Apple Can't Fix Siri but Still Owns AI's Toll Road

The most valuable company on Earth cannot make its voice assistant hold a conversation without cutting people off mid-sentence. That is the headline Apple hoped would never be written, but the evidence is now too abundant to ignore. On March 20, 9to5Mac reported that Apple’s Gemini-powered Siri upgrade “could still arrive this month” — a formulation so hedged it practically concedes the point. The first iOS 26.4 beta shipped in February with zero new Siri features. Internal testing has revealed that the assistant still cuts off users mid-sentence during fast speech and struggles with complex multi-step requests. Worse, the system sometimes falls back to ChatGPT for answers that its own Gemini-trained models should be handling, a debugging symptom that suggests the architecture is not merely unfinished but architecturally confused. Bloomberg’s Mark Gurman has reported that Apple’s research showed ChatGPT outperformed Siri by at least 25 percent in accuracy, a gap that two years of investment and a billion-dollar partnership with Google have failed to close.

Yet here is the paradox that makes this story more than another chapter in Apple’s AI embarrassment: the company that cannot build a competitive AI assistant is on pace to earn more than $1 billion this year from the companies that can. Apple collected nearly $900 million in App Store commissions from generative AI apps in 2025 alone, with monthly fees climbing from $35 million in January to $101 million by August. ChatGPT accounts for roughly 75 percent of that revenue. Every time a user subscribes to ChatGPT Plus through the iOS app, Apple takes 30 percent in the first year and 15 percent thereafter — a tax on the intelligence its own product cannot deliver. OpenAI, Anthropic, Google, and xAI are collectively spending hundreds of billions of dollars to build the AI systems that run on Apple’s platform, and Apple is skimming a billion dollars a year for the privilege of distribution. The biggest moat in artificial intelligence, it turns out, belongs to the company that cannot even ship a working LLM.

Two years late and counting: the architecture that keeps breaking

The story of LLM Siri is a masterclass in how engineering ambition can collide with organizational inertia. Apple first committed publicly to a fundamentally new Siri powered by large language models in mid-2024, when Tim Cook promised that a reimagined assistant would arrive as part of the Apple Intelligence suite. The original target was iOS 19 (later renamed iOS 26 under Apple’s new versioning scheme), with a spring 2025 launch intended to coincide with new hardware. That deadline came and went. Apple’s engineers had attempted a hybrid architecture that bolted LLM capabilities onto Siri’s legacy intent-matching system — the same rigid, rule-based framework that has powered Siri since 2011. The hybrid approach failed catastrophically. Bloomberg reported that roughly one-third of test queries produced errors, hallucinations, or nonsensical responses because the old and new systems fought each other for control of the interaction. A query that triggered Siri’s legacy intent classifier would bypass the LLM entirely, producing the same robotic responses users have complained about for a decade. A query that reached the LLM would sometimes loop back to the legacy system for execution, creating latency spikes that made the assistant feel slower than its competitors.

The January 2026 announcement of the Google Gemini partnership was Apple’s tacit admission that its internal models were not ready. Under the multi-year agreement, Apple will use Google’s 1.2-trillion-parameter Gemini models as the foundation for training its own Apple Foundation Models — the AI backend that powers the new Siri. The deal is structured as a cloud computing contract worth an estimated $5 billion over its lifetime, with Apple paying approximately $1 billion annually for access to Google’s model weights, training infrastructure, and cloud technology. The system will run on Apple’s Private Cloud Compute infrastructure, which processes queries in secure enclaves that Apple says neither it nor Google can access after computation. Privacy is the value proposition that justifies the Rube Goldberg architecture: rather than running Gemini directly, Apple is distilling Google’s models into smaller, Apple-branded models that run partially on-device and partially in Apple’s privacy-preserving cloud. The technical challenge is compressing a frontier model’s capabilities into hardware-constrained packages without losing the conversational fluency that makes ChatGPT and Gemini compelling.

That challenge remains unmet. The iOS 26.4 beta that shipped in February contained none of the promised LLM Siri features. Apple’s engineers have since shifted focus to iOS 26.5, targeted for May, with the full conversational Siri now expected in iOS 27 in September 2026. Some features — “world knowledge” search capabilities and on-screen awareness that lets Siri understand what you are looking at on your phone — may arrive in a late-March iOS 26.4 update if internal quality gates are met. But the marquee capability, a Siri that can hold a genuine conversation and execute multi-step workflows across apps, has been pushed back by at least six months from Apple’s most recent promise. One proprietary calculation illustrates the cost of delay: if Apple’s 2.2 billion active devices each generated just one additional AI-assisted query per day — a conservative estimate given ChatGPT’s 900 million weekly active users — and Apple monetized those queries at even one-tenth the rate of a ChatGPT Plus subscription, the missed revenue opportunity from each month of Siri delay exceeds $500 million. Apple is not just losing a feature race. It is losing the economic value of its own installed base.

Follow the money through Cupertino’s AI toll booth

The financial mechanics of Apple’s AI toll road deserve closer examination because they reveal a business model that may be more durable than any AI assistant. Apple charges developers a standard 30 percent commission on in-app subscription purchases in the first year, dropping to 15 percent for subscribers who renew beyond twelve months. For generative AI apps, this means every ChatGPT Plus subscriber who signs up through the iOS app at $20 per month sends $6 to Apple — or $72 per year — without Apple writing a single line of AI code. OpenAI reportedly has more than 50 million consumer subscribers as of February 2026, alongside 9 million paying business users. If even 30 percent of those consumer subscribers pay through iOS — a plausible estimate given iPhone’s dominant share of premium smartphone users — Apple’s annual take from ChatGPT alone approaches $1 billion.

The economics extend far beyond OpenAI. App Store commission data shows that ChatGPT accounts for approximately 75 percent of generative AI subscription revenue on iOS, with xAI’s Grok contributing about 5 percent and the remainder split among Anthropic’s Claude, Google’s Gemini, Perplexity, and dozens of smaller players. The total addressable market is expanding rapidly. ChatGPT’s weekly active user base has grown from 400 million to 900 million in a single year — more than 10 percent of the global population now uses the service every week — and every new AI app that gains traction on iOS adds another tributary to Apple’s revenue stream. U.S. advertisers are projected to push $57 billion through AI-powered platforms this year, a 63 percent leap, and much of the consumer-facing spend will flow through mobile applications that Apple taxes at the point of purchase.

The Google Gemini deal adds a second revenue dimension that most analysts have overlooked. Apple is paying Google an estimated $1 billion per year to train its foundation models, but the deal is not exclusive — Apple retains its existing relationship with OpenAI, which powers the ChatGPT integration within Siri for complex queries. This means Apple is simultaneously a customer of Google and a distribution partner for Google’s competitor, collecting commissions on ChatGPT subscriptions with one hand while paying for Gemini model access with the other. The net effect is that Apple has positioned itself as the AI industry’s universal intermediary: it profits from every model that ships on iOS, it trains its own models using the best available external technology, and it maintains optionality to switch providers if the competitive landscape shifts. For Google, the deal is worth far more than $5 billion in direct revenue because it positions Gemini as the training backbone for 2.2 billion Apple devices — a distribution footprint that even Google’s own Android ecosystem cannot match for premium users. Fortune reported that the deal represents a strategic loss for OpenAI, which had spent months lobbying to become Apple’s primary AI partner and now finds its most lucrative distribution channel also funding a competitor’s model training.

The toll-road analogy is imperfect in one crucial respect: toll roads are commodities, and Apple’s distribution advantage is anything but. The App Store is the only sanctioned channel for installing applications on more than a billion iPhones worldwide (the EU’s Digital Markets Act notwithstanding, with sideloading adoption hovering near single digits in regulated markets). Any AI company that wants to reach iPhone users must either accept Apple’s commission structure or build a web-only experience that sacrifices push notifications, Siri integration, and the seamless payment infrastructure that drives conversion. For the AI labs racing to acquire paying subscribers — a race that now determines which companies survive the current capital expenditure cycle — Apple’s platform tax is not optional. It is the cost of reaching the world’s highest-spending consumer demographic.

The three cracks in Apple’s moat

The bear case against Apple’s AI toll road is not that it will collapse overnight but that it will erode gradually as the industry’s center of gravity shifts away from mobile app distribution. Three forces could drive that erosion, and each deserves serious scrutiny from anyone who assumes Apple’s current position is permanent.

The first is regulatory action. The European Union’s Digital Markets Act already requires Apple to allow alternative app stores and sideloading on iPhones sold in Europe, and the U.S. Department of Justice’s antitrust case against Apple — filed in March 2024 and proceeding through discovery — targets the App Store’s commission structure as a core component of Apple’s alleged monopoly power. If U.S. courts or regulators force Apple to reduce its take rate or allow meaningful sideloading, the 30 percent toll that funds Apple’s AI revenue could shrink to 15 percent or less. For context, Epic Games’ lawsuit established that Apple’s effective commission rate on game purchases must drop in certain circumstances, and the AI app category is likely to face similar pressure as its revenue scale attracts judicial attention.

The second crack is the rise of agentic AI interfaces that bypass the app paradigm entirely. OpenClaw’s explosive growth — 280,000 GitHub stars in five months, making it the most popular open-source project in history — demonstrates that users are willing to interact with AI agents through messaging platforms, voice interfaces, and desktop applications that never touch the App Store. If the dominant AI interaction model shifts from “open an app and subscribe” to “issue a voice command and let an agent handle it,” Apple’s commission structure becomes irrelevant because there is no app to tax. The irony is that Apple’s own Siri delays are accelerating this shift: every month that LLM Siri fails to ship is another month that users build habits around ChatGPT’s native app, OpenClaw’s WhatsApp integration, or Perplexity’s always-on personal computer agent. These alternative interaction surfaces do not route through Apple’s payment infrastructure, and the longer Siri remains broken, the more entrenched they become.

The third risk is that Apple’s reliance on Google creates a dependency that limits its strategic flexibility. Training Apple Foundation Models on Gemini means that Apple’s AI capabilities are, at a fundamental level, a derivative of Google’s research. If Google decides to restrict access to its most advanced model weights — a plausible scenario if Gemini’s competitive advantage depends on exclusivity — Apple would be forced to find an alternative training partner on short notice or invest dramatically more in its own model development. Apple has approximately 3,000 employees working on AI research, a formidable team by most standards but a fraction of the 5,000-plus researchers at Google DeepMind or the resources available to OpenAI after its $110 billion funding round. The Gemini partnership is a shortcut that buys time, not a permanent solution. If Apple’s internal models do not reach competitive parity within two to three years, the company risks becoming permanently dependent on a rival for its most strategic product capability — the intelligence layer that mediates between 2.2 billion users and every application on their devices.

The roadmap from here and what operators should watch

Apple’s AI strategy for the next twelve months will reveal whether the toll-road model is a temporary windfall or a durable competitive structure. Several signposts deserve close attention from developers, investors, and enterprise operators who build on Apple’s platform.

The first and most important signal is the quality of LLM Siri at launch. If the iOS 26.5 release in May delivers a conversational assistant that genuinely competes with ChatGPT on accuracy, response time, and multi-step task execution, the market will reward Apple with a re-rating of its AI narrative. The 25-percent accuracy gap that Bloomberg reported is not insurmountable — it roughly corresponds to one model generation of improvement at the current pace of frontier lab progress — but closing it requires flawless execution on the distillation pipeline from Gemini to Apple Foundation Models. Any delay beyond May pushes the full launch to iOS 27 in September, which would mean Apple has spent more than two years promising an AI assistant it cannot deliver. For enterprise IT departments evaluating whether to build Siri Shortcuts integrations or default to third-party agent platforms, the May milestone is the decision point.

The second signal is Apple’s developer economics for AI. At WWDC 2026 in June, Apple will almost certainly announce new APIs that let third-party AI models integrate more deeply with Siri, the App Intents framework, and on-device inference through Core ML. The question is whether Apple will also adjust its commission structure for AI subscriptions. A reduced take rate — say, 15 percent from day one rather than 30 percent — would simultaneously reduce Apple’s immediate revenue and increase its long-term platform stickiness by attracting more AI developers. The precedent exists: Apple already offers reduced commissions for small developers and certain content categories. Extending that treatment to AI subscriptions would signal that Apple prioritizes ecosystem growth over short-term extraction.

The third signal is the competitive response from the AI labs themselves. OpenAI’s decision to hire OpenClaw creator Peter Steinberger and integrate agent capabilities directly into ChatGPT suggests a strategy of making the ChatGPT app so indispensable that users will pay Apple’s commission willingly. Anthropic’s Claude, Google’s Gemini app, and xAI’s Grok are all investing heavily in mobile experiences. If these apps collectively drive more than $2 billion in annual App Store commissions by 2027 — a trajectory the current growth rate supports — Apple’s AI toll road becomes one of the most profitable businesses in technology without requiring Apple to build a single model of its own.

For developers and operators navigating this landscape, the actionable checklist is clear:

• Build for Siri Shortcuts now but design for portability. If LLM Siri launches with genuine capability in May, early adopters of the App Intents framework will have a distribution advantage. But hedge by also supporting OpenClaw’s agent protocol and ChatGPT’s function-calling API, because no one knows which interaction surface will dominate by year’s end.
• Budget for Apple’s commission in your AI subscription pricing. If you sell an AI product on iOS, your effective margin is 70 cents on the dollar in year one. Price accordingly or risk subsidizing Apple’s AI strategy with your own operating losses.
• Watch the EU sideloading numbers. If alternative app store adoption in Europe exceeds 10 percent of iPhone users by Q3 2026, the regulatory pressure for global sideloading will become irresistible, and Apple’s commission leverage will decline.
• Monitor the Gemini dependency. If Apple announces a second model training partnership at WWDC — with Anthropic, Mistral, or an internal model effort — it signals that the Google relationship is less exclusive than the January announcement implied. That would be bullish for Apple’s long-term AI independence and bearish for Google’s cloud revenue projections.

Apple’s position in the AI era is a paradox wrapped in a business model: the company that ships the worst frontier AI assistant among Big Tech also runs the most profitable AI distribution channel on the planet. Whether that paradox resolves in Apple’s favor depends entirely on whether Cupertino can close the gap between the intelligence it taxes and the intelligence it builds. The clock is ticking, and every month of Siri delay adds another billion dollars to the rivals’ subscription totals — and another hundred million to Apple’s commission receipts.

In other news

OpenClaw’s “ChatGPT moment” reignites the commoditization debate — CNBC reported that OpenClaw’s explosive rise to 280,000 GitHub stars, crowned by a showcase at Nvidia’s GTC keynote, is sparking concern that AI models are becoming commodities. Forrester analyst Charlie Dai noted that attention is shifting from foundation models to agent frameworks, a trend that could undermine the investment thesis behind richly valued AI labs.

Senators Warren and Blumenthal probe Nvidia’s $20 billion Groq deal — Two Democratic senators sent Nvidia CEO Jensen Huang a letter questioning whether the company’s licensing agreement with inference chip startup Groq was structured to evade antitrust review. The deal hired 80 percent of Groq’s engineering staff, including founder Jonathan Ross, without filing for mandatory Hart-Scott-Rodino merger review.

Meta reportedly planning 20 percent workforce reduction — Meta is considering layoffs that could affect approximately 15,000 employees to offset AI infrastructure spending projected between $115 billion and $135 billion for 2026. The potential cuts would represent Meta’s largest restructuring since the 2022–2023 layoffs, with Reality Labs already shedding 1,500 positions earlier this year.

Google rolls out Gemini upgrades across Workspace — Google announced expanded Gemini AI integration across Docs, Sheets, Slides, and Drive, starting in beta for AI Ultra and Pro subscribers. The update enables context-aware editing, formula generation, and slide creation directly within Workspace applications.

Replit triples valuation to $9 billion in six months — The agentic AI coding platform raised a $400 million Series D led by Georgian Partners, with participation from Andreessen Horowitz, Coatue, and celebrity investors including Shaquille O’Neal and Jared Leto. Replit now serves over 50 million users and targets $1 billion in annual recurring revenue by year’s end.