Apple Just Settled One Question. Microsoft Opened the Other.

A week of announcements has clarified where everyday AI best lives. The institutional contest is just beginning.

Last week, Apple opened WWDC with Siri AI and the next generation of Apple Intelligence. A few days later, Satya Nadella published an essay on the future of the firm in an AI-driven economy. Reading the second against the first, I realized the two were making the same argument from opposite ends of the user experience. Apple had shown what an AI ecosystem looks like when it serves an individual through a device. Microsoft was describing what the same ecosystem looks like when it serves an institution at scale. The framework I sketched earlier this year, which separated everyday AI from game-changing AI, sharpens under this new evidence. Both layers now have a reference architecture. One of them is, for practical purposes, now settled. The other is not. In today's Dispatch, I work through what Apple and Microsoft have each made clear about the ecosystem era of AI, and what remains contested.

The big picture

Over the weekend, Satya Nadella published an argument about where value lives in an AI-driven economy. The short version: the AI model itself is becoming a commodity. The asset that appreciates is the feedback loop around it. For the first time, organizations can build an iterative loop between human judgment and AI systems, not just use chatbots to make people slightly more productive. The durable asset is not which model a firm picks but the ecosystem that surrounds it: the workflows, domain knowledge, and accumulated context that make a model useful in a specific institutional setting. Nadella calls the two sides human capital and token capital, and he argues that they compound together rather than substitute for one another.

The structural claim underneath the vocabulary matters more than the vocabulary itself. If AI models are becoming a commodity, then the value moves to whoever owns the loop. Whoever owns the loop owns the context, the data, the workflows, and the judgment that makes the model useful. The model is replaceable. The loop is not.

This framing arrived after Apple’s WWDC 2026 keynote, where Apple demonstrated the same insight applied to individuals across their ecosystem of devices: iPhones, iPads, and Macs. Nadella was writing about the future of businesses. Apple had already shown what the same logic looks like when applied to individual consumers.

Apple’s Answer

Apple announced a new family of foundation models, AFM 3, organized around a clear architectural premise: keep the loop on the device, and use the cloud only when the device cannot handle the request alone. The family includes five models. A three-billion-parameter dense model, AFM 3 Core, runs on a wide range of existing hardware. A larger twenty-billion-parameter on-device model, AFM 3 Core Advanced, uses a sparse architecture that activates between one and four billion parameters per request, depending on the complexity of the task. The full model lives in flash storage, with active experts loaded into working memory as needed. A server-based model, AFM 3 Cloud, runs on Apple Silicon inside Apple’s Private Cloud Compute. A diffusion model handles image generation. The most capable cloud model, AFM 3 Cloud Pro, runs on NVIDIA GPUs hosted in Google Cloud, with Private Cloud Compute extended to that infrastructure to preserve the same privacy guarantees.

Above these models sits the new Siri, which Apple is calling Siri AI. The job of Siri AI is orchestration. It decides which model handles which request, based on what the task requires and what the device can do locally. The user does not see this decision. The user sees an assistant that holds the personal context of the device and acts on it.

Two design commitments hold the architecture together. The first is that personal context lives on the device and does not leave it. The information that makes the assistant useful, including messages, calendar, photos, files, and on-screen content, is local. The second is that even when a request goes to the cloud, the data is encrypted in transit and not retained. Private Cloud Compute is not a hosting environment in the conventional sense. It is an attested, cryptographically verifiable infrastructure that Apple has designed to make user data unreachable, even to Apple itself.

Apple’s answer to the question of where the loop lives is therefore concrete. The loop lives on the device. The device holds the personal context. The cloud is a temporary extension of the device for the limited tasks the device cannot perform alone.

The Loop in Practice

I am running the first developer beta of iOS 27 and iPadOS 27 to see what this looks like in operation. The most useful demonstration so far came from a Google Sheet open on my iPad. The sheet held Zillow data for a house we listed for sale: views, saves, showings, days on market, etc. I asked Siri AI to tell me what the data meant.

The response did several things at once. Siri AI interpreted the listing performance against typical patterns at this stage of a sale. It then pulled in context from text messages between my wife and me, where we had discussed expectations and concerns. It also pulled context from a separate thread with our real estate agent. Siri AI recommended next steps consistent with the outcome we had been discussing, and offered to draft a set of follow-up questions and send them to our agent.

The recommendations were genuinely insightful and useful, but they are not the point. The point is the architecture. A single assistant reasoned across a spreadsheet in a third-party app, two separate message threads with two different people, and the underlying intent of the conversation, then proposed a specific action and offered to execute it. None of that personal context left the device. There was nothing to upload, nothing to grant additional permission for, nothing to copy and paste into a prompt.

This is the loop Nadella was describing. It is running on my Apple device.

The Device as Infrastructure

The structural point is easy to miss inside the Apple Intelligence demo. Apple did not solve the context-and-privacy problem by building more data centers. Apple solved it by treating hundreds of millions of devices it had already sold as the infrastructure. The iPhone, iPad, and Mac are not endpoints calling into a centralized intelligence. They are the loop. The data center, for everyday use, is the Apple-made device itself.

The contrast with the current alternatives is sharp. To get a comparable result from ChatGPT, Claude, or Gemini today, the user has to do the integration work: connect documents, grant access to mail and messages, paste in context, and accept that the privacy guarantee is essentially an unverifiable promise not to train on the data. The model is powerful. The loop is not closed. The user is the integration layer.

Apple’s approach reverses that arrangement. The loop is closed by default. The integration is the operating system. The privacy guarantee is architectural rather than contractual. For the everyday tasks that make up the majority of what individuals actually want from AI, summarizing a document, drafting a reply, reasoning across personal context, taking a small action, this is sufficient. It does not need a frontier model or new data centers. It needs a closed loop with the user’s context inside it.

That is what Apple will ship in September.

The Enterprise Question

Apple and Microsoft now agree on the structural insight. The ecosystem is the asset. The model is less critical. Where they differ is which ecosystem they intend to own.

Apple is competing for the individual consumer. The loop runs on the device. The context is personal. The tasks are everyday. This maps onto the framework I have used before, which separates everyday AI, the productivity layer that makes individual work marginally faster and more coherent, from game-changing AI, the systems that reorganize how an institution operates. On the everyday side, the contest is now largely settled. Apple delivers the device, the operating system, the application layer, and the data graph of the user. No other competitor can bring those pieces together.

Microsoft is competing for the firm. Nadella’s essay is a wager that the same ecosystem logic, applied to businesses, will determine who captures the value of game-changing AI: the workflows, judgment, and proprietary knowledge that make an organization differentiated. Microsoft’s position is that Azure, the Office 365 footprint, and the installed base of Windows in the corporate world give it the right starting point to host the enterprise loop. Google is making a similar wager, with Workspace and Gemini occupying the same conceptual ground as Microsoft.

The enterprise contest is not a question of who has the best AI model. It is a question of who can credibly host the institutional loop that makes a model useful inside a specific organization with specific business needs. That question is not yet settled.

The final word

The ecosystem era of AI is here. The model is becoming the commodity. The loop, the context, and the privacy posture around AI models are the durable assets.

On the consumer side, Apple has closed the question. The combination of on-device models, an orchestration layer that picks among them, and a privacy architecture that keeps the context local is the answer to what most individuals want from AI most of the time. It does not require new data centers. The data center is the device.

On the enterprise side, Microsoft and Google are now stating their respective theories of the case. The institutions that will use these systems, including most universities, are still working out which loop they want to live inside and what they are willing to give up to do so. That is a consequential contest, and it will take a while to resolve.

For now, I am waiting on the second developer beta.