Microsoft FY 3Q'26: Multi-Model Mirage, Copilot Momentum
Programming Note: While I typically don’t write on company specific topics on Sunday, I’m making an exception during the earnings season.
Microsoft’s own narrative about their future has evolved quite a bit following the ChatGPT moment in 2022. While they used to highlight how their exclusive access to OpenAI models would give them a decisive competitive advantage in attracting customers, they have noticeably shifted the tone to emphasize the importance of having a multi-model environment for customers. Notice the following from their FY 3Q’26 earnings call:
Over 10,000 customers have used more than one model on Foundry. 5,000 have used open source models, and the number who have used Anthropic and OpenAI models increased 2x quarter-over-quarter.
Microsoft hasn’t quite spelled it out clearly for investors, but once you triangulate these data points with previous disclosures, a more interesting picture emerges.
During the FY 2Q’26 call, Microsoft actually disclosed that ~1,500 customers were using both OpenAI and Anthropic models. Since they mentioned that such customers have doubled QoQ, we can estimate that the number has reached 3k customers now. That provides better context for understanding that more customers are using at least one open-source model than are exclusively using OpenAI and Anthropic’s models.
Perhaps the more interesting takeaway that’s (deliberately?) not clear from Microsoft’s commentary during the call is that the vast majority of their Foundry customers are still likely using a model from just one company instead of adopting the multi-model environment Microsoft wishes customers would adopt. Microsoft hasn’t disclosed their total number of Foundry customers during the call, but they did mention in the FY 1Q’26 call that Azure AI Foundry had 80k customers, which is ~80% of the Fortune 500. If only 10k customers used more than one model last quarter, that likely means ~90% of customers are still using just one model. Maybe multi-model adoption will gain more momentum over time, but I wanted to highlight that the data indicates Microsoft’s own customers are still far away from such a world.
While both Microsoft and Amazon propagate the value of a multi-model environment, they should perhaps have more appreciation for the appeal of standardization on a single stack, since they themselves were large beneficiaries of such a reality in the earlier cloud era. Everyone similarly argued that enterprises must be multi-cloud, yet most CIOs rationally pay the lock-in premium and standardize, using a second cloud only for narrow specialty workloads. The lock-in risk is abstract and deferred, but since the operational tax of multi-cloud is concrete and recurring, there is a gravitational pull to standardize your IT stack mostly on a single cloud. I suspect most customers may gravitate to a single frontier model for the same reason they gravitated to a single cloud, even though you can make a compelling argument for why it can be quite risky for an enterprise to do so over the long term. Microsoft wants the harness layer to be decoupled from the model layer and would love their customers to have a multi-model environment, but the AI labs not only have little interest in being commoditized but are also clearly indicating that it will be harder and harder to untangle the harness layer from the model layer. Notice the following exchange between Ben Thompson and Sam Altman in the recent Stratechery interview:
Ben Thompson (BT): How important is the harness, the runtime around the model, the tools, state — to your point, a very important word to you — memory, permissions, evals, to making agents actually work?
Sam Altman (SA): Hard to overstate how critical it is. I no longer think of the harness and the model as these entirely separable things, like my experience of using these, I am very aware of the fact that I don’t always know when I fire something off in Codex and it does an amazing thing for me. I don’t know how much credit —
BT: Was it that the model is amazing or the harness was amazing?
SA: Yeah, exactly.
BT: To what extent is the harness developed in conjunction with the model? Where does that integration happen? Is it in post-training? Is it in the prompt? What makes this integration work?
SA: Both of those. It’s not really part of the pre-training process but I would say you can look at it — there’s a more interesting thing here which is the fact that we’ve seen examples of this many times in the past of where things that we thought were very separable get baked in more and more and more. Like the way we initially thought about tool-calling, which is now a critical part of how we use these models, was not something that we thought about deeply integrating into the training process and over time we’ve done more and more of that.
I would also suspect that model and harness come together more over time and I would for that matter, I would expect that pre-training and post-training eventually come together more over time as well. It’s such a cliché to say, but I’ll do it anyway, because I think it’s very, very true — we’re so early in the paradigm of all of this, this is still like the Homebrew Computer Club days of how much this is like really matured as an industry.
BT: This is why I think so interesting, I wrote about this a few weeks ago, in any value chain, ultimately a point of integration emerges that that’s where it’s really important, these two pieces have to go together to make it work. And over time, that’s obviously where a lot of value collects — my thesis then is that this harness-model integration is the key point. It’s to your interest, but it sounds like you agree.
SA: It is to my interest, I do agree, but I also would say even more broadly, what you care about is that you go type into Codex what you want to happen and that it happens.
BT: You don’t care about the implementation details.
SA: I don’t think you do.
The more pressing matter for Microsoft is their own cash cow software business, which is why they continue to highlight that they expect to balance the incoming supply of compute between Azure and their own first-party workloads. Microsoft doesn’t have a frontier model of their own, but Satya Nadella highlighted that they have access to OpenAI’s royalty-free frontier model with all the IP rights until 2032. The implicit assumption is that by the time those rights expire, Microsoft will be able to build their own frontier model. I’m not sure how much credence or confidence investors should have in Microsoft’s ability to build a frontier model on their own when they noticeably lag behind today’s frontier models despite having such unfettered access to OpenAI’s IP. Nonetheless, it does appear that Copilot has started to see noticeable momentum. From the call (emphasis mine):
We have seen a surge in usage of our first-party agents with monthly active usage up 6x year-to-date. Copilot queries per user were up nearly 20% quarter-over-quarter. To put this momentum in perspective, weekly engagement is now at the same level as Outlook, as more and more users make Copilot a habit.
Copilot engagement at the same level as Outlook? Now that’s pretty impressive. Moreover, while investors worry about the seat-based subscription model in a world of agents, Microsoft reassured investors that the transition won’t be too messy for them. From the call (emphasis mine):
“customers want predictability, especially for budgets and procurement. And the seat-based pricing is just entitlements to some consumption, right? And so that’s, I think, the way to think about it, which is, there is some base usage rights that get bundled in or packaged in to seats, it’s a convenient way for people to buy some essentially consumption packs that happen to be assigned to seats or agents.
And then beyond a certain level, there’s overages that go into pure consumption. And even there, if you have commitments, long-term commitments to consumption, you get discounting that is appropriate with it. So I feel like that’s the direction of travel.
And then -- the other thing you mentioned is how is this going to -- from a customer perspective, they’re going to evaluate it by evals. Where are they seeing the value of tokens, as simple as that. So where they see the outcome, the eval and the token, whether it’s improving revenue, improving efficiency, and that’s what will refine. Like when we talk about IT budgets, IT budgets are going to have to be reshaped by a combination of business outcomes, making their way into IT budgets and maybe reallocation from other line items on the income statement like OpEx.”
One concern I do have about enterprise AI revenue momentum in 2026 is that the current tokenmaxxing era will inevitably go through a period of optimization once the hype settles. Today, everyone is still figuring out how to embed AI into their workflows or how to create entirely new workflows that are only possible due to AI, but as the early phase of experimentation winds down, enterprises will look much more closely at how to optimize token spending. How such optimization will affect revenue growth ambitions for AI labs remains to be seen. However, even though Nadella thinks enterprise customers will allocate IT budgets based on the revenue or cost improvements they see, I suspect the reality may be that when everyone has access to the same tools, those customers may not see much sustained improvement in a relative sense. That is, of course, not to say enterprise customers won’t see value from using AI, but sustained value is unlikely to accrue to customers who all have access to the same tools. The end result might just be lower margins for everyone as enterprises feel forced to hand their margins to the machine gods just to keep up with their competitors.
Subscribers get the daily journal and five+ years of Deep Dives, i.e. full-length analyses with financial models on 65+ companies. The daily is just how I think out loud between the Deep Dives!
Current Portfolio:
Please note that these are NOT my recommendation to buy/sell these securities, but just disclosure from my end so that you can assess potential biases that I may have because of my own personal portfolio holdings. Always consider my write-up my personal investing journal and never forget my objectives, risk tolerance, and constraints may have no resemblance to yours.
My current portfolio is disclosed below: