Apple’s AI Advantage Is the Operating System, Not the Model

Alex Kantrowitz’s case for Apple before WWDC starts with a constraint on the panic. Apple has not yet paid a visible business price for being behind in AI. The company, he noted, is still an “iPhone selling machine” before any successful AI rollout: more than $140 billion in its first quarter, $111 billion in the most recent quarter, and a stock up 56% over the prior year. There is no successful AI-native consumer device yet that has displaced the phone. For now, Apple’s AI weakness has not become an Apple business crisis.

Ranjan Roy agreed that the business remains healthy in the most basic sense: Apple keeps generating “ungodly amounts of revenue” and keeps users locked into the ecosystem. But he separated that from product vitality. Apple has done this, he argued, without “genuine innovation on the actual product side” in years. Kantrowitz pushed back only partway. Under Tim Cook, he said, Apple has been a refiner rather than a revolutionary: better cameras, better battery life, more computing power, better devices. But the revolutionary AI computing device has not arrived from OpenAI, Meta, Google, Amazon, or anyone else.

That absence is the opening. If the AI device is still undefined, the iPhone itself may become the AI device — not through a new form factor, but through integration into the interface users already inhabit.

The reported WWDC plan, as Kantrowitz described it from Mark Gurman’s reporting, centers on a new Siri known internally as Project Campo. The goal is to move Siri from a voice-control system into a “do-it-all AI companion” capable of handling tasks across iOS, iPadOS, and macOS, including both Apple apps and third-party apps. The new Siri is reportedly powered by a Gemini model from Google as part of a billion-dollar agreement, with much of the system hosted on Google servers — an arrangement Kantrowitz said may raise privacy questions given Apple’s long-running emphasis on safeguarding user data.

The more important part is not that Apple has its own frontier model. It is that Apple controls the mobile operating system. The reported interface change would let users swipe down from the top center of the iPhone to open a new “Search or Ask” experience. From there, users could launch apps, start text messages, ask about the weather, add calendar appointments, search notes, trigger shortcuts inside apps, or search the web using AI. Voice remains an option, but the interface is built around getting things done by typing as well.

For Kantrowitz, that is the right shape. Meta may want to build “the best personal agents,” as Alexander Wang said at a Bloomberg conference, but Meta’s agent still requires a user to open an app on an iPhone. Apple’s agent can be one swipe away from the operating system itself.

Roy called the proposed functionality “the most basic AI stuff imaginable” and, more pointedly, “the most basic user experience stuff imaginable.” Dictation, AI search, and app-adjacent assistance already exist for many users in other tools. He uses Whisper Flow because Apple’s native dictation is bad, and he already does much of his search through AI. But he accepted the larger distribution point: many users still do not use AI in an integrated way, and Apple has a rare opportunity to make AI ambient simply by putting it where users already are.

The issue is whether the basicness is a weakness or a discipline. Roy suggested it may be good if Apple is now announcing things it can actually ship rather than advertising an idealized future too early. Kantrowitz’s version was similar: this is less ambitious than the original Apple Intelligence vision, but it appears better matched to what current models can actually do.

The core disagreement over Apple was not whether the reported Siri design makes strategic sense. It was whether Apple can build it.

Ranjan Roy’s skepticism came from the practical complexity of agentic systemsAI systems designed to take actions across tools rather than only generate text.. Having spent the past year and a half building such systems at Writer, he argued that accessing “any app on your phone” and routing across data from many systems is difficult and unpredictable. It is one thing for a model to write an email from a prompt. It is another for Siri to know which app owns the relevant data, whether a user’s calendar is Apple Calendar or Google Calendar, whether email lives in Mail or Gmail, and what tool should be called for a given task.

That routing problem becomes harder as the universe of possible actions expands. Roy asked whether Apple would launch narrowly — email and calendar appointments on day one — or whether Siri would be expected to pull up a YouTube video by voice prompt, coordinate across third-party apps, and work across a user’s whole phone.

Alex Kantrowitz drew a distinction between the ambition and the proof. Apple is reportedly planning to let users issue multiple commands to Siri at once, such as checking the weather, creating a calendar appointment, and sending a message in one prompt. That, he said, matches the current capabilities of systems like ChatGPT and Claude, which can handle compound requests and maintain enough context to route through a series of tasks. The ambition is no longer fantasy. But when Roy asked whether Apple could pull it off, Kantrowitz’s answer was blunt.

Roy accepted that as the right standard. Any company of Apple’s scale and history should be able to execute this kind of product. But Apple’s own record with Siri and Apple Intelligence makes confidence difficult. Roy said he wanted to believe in Apple, but had been disappointed too many times. He pointed back to the Bella Ramsey Apple Intelligence ad, which implied a level of AI usefulness that the product did not deliver. In retrospect, he said, that rollout suggested Apple “fundamentally misunderstood the actual technology” at the time.

Kantrowitz’s more favorable read was that Apple has made progress because it now understands the technology better. It is building toward current model capabilities rather than toward a speculative dream produced in Cupertino. It may also have Google deeply involved. Because the new Siri reportedly relies on Gemini, Kantrowitz argued Google has reputational reasons to make sure it works: Gemini’s name and technology are implicated.

Roy complicated that point. Google already integrates assistant-style AI across Android, and even Alexa Plus on an Echo Show can do much of what Apple is reportedly preparing. From a technology standpoint, this should be table stakes. But Google’s incentives are mixed. If the new Siri becomes a defining iPhone feature, Google is strengthening a hardware rival. Kantrowitz countered that Google still gains leverage. If Apple depends on Gemini-derived systems, each future Google model gives Google renewed power in the relationship. Apple may want to distill from the next stronger Gemini model, and the one after that. That makes Google both supplier and strategic constraint.

The most troubling reported detail was Apple’s internal labeling of the new Siri as a “beta” and “preview,” with the possibility of a waitlist for new features. Roy said he preferred honest caution to a glossy marketing campaign that suggests everything is finished. But the line gave him “PTSD” from the first Apple Intelligence launch, which he called one of the worst product rollouts he could remember without claiming it was the worst in history. Kantrowitz said the beta-and-waitlist detail made him less excited.

His summary had three parts. First, AI does not matter to Apple’s business as much in the short term as the market narrative suggests; absent a true AI device, Apple is safe for now. Second, the reported WWDC direction is positive because Apple appears to be designing for what the technology can actually do. Third, the execution risk remains large, and the possibility of a waitlist is a warning sign.

Ranjan Roy floated a “Dutch disease” analogy for Apple: a resource curse in which a dominant asset distorts the ability to develop other parts of an economy. Apple’s natural resource is the iPhone ecosystem. Because the iPhone keeps selling and users keep buying adjacent products, the company may lack the internal urgency to build fundamentally new capabilities.

Alex Kantrowitz resisted making that the whole explanation. Other big tech companies have natural resources too. Google has search, which Roy called “the greatest natural monopoly of all time,” but Google has invested in AI for a decade and has recently reasserted itself. The difference, Kantrowitz argued, is that strong incumbents can stay on top if they continue to innovate rather than sit on their advantages.

For Kantrowitz, Apple’s deeper issue is cultural. He described Apple as “very high on its own supply.” The company believes its culture produced the iPhone, and while that is partly true, the iPhone also came from Steve Jobs. Apple has continued as a refined Jobs-era culture without Jobs himself, and Kantrowitz argued that pride in its way of doing things has made it reluctant to try other models.

Roy joked that Google’s AI transformation could be traced to Sundar Pichai’s McKinsey background and an orientation toward organizational design and change management. Kantrowitz took the organizational point seriously. Google’s model — a central AI engine room working with product areas like Gmail, Maps, and Search — may help improve specific products. But building a unified assistant is different. If the future is a “super app” or operating-layer agent, product groups may have to subordinate their own priorities to the central assistant. That is hard in any large company where individual products want users to spend time inside them.

Roy pointed to one reported Siri feature as especially significant: a dedicated Siri app for conversations and continuing past chats, effectively a ChatGPT-style assistant inside Apple’s ecosystem. He asked whether Siri becomes Apple’s super app one or two years from now. Kantrowitz said that is likely Apple’s ambition, though he added that iOS is already a super app in a sense. The strategic question is whether the chatbot or AI agent becomes the operating system. If it does, Apple needs an answer at the operating-system level, not merely another app.

Anthropic’s blog post, shown in the source under the headline “When AI builds itself,” argued that AI-assisted engineering is accelerating. Alex Kantrowitz cited the company’s claim that Anthropic engineers now ship eight times as much code per quarter as they did from 2021 to 2025, with the help of AI models such as Mythos. The same material warned that rare model misalignment could compound as models help build their successors, becoming more frequent and less understood until humans lose control.

Anthropic also caveated its own metric in the on-screen text. Lines of code are an imperfect measure of productivity because they capture quantity rather than quality. The company said the 8x figure is almost certainly an overstatement of true productivity gains, but still indicates acceleration. In a March 2026 poll of 130 Anthropic research employees, the median respondent estimated they produced around four times as much output with Mythos Preview as they would have without AI models on the same kinds of projects.

Kantrowitz summarized the argument this way: AI has made engineers much more productive; humans are still in the loop; eventually they may not need to be; and that possibility creates risk. Ranjan Roy’s reaction was exasperation. He sees the recursive-self-improvement warning as part of a repeated communications pattern from Anthropic: the company stresses that its models may pose extraordinary risks while, in Roy’s view, still moving aggressively toward scale, enterprise deployment, and a possible public-market outcome.

Roy did not dispute that AI coding tools have improved. When Kantrowitz asked whether Anthropic’s coding tools have gotten much better and enabled people to do more, Roy said yes over the past six months, though he called the 4.8 release “a pretty big dud.” His objection was not to the existence of progress. It was to the combination of danger rhetoric and full-speed commercialization. If Anthropic truly believes the technology poses significant societal risk, Roy asked, why raise enormous sums, push toward an IPO, and publicize the danger? He framed the next few months as a race to see who gets out first among SpaceX, Anthropic, and OpenAI, and treated Anthropic’s messaging in that context.

Kantrowitz offered the strongest version of Anthropic’s defense. If a company believes the technology is dangerous, it may still push forward because only a leading lab has influence over deployment. A lagging model company warning the frontier to slow down would be ignored. A leader can shape norms, get governments and industry to listen, and make safety demands from a position of relevance. On that logic, staying at the frontier is not hypocrisy but a condition of influence.

Roy called that the “I alone can fix it” argument. He said OpenAI and Anthropic have used versions of the same narrative from the beginning: the technology is incredibly dangerous, but they must continue advancing it because they are the responsible actors. That narrative, he argued, has helped them tremendously. He contrasted it with Google’s posture. Sundar Pichai has described AI as fire, but Roy said he has not heard Sundar use the same constant existential-risk framing while racing ahead.

Kantrowitz connected the issue to a point he said Geoffrey Hinton had made in a prior interview on the same feed: whether Anthropic can maintain a safety mission if it becomes a public company. As Kantrowitz recounted Hinton’s view, after an IPO a company has a legal obligation to shareholders to maximize profit, not AI safety. Kantrowitz struggled to see how those obligations continue to point in the same direction.

Roy replied that a company also has an obligation not to destroy humanity, calling his own phrasing hyperbolic but insisting the risk framing cuts both ways. If the technology truly poses significant societal risks, investors should view that as a risk too. He also argued that being public could force more transparency around the real stakes.

Kantrowitz pressed with a concrete scenario. Suppose Anthropic is public, has Mythos, and OpenAI releases a dynamic coding model. Anthropic might believe releasing Mythos would increase cyber risk but also help it grow faster and compete. If it holds Mythos back, could shareholders sue? Kantrowitz’s point was that public-market pressure makes restraint harder, even in cases where risk is not framed as the end of the world but as a serious increase in cyber danger.

Roy was not persuaded that shareholders would sue a company for declining to release a model it had itself marketed as posing national-security-level risks, especially if a competitor was being careless. He still views Anthropic’s public warnings largely as marketing from a capitalist company. But he also accepted that if risks are real and material, the company would have to act with safety in mind.

The strongest reconciliation came from a quote Kantrowitz read from an on-screen excerpt attributed to Ethan Mollick, a professor at the University of Pennsylvania’s Wharton School. Mollick said that while some Anthropic critics see fluff and marketing in the company’s safety pronouncements, many inside are “true believers.” He described AI labs as mixtures of a trillion-dollar company apparatus — marketing teams, lawyers, normal corporate machinery — with researchers building the next models and “philosopher kings” worried about the future. Those groups, Mollick said, are in conflict with one another at times.

That explanation changed Roy’s tone. It gave him a way to reconcile the contradictions he had been pointing at for months: Anthropic can be both a company with sophisticated marketing and a place where some employees genuinely believe they are operating a technology that could pose unprecedented risks.

The Microsoft-OpenAI relationship, in Alex Kantrowitz’s framing, has shifted from partnership into direct competition. Microsoft’s Build messaging emphasized that the company is not simply wedded to OpenAI and intends to build serious AI capabilities of its own. Kantrowitz cited an on-screen excerpt from The Verge characterizing the relationship as a “drama-filled marriage” that became a “situationship” and effectively separated in late April, while Microsoft remains OpenAI’s primary cloud partner “for now.”

The clearest statement came from Mustafa Suleyman, quoted on screen from The Verge, saying Microsoft wants to prove it can become one of the top four AI labs in the world. Suleyman named the three that matter today as Google DeepMind, OpenAI, and Anthropic, adding that Microsoft is “not one of them at the moment.” Ranjan Roy noted the conspicuous absence of Meta from that list.

Roy’s emphasis was distribution. Microsoft, like Apple, owns enormous end-user real estate. Copilot is present across Office and other Microsoft products, with tens or hundreds of millions of users exposed to it. If Microsoft figures out the product, it could become newly interesting in AI in the same way Google returned to relevance over the past year.

Kantrowitz agreed Microsoft has a chance because it has products, Office distribution, and access to OpenAI’s IP. But he questioned the strategic need to build frontier foundation models. Why should Microsoft try to become a top model lab at all?

Roy’s first answer was organizational: Suleyman was brought in to do that. He is not at Microsoft merely to build another Copilot feature. Roy also mocked Microsoft’s naming sprawl — Microsoft IQ, Work IQ, Fabric IQ, Foundry IQ, Web IQ — as classic Microsoft product packaging. But beneath that, he was skeptical of the foundation-model strategy. In his world, model routing has quickly become dominant: find the right model for the right problem, optimize cost, understand token economics. If that is the direction of the market, Microsoft may be better off letting frontier labs fight while it routes intelligently across models.

Kantrowitz offered one strategic reason Microsoft may not have that luxury. He pointed to Satya Nadella’s comment to Ben Thompson that OpenAI and Anthropic will, over time, build their own infrastructure because “it makes sense.” Nadella said he did not want to allocate all Microsoft compute to one player and that hyperscale infrastructure must be spread around the world and across the United States, not concentrated in a single enormous Texas buildout.

Kantrowitz read that as a warning about future cloud competition. In his inference, if OpenAI builds what he called an AI cloud — servers plus models on top — then Microsoft cannot simply be Azure plus access to other people’s models. It may need its own foundation models to answer a more vertically integrated OpenAI offering. He tied the point to OpenAI having talked about building its own cloud service and to Nadella’s assumption that frontier AI companies will build more of their own infrastructure.

Roy still called that a big if. He said OpenAI’s potential AI-cloud business model has not been widely discussed heading into what he called “hot IPO summer” for Anthropic, OpenAI, SpaceX, and others. But the fact that Nadella was addressing the possibility struck him as significant. “What does Satya know?” Roy asked. Kantrowitz replied that Nadella has reason to know plenty, adding that Microsoft has access to OpenAI’s IP “till 2032.”

The resulting picture is no longer a simple dependency map. Microsoft has powered OpenAI through Azure and benefited from OpenAI’s models. Kantrowitz sees OpenAI as potentially wanting to own more of the infrastructure layer. Microsoft, under Suleyman’s stated ambition, wants to own more of the model layer. Each is moving toward the other’s territory.