Copilot? Isn’t that the app I make spreadsheets in? No… wait… it’s my email app right?

Microsoft's problem, I think, is in significant part that they are the big commercial player trying for a local AI play. Like, your local Windows machine does AI inference. In Anthropic's business model, the inference is cloud-based.

Local is more hardware-intensive, because the capacity utilization of hardware is going to be lower for local AI. If you stick a piece of hardware in a datacenter and lots of people share it, you need less hardware, because when one person isn't using that hardware, another can be. If you would use local AI hardware 1% of the time, then it costs only about one-hundredth the amount from a hardware standpoint to have people sharing parallel compute hardware in a datacenter as to do everything locally. So as long as hardware prices, like shortages of memory, are a constraining factor (or cooling, for that matter, or maybe power if you're talking about laptops on battery, all of which have cloud-based approaches getting an advantage) Microsoft's going to have a harder time of it than the cloud guys.

Microsoft (and local AI in general) does better if people really want low-latency or always-doing-work load, or reliably always-available services, or services where data privacy is critical. There, local AI has the advantage over cloud-based or at least erodes the cloud-based advantage. Right now, I think that that's just not generally where the state of affairs is. Could change in the future, but I think that they're just going to have a hard time of things in the near term. My guess is that Microsoft's relative potential improves as memory prices come back down.

I think that running local LLMs would be great. But the simple fact is that for most users, it's just too costly to make sense for a lot of applications with current memory prices.

I got a Framework Desktop, 128GB, specifically to do local generative AI stuff, in 2025. At the time, the system was $2,500, which is already going to be pricey for a number of people for a single-purpose computer. In the months since that shipped, the price on the exact same hardware configuration has gone up to over $6,500. That's just not a price that a lot of people are going to be willing to pay for a PC. If component supply rises and prices drop back down, then I think that the calculus changes for local AI.

AI companies are acquiring more memory than the entire rest of the world uses. If we want to do the same thing that we could do in the cloud locally and have capacity utilization of 1% on that hardware, then we need a hundred times as much memory as we do with a cloud-based compute approach. That's...a kind of staggering number.

EDIT: Oh, one major exception that could favor local compute, if R&D produces some major improvements in this direction. Right now, the way most models work, we have a very-expensive-to-generate model that's static, unchanging. This model does not change as one does inference on it. This makes it fairly efficient for a single compute node in an AI datacenter somewhere to have this one model loaded onto it, and then be used by many users who are all wanting to use the same model.

However, if the same model isn't being used by many users, then the (expensive) cost of loading a new model onto memory attached to parallel compute hardware for each prompt has to be paid.

If someone comes up with major improvements that can be derived from mutating a model, from updating it as it is used --- and I would say with some confidence that human-level AI will require some mechanism giving it more ability to learn after the initial training phase is complete than is presently the case for LLMs as used in 2026 --- this is something that may greatly shift the balance in favor of local AI. It's something that we, as humans, do.

If I were Microsoft, I might seriously look into advanced AI R&D, where models are updated during use. It's not just that it's an area with potential, but that it's an area that might advantage Microsoft's strengths (control of the local computer, doing local compute) relative to other major AI companies. It only takes one major breakthrough that makes everyone very badly want to have an constantly-updated model to drastically alter the economics of AI compute.

Though...if that happens, as I point out above, that'll likely set off an even greater RAM crisis than happened with cloud AI compute...