Review: Can Microsoft Compete? Part II
The next two layers: LLMs and Agents.
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: $421, $3.1tn.
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.
Part 1 of this review introduced the layers of AI and dived into two, silicon and infrastructure. It concluded that
Microsoft is behind in silicon 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.
Part 2 dives into the middle two layers: LLMs and Agents.
LLMs
Large Language Models are the engines behind AI. Is this a winner-take-most market, or will LLMs be commoditised? If it’s winner-take-most Microsoft has a real problem. It doesn’t own a leading LLM, so it won’t be the winner; in order to sell AI to its customers, therefore, it will have to buy from the winner, who will have the whip hand in negotiations and might bypass Microsoft entirely to own the customer relationship directly.
Winner-take-most could come about in two ways:
One LLM lab reaches artificial general intelligence first, it could pull ahead of the others in performance terms. AGI isn’t well-defined, but in this context it means the LLM would improve itself faster and at lower cost than is possible with humans managing the process.
Several LLM labs run out of money.
I think it is vanishingly improbable that either of these things will happen, for a variety of reasons.
The first is that it is in literally nobody’s interests for one LLM to dominate the future. Here’s a short list of people who would absolutely hate this outcome:
Every consumer.
Every enterprise.
Every voter.
Every government.
AI is an incredibly potent technology. AGI will be terrifying. The idea that one company would be allowed to monopolise that sort of power is effectively impossible. Governments, customers, competitors, and regulators will work hard to ensure this doesn’t happen. And if it looks like it’s going to happen the company concerned will be broken up, so winner-take-most isn’t even really in the interests of the winner.
The next reason is simply that I don’t think technology works this way. Technology seems to be inherently replicable. Like the 4-minute mile or the 2-hour marathon, once one person can do something, others follow. I can’t think of any major technologies that were cutting edge 5, 10, or 20 years ago that have not been replicated. I am not sure AI will be any different.
The evidence so far suggests that LLM advances are indeed replicable. There are perhaps 4-8 frontier labs. None has so far carved out a significant lead - on the contrary, models from these labs leapfrog each other in capability. Further, xAI is only 3 years old and Meta has stumbled, fallen behind, and is apparently catching up again. None of these datapoints argues for winner-take-most. On the contrary, the evidence suggests that hiring the right people and throwing money at them gets you a seat at the front of the race. And the number of “right people” will inevitably grow over time.
Equally important is the proliferation and performance of second-tier models. While the frontier models are breaking new ground at great cost, second-tier models are about a year behind and cost far less. This one-year gap varies by task and over time, but it doesn’t appear to be trending longer. This is critically important. A certain level of intelligence is required for any given task. Eventually, many models will be good enough for each task. Intelligence that is impossible this year might be frontier next year and commoditised the year after. In addition there is evidence that the software surrounding a model can be as important as the model itself: for example, Microsoft Critique uses multi-agent review to achieve better results than the underlying models alone, suggesting that combinations of commoditised models may become extremely capable.
I believe all of the above argues strongly for a rolling wave of commoditisation behind the ever-advancing frontier LLMs. Microsoft seems to agree: I think their strategy is to become the distributor of commoditised AI. This is a strategy shioft from the early days of AI in 2022/3. Back then they built everything on OpenAI models, arguing this drove utilisation and efficiency. Why has the strategy changed? A cynic might argue it’s because they don’t have a frontier model of their own and their relationship with OpenAI has become rather less cozy. But there might be a simpler answer: initially Microsoft didn’t have a choice, but 18 months later they did. ChatGPT kicked off the AI era in November, 2022. It took about a year for all the other frontier models to become available and months more for powerful open-weight models to proliferate. But proliferate the models did, and in May 2024 Microsoft publicly announced a model-agnostic architecture for the first time. Today there are an over 11,000 models available on Microsoft’s platform. That number alone is strong evidence that LLMs will commoditise.
To distribute commoditised AI, Microsoft puts together several things. First, the models, available via API. Second, the harness in which the models run. Third, the reams of enterprise data stored in Microsoft’s systems. And fourth, the customers.
Who benefits? The commoditised LLM benefits from broad distribution. The customer benefits from access to cheap models and from not locking themselves into a frontier lab that might not be competitive in the future. And Microsoft benefits, because amongst other things:
Inference explodes. Models are cheap to run. Labs don’t have to invest in distribution. Low costs drive AI adoption.
Costs can be optimised by routing tasks to the optimal model.
Performance can be optimised by using domain-specific architecture and/or more than one model.
By building proprietary low-cost models and agents for specific use cases, Microsoft can build vertical niches with full-stack economics.
There are, of course, counterarguments to the commoditisation thesis. Perhaps the best is the most obvious: the independent frontier labs are growing revenue like weeds. Anthropic’s run-rate revenue reached $30bn in April. This is over 10% of the revenue run rate Microsoft has taken decades to build, and is up 3x in 4 months. OpenAI is reportedly on a similar trend. On the face of it, exploding revenues suggests these will be powerful companies in the future. But the more revenue the labs make the longer they stay alive, the more models there are, and the more likely commoditisation is. Also, I am not arguing that the frontier becomes commoditised. Indeed, the labs may be able to keep growing revenue by solving ever-more impressive problems at the frontier. But behind the frontier I predict rolling model commoditisation as domain-specific systems are built on second-tier models to solve problems that were frontier 12 months ago. That is what I think Microsoft is positioning for.
Another counterargument is that models trained on data nobody else has access to are likely to outperform with regard to that dataset/use case. I don’t think this argues against commoditisation of models in general, but it is an argument in favour of companies like Meta and Google driving increasing performance out of their legacy businesses.
A third counterargument is that vertical integration might be key to optimising model economics. Complex queries are exponentially more expensive to run than simple ones, so a small fraction of super-users can dominate the cost of running a model. This means flat pricing does not work and even consumption-based pricing doesn’t perfectly solve the issue since inference pricing depends on datacentre utilisation: a complex task run at a slow time is cheaper than a complex task run at a busy time. Demand prediction is key and adding a distributor between the model and the customer obfuscates the data. I don’t think this kills a multi-model architecture. I find it very hard to believe that Microsoft can’t figure out a pricing model that produces stable average margins across its huge customer base. Indeed, they have already begun to introduce consumption-based pricing, which will help with this.
Finally, Mark Zuckerberg argues that for competitive and safety reasons, frontier AI won’t always be fully available via APIs. I can believe that. But I think second-tier models are likely to remain available via API, at least to grade-A distributors like Microsoft who can ensure safety and provide a route to market. In fact, as second-tier models get more powerful, I could imagine strict regulation of how they are distributed, which would benefit scale players like Microsoft.
The endgame may be commoditisation even in an AGI scenario. AGI should reduce the cost of developing frontier models. In addition, compute is only going to get cheaper. That argues for more and cheaper frontier models, not fewer and dearer. Cheaper frontier models would lead to a massive expansion in TAM for the industry, and probably excellent economics for the distributors.
OpenAI
What does commoditisation mean for OpenAI? Microsoft’s contractual relationship with OpenAI keeps being weakened, so I don’t ascribe much value to it. However Microsoft does have a strategic interest in keeping multiple model providers alive, and it has an equity stake in OpenAI.
OpenAI’s frontier model performance is, broadly, matched by its main competitors. ChatGPT has a huge consumer user base (>800m) but engagement is low, there is no network effect, and switching costs for AI chat assistants are nil (I’ve personally gone from ChatGPT to Grok to Copilot to Claude and halfway back to Copilot). Unlike Meta and Google, OpenAI has no legacy business on which to unleash its LLMs to drive productivity and cash flow. OpenAI does have growing enterprise revenues but appears to be behind Anthropic, and anyway both have to compete with Google which can afford to subsidise its Gemini LLM aggressively, as it demonstrated when it reportedly undercut Anthropic when Apple was looking for a model to power Siri. OpenAI is burning cash fast as it invests to advance the frontier, and even on its own projections won’t achieve cash flow breakeven before 2030. Rolling commoditisation may prevent independent frontier labs like OpenAI and Anthropic from selling legacy models at prices that would support continued investment in their frontier models. If this is the case, they will have to generate a return on each frontier model in the year or two before its capabilities are commoditised.
None of this reads well. But big cash generative firms like Microsoft, Amazon, and Nividia have two good reasons to keep OpenAI alive: they benefit from advancing the AI frontier because this expands their TAM, and they benefit from preventing Google developing a vertically-integrated monopoly. Arguably, it is better that they share the costs than that each bears them alone. My base case is that they will keep OpenAI alive, at least for now.
In the long run I see three possible outcomes:
There is a positive return on developing frontier LLMs, in which case OpenAI will survive and Microsoft’s equity will be worth something.
There is a negative return on developing frontier models, in which case OpenAI will die and Microsoft will buy it, shut down the frontier effort, and sell OpenAI’s legacy models to customers at strong margins.
Only Google and maybe Meta can earn a return on frontier models, by deploying them to improve their legacy businesses. I think this is unlikely - if there is a return on developing frontier models it will apply across the economy and thousands of companies will pay for it - but if it happens Google becomes an AI monopoly and gets broken up.
Conclusion on LLMs
While frontier models will remain an oligopoly with capital as a barrier to entry, trailing models will commoditise. As the frontier moves forward, more and more use cases will be addressable by commoditised models. This rolling wave of commoditisation will be wonderful for customers but also potentially very good for distributors. Microsoft will distribute these commoditised models by allowing customers to build bespoke apps and agents on them and by bundling cheap intelligence into existing services. Customers will choose Microsoft over independent labs to avoid getting locked into a single model vendor and to benefit from using the right model for each specific task. As commoditised models improve, Microsoft’s TAM will increase. Eventually every process in every enterprise will be addressable.
Agents
This section is about everything Microsoft does to make agents possible, including its cloud platform, productivity suite, vast data lake, proprietary LLMs, and finally the agents themselves.
Why am I lumping all of these things together? Because while several of them are powerful standalone businesses in their own right, I think it will become increasingly clear that they all exist to support agents. In the AI era agents will drive productivity, agents will expand the TAM, and agents are where the money is.
Cheap models make powerful agents economically viable. As the AI frontier rolls forward agents will be able to do more and more complex tasks. In addition as silicon advances and datacentres become more efficient, cost per token will continue to fall. This combination - more powerful agents that cost less - will cause an explosion in use cases. To capture this opportunity, Microsoft has developed products to address every stage in the process of creating, deploying, using, and managing an agent:
Azure is Microsoft’s cloud platform and the foundational software infrastructure layer. It provides the global cloud infrastructure, security, identity, and scalable compute on which every other product in the agentic stack runs.
OneLake stores enterprise business and operational data, natively or via mirroring. Fabric ingests, transforms, analyses, and models this data to surface business-ready insights. Fabric IQ adds business semantics so te data has consistent meaning. Foundry IQ turns scattered enterprise content into usable data for agents. Work IQ uses Microsoft 365 emails, meetings, and documents to help agents understand a user’s role, relationships, work patterns, and how their organisation actually operates.
Foundry lets developers build, deploy, and govern code-heavy agents. Copilot Studio lets users build low-code agents that live in Microsoft 365 apps. Agent 365 lets organizations register, manage, secure, and monitor agents from any source.
Microsoft 365 Copilot embeds agents directly into Word, Excel, Outlook, Teams, and other productivity tools.
Microsoft Agent Factory is a program (and procurement bundle) that brings Work IQ, Fabric IQ, and Foundry IQ together under a single metered plan, helping organizations move quickly from experimentation to deployment.
One important caveat: the customer, not Microsoft, owns the data in OneLake and Work IQ. Microsoft does not use it for training models or agents that it makes available to other customers. The data can only be used to train models and agents for that specific customer, with that customer’s permission. Nonetheless Microsoft’s agent stack is immensely powerful, combining infrastructure, data, data preparation and analytics, human and organisational insights, tools for building simple and complex agents, surfaces on which to deploy them, and tools to control and manage them.
Microsoft is building two business models on this platform:
The platform, sold to customers to create their own agents.
Proprietary agents, sold to customers directly.
Customers building their own agents will be a huge business. As a matter of competitive necessity, every enterprise will apply agentic AI to every process to drive efficiency. Each enterprise has specific use cases and ways of working, so they will develop specific agents that suit their needs.
Building proprietary agents is potentially transformational. Microsoft has always been a horizontal software company. It builds software that can be used by all enterprises across all sectors. Outlook, Excel, security - all enterprises need these. The first thing proprietary agents can do is transform the capabilities of the existing horizontal productivity suite. By adding AI chat to Windows and agents into major applications like Excel, Microsoft increases the value-add and pricing power of its legacy products.
The second thing proprietary agents can do is help Microsoft go vertical. By building agents that automate sector-specific workflows, Microsoft can build large vertical software businesses. Microsoft Copilot inside Excel now does AI-powered DCF modeling and financial statement parsing. Copilot inside Word does contract review and case law research. Developing vertical software no longer requires domain engineers and years of development; it now requires an agent and a customer relationship. Microsoft has both.
To date Microsoft’s proprietary agents don’t have a great reputation. For example, Copilot chat has a reputation for hallucinating. I think this has caused analysts to underestimate the potential of this business. To date, Copilot has run on OpenAI’s GPT series of models. These are highly competitive but Copilot underperforms because Microsoft has limited its context window (presumably to control costs).
Three things will make Microsoft’s proprietary agents vastly better. The first is Microsoft’s move to a seat + consumption pricing model, which will allow Microsoft to charge more for users who need better performance. The second is cheaper inference. And the third is the rolling commoditisation of LLMs, which will be able to address exponentially more use cases over time.
As Satya Nadella said on the 3q26 earnings call: Agent Mode in Excel “sort of didn’t work until it started working” because the model got better. Many use cases will follow this pattern.
Competition in agents
Microsoft is not the only company offering an agent platform or proprietary agents. At minimum, it will have to compete with the other hyperscalers plus Anthropic and OpenAI, and a host of single-product startups building agents for specific uses. I think Microsoft is well-equipped to compete, with advantages that include:
Distribution. Microsoft can bundle agents into existing subscriptions and products. More on this in Part III.
Trust. For decades, enterprise IT departments have been built around Microsoft’s data privacy, permissioning, and security architectures. This is not an unassailable moat, but it will take time for each competitor to build trust.
Data. With a customer’s permission, Microsoft can build agents that know who you are, what you have worked on, who you interact with, what permissions you have, and how your organisation works.
Scale. Microsoft can leverage the cost of developing and maintaining its agent platform and its proprietary agents across thousands of existing customers, globally.
Switching costs. Agents reduce these initially, by handling migration tasks that used to be a headache. But agents that handle workflows will become deeply embedded in how each organisation operates. Once large numbers of agents are working together to coordinate and execute complex tasks, switching costs will be significant. Customers know this, which is an incentive to work with Microsoft rather than getting locked into Anthropic and OpenAI (which may not stay solvent, let alone with the LLM race).
Despite this, customers will likely build agents on Microsoft’s platform and experiment with other vendors. Some of these new relationships might become sticky, or Microsoft might win these customers back as its proprietary models and agents improve. Either way, the market is big enough for several players.
Overall, I think that as the infrastructure and LLM layers commoditise, value will shift to “compound systems” like Microsoft’s agent platform.
Economics
If that is right, then the economics are very attractive.
For starters, the market is huge. As agents do more of the work that humans can do, enterprises will devote more of their operating budget to tech spending. Agentic efficiency will also help them develop new products and services, growing revenue and boosting spending power. To get an idea of the scale, consider Microsoft 365 Copilot. Microsoft 365 has 450m customers. Over 20m of them pay for M365 Copilot. That number rose 35% in the last quarter. Paid M365 Copilot costs $30 per month. That makes it a $7bn product already. As it gets more powerful, I expect adoption and possibly pricing to rise. M365 Copilot alone could theoretically be a $50-100bn product. That’s just the general, horizontal Copilot. There are others for specific verticals, and that’s before you consider all the other agents being built on Microsoft’s platform.
If the competitive arguments above are right then margins should be strong. The platform itself has some of the economic characteristics of legacy software. Each new customer involves additional compute costs, which can be covered by consumption-based pricing. However there is no additional software development cost for each new customer. Like Windows and Office, the software component of Azure, Foundry etc. can be sold millions of times at near 100% gross margins. The cost of coding is falling fast, but that’s been true for decades; code has never been Microsoft’s competitive advantage. As long as there are some remaining advantages around distribution, trust, data, scale, and switching costs, margins should be attractive.
On the proprietary agent side, the same applies: the same advantages, and the same low marginal cost of the software component. Some of the more complex agents may be able to achieve outcome-based pricing, capturing a portion of the value they deliver rather than a fixed price. In addition, the cost to run agents will fall as models commoditise and compute gets cheaper. All of this suggests strong margins.
One concern is that Microsoft does not have a frontier LLM itself and therefore has to pay external providers for frontier intelligence. The benefit of this is that it doesn’t have to fund the essentially speculative development of frontier models. The downside is that once the OpenAI deal expires in 2032, Microsoft might have to pay market price for frontier intelligence and its gross margins might fall. But by then, I think Microsoft will have built an immense business distributing commoditised models. These will be hugely capable and cheap, allowing high gross margins for a distributor with a moat. In addition, Microsoft is building proprietary models optimised for specific large use cases. These are cheap to develop and efficient to run, and will give Microsoft full-stack economics in specific verticals.
Another concern is that for decades, Microsoft has sold its products on a per-seat basis. If one agent can do the work of several people, customers will need fewer “seats”. But if agents make humans more productive enterprises might employ more of them, not fewer (Jevons Paradox). Second, Microsoft is already migrating to a seat + consumption model, and agents will consume a lot.
On the 3q26 results call, Microsoft’s CFO Amy Hood said that AI margins are better now than they were at the same stage of the cloud transition. She implied that this isn’t widely understood, and that consumption-based pricing will improve margins. If she’s right, these statements have enormous importance.
Conclusion on agents
Agents vastly expand Microsoft’s TAM. Microsoft offers a full stack for third parties to build their own agents without locking themselves into one model provider. It is also developing its proprietary agents which improve the value offered by its productivity suite and allow it to go much deeper into large verticals where it will own the full stack. As commoditised LLMs get more and more powerful, Microsoft will be able to offer better and better proprietary agents at strong margins.
Agents are Microsoft’s future. In Part III, we will examine Microsoft’s superpower: how it will distribute them. We will also look at the competitive advantages of combining different layers of the AI stack, and briefly discuss valuation.
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
sounds like Microsoft is doing a little bit what Palantir is doing.