proprietary agents and roll them out for free, without asking, and within existing security and permissioning arrangements. Finally, Microsoft is building cheap, optimised LLMs for specific, large use cases. In these cases, it will have genuine full stack economics. As cheap models get better and better, I expect Microsoft to address more and more use cases with this full-stack architecture.
I don’t mean to dismiss Google. It is certainly a formidable competitor. Gemini reasons better than Copilot, Google’s cloud business has built strong relationships across enterprise, and there is plenty of anecdotal data to suggest that Google is willing to discount heavily to encourage enterprise to switch. And there are advantages to the full stack. But Microsoft has advantages too. I am not sure Google’s full stack gives it a unique right to win in enterprise.
Conclusion
Here is my Microsoft hypothesis per layer:
Silicon: Microsoft is behind but catching up with Google and Amazon.
Infrastructure: Likely to commoditise eventually, especially behind the frontier, creating a new utility compute industry and allowing hyperscalers to return to more capital-light ways.
LLMs: A rolling wave of commoditisation, with frontier models never more than a year or two ahead of cheaper models. Microsoft will be the distributor of increasingly powerful commoditised AI.
Agents: A goldmine, with a deep cloud stack allowing customers to develop, orchestrate, monitor, and control agents, plus extremely valuable horizontal and vertical proprietary agents. TAM is every process in every enterprise.
Distribution: Microsoft’s deepest moat. Microsoft will bundle increasingly powerful intelligence into existing products, wrapping it in trust and pricing it cheaply. It will be hard for IT departments not to buy it.
AI is moving fast and the competitive dynamics are fluid. It is hard to develop confidence in where profits will eventually be made and who will make them. One strategy for dealing with this would be to reformulate investment theses whenever there is new news. I think this leads to high trading costs and chasing the latest hot stock. I prefer to form a 5-10 year high-level thesis based on an understanding of competitive dynamics and moat formation. That thesis can be assessed against short term newsflow, but it is important to give it time to play out. It is also important to diversify exposures.
Microsoft is well positioned to be the distributor of commoditised AI to enterprise. The TAM is huge and the right pricing models combined with Microsoft’s distribution advantages should ensure attractive margins. With revenue growth and operating leverage, I expect that Microsoft can compound earnings at 15-20% for many years despite rising depreciation costs. At 21x June 2027 earnings (less if you give credit for the Open AI equity stake), I think Microsoft is a compelling investment.
To end: as I said at the start, I have not addressed the ultimate bear case - that artificial general intelligence renders all competitive advantage obsolete and reduces returns across the entire tech industry to the cost of capital. I have three things to say on this:
I see this as a low-probability outcome for various reasons; in fact it’s equally likely that Microsoft ends up making good margins distributing commoditised AGI.
If it does come to pass, affordability and quality of life will improve dramatically regardless of how an investment in Microsoft works out.
This risk is easy to hedge. Plenty of assets will do well in a scenario in which humans have more discretionary spending power and more free time.
Links to previous Reviews
Thanks for reading - if you enjoyed reading this please subscribe, like, and restack, and do get in touch if you have questions.
Pete