Apple Intelligence Was Supposed to Change Everything. What Happened?

In January 2026, Apple quietly admitted something that would've been unthinkable two years earlier: it couldn't build competitive AI on its own. The company announced a multi-year partnership with Google to integrate Gemini models into Apple Intelligence, including an on-device version of Gemini for Siri. According to Bloomberg's Mark Gurman, Apple pushed the deal forward after its own internal foundation model — codenamed Ajax — failed to keep pace with what Google, OpenAI, and even Meta were shipping. Let that sink in. Apple, the company with over 2.2 billion active devices, the company that spent decades building the best mobile silicon on earth, the company that controls the most valuable distribution channel in technology — couldn't build AI models good enough to compete. This isn't a story about a company that ignored AI. It's a story about a company that tried and stumbled. And the reasons why are a warning to every organization that thinks "we'll do it ourselves" is always the right strategy. ## The Promise When Apple announced Apple Intelligence at WWDC in June 2024, the pitch was compelling. On-device AI powered by Apple's own models. Private Cloud Compute for heavier tasks, running on Apple silicon servers with end-to-end encryption. No data leaving Apple's ecosystem. Privacy-first AI. This was supposed to be Apple's killer differentiation. While Google and OpenAI vacuum up user data to train better models, Apple would prove you could build great AI without compromising privacy. The messaging was perfect for Apple's brand: we don't need your data, we're just that good. Tim Cook told Good Morning America that generative AI had "great promise" while noting potential dangers. On an earnings call in February 2024, he said Apple was spending a "tremendous amount of time and effort" on AI features. The implication was clear: something big was coming. What we got instead was... notification summaries. ## The Slow Rollout Apple Intelligence launched on October 28, 2024, with iOS 18.1. The feature set was thin. Writing Tools that could proofread, rewrite, or summarize text. A slightly updated Siri with a new visual design and the ability to type instead of speak. Notification summaries. A photo cleanup tool. Compare this to what Google and OpenAI were shipping at the same time. ChatGPT had voice mode, vision, custom GPTs, an app store, canvas, and code interpreter. Google had rolled Gemini into Search, Gmail, Docs, and Workspace with features that actually changed how people worked. Apple's writing tools were fine. The notification summaries were... problematic. Users quickly discovered that the AI was generating summaries that misrepresented the content of the original messages, sometimes in dangerous ways. News notifications were summarized inaccurately. Sarcastic texts were taken literally. The feature became a meme. More features trickled out over subsequent releases. Image Playground arrived in December 2024 with iOS 18.2, along with Genmoji and a ChatGPT integration for Siri. But the image generation was cartoonish and limited. You couldn't generate photorealistic images — Apple presumably chose to avoid the liability — and the results looked like clip art compared to what Midjourney and DALL-E were producing. The language support rollout was glacially slow. U.S. English only at launch. British English, Australian English, and a few other English variants didn't arrive until December 2024. Chinese, Japanese, French, German, and other major languages didn't get support until March 2025 with iOS 18.4. For a company that sells billions of devices globally, launching an AI system that only worked in American English for the first two months was baffling. ## Why It Went Wrong The root cause is structural, and it goes back years. Apple's culture of secrecy — the same culture that produced incredible product launches — actively sabotaged its AI efforts. Trevor Darrell, a UC Berkeley professor, noted that Apple's secrecy "deterred graduate students" from wanting to work there. The best AI researchers want to publish papers, attend conferences, and build their reputations. Apple wanted them to work in a vault. This created a talent pipeline problem. Google, Meta, OpenAI, and Anthropic attracted the best researchers because they let them publish. Apple attracted great hardware engineers and great software designers, but struggled to build a world-class AI research team. When ChatGPT launched in November 2022, Apple executives were reportedly "blindsided." The company had been working on ML features for years — the Neural Engine has been in iPhones since the A11 Bionic in 2017. But there's a massive gap between "ML features embedded in the camera app" and "a general-purpose foundation model that can hold a conversation." Apple tried to catch up. The Ajax project was supposed to produce competitive foundation models. But by the time Apple Intelligence shipped, their on-device model was beating small models from Mistral and Microsoft while "roughly matching" GPT-4 on their server model — by their own evaluation. In a world where GPT-4 was already old news, matching GPT-4 wasn't enough. The privacy-first architecture, while admirable, also imposed real constraints. On-device models are limited by the phone's memory and processing power. Apple's on-device model had to be small enough to run on an A17 Pro chip with 8GB of RAM. That's a severe constraint. You can't fit a frontier-quality model in 8GB. The result was a model that could do text summarization and rewriting but couldn't hold a candle to the conversational abilities of Claude or GPT-4. Private Cloud Compute solved the size constraint for server-side tasks, but Apple's infrastructure wasn't built for this. They'd never operated cloud AI at scale. Google has been running massive ML workloads for over a decade. OpenAI has Microsoft's Azure infrastructure. Apple was starting from scratch, running Apple silicon servers that were powerful per-chip but unproven at datacenter scale. ## The Gemini Concession The January 2026 Google partnership was the clearest signal that Apple's internal efforts had fallen short. According to reporting, Apple will integrate a version of Google's Gemini — a model with roughly 1.2 trillion parameters — as an on-device model for Siri. That's dramatically larger and more capable than anything Apple built internally. The deal presumably includes access to Google's cloud infrastructure for heavier tasks, while Apple maintains control of the privacy layer through its devices and Private Cloud Compute. Apple already had a ChatGPT integration (GPT-4o powered Siri's system-wide ChatGPT pass-through since December 2024), but that was a bolt-on. The Google deal goes deeper — it's about the foundation model itself. For Apple, this is a pragmatic choice. They couldn't close the gap fast enough. Better to ship a competitive product using someone else's model than to ship an inferior product using your own. The Foundation Models API, announced at WWDC 2025, lets third-party developers access Apple's on-device models — but if those models are partly Google's, the "Apple-built" narrative gets complicated. For Google, the deal is a massive win. Getting Gemini onto 2+ billion Apple devices as the default AI powering Siri is potentially worth more than any ad deal. It's reminiscent of the Search deal where Google pays Apple billions annually to be the default search engine in Safari. Now Gemini might become the default AI brain behind Siri. ## What Apple Still Has I want to be fair. Apple isn't dead in AI. They have real advantages that nobody else can match. **Distribution.** 2.2 billion active devices. No other AI company can get their model into that many hands overnight. When Apple Intelligence ships a new feature, it doesn't need an app download or a subscription — it's just there. **The Neural Engine.** Apple's on-device ML hardware is genuinely excellent. The A-series and M-series chips have Neural Engines that can run billions of operations per second with remarkable efficiency. For on-device inference of appropriately sized models, Apple's hardware is best-in-class. **Privacy architecture.** Private Cloud Compute is a genuinely novel approach. The idea that cloud-side AI processing can be verifiable, encrypted, and stateless is something no other company has shipped. For users and enterprises that care deeply about data privacy, this matters. **Ecosystem integration.** Nobody does system-level integration like Apple. AI features that work across Messages, Mail, Photos, Safari, and Siri in a unified way — that's harder than it sounds, and Apple's vertical integration makes it possible. ## What It Means Apple's AI stumble is instructive for the whole industry. It shows that hardware excellence, distribution, and brand don't automatically translate to AI leadership. Models matter. Data matters. Research culture matters. The company that should've been best positioned to win the on-device AI race — the company that makes the best mobile chips on earth and controls the entire stack from silicon to software — ended up licensing models from Google. That's a humbling outcome. Going forward, Apple's AI strategy will be hybrid: their own on-device capabilities for privacy-sensitive tasks, Google's Gemini for the heavy lifting, and ChatGPT as a system-wide opt-in for complex queries. It's a reasonable strategy, but it puts Apple in the unfamiliar position of depending on partners for a core technology. For the first time in a long time, the most important software on your iPhone won't be built by Apple. Whether that matters to consumers is the billion-dollar question. My guess? Most people won't care, as long as Siri actually works. And maybe that's the real lesson: users don't care who built the model. They care if it answers their question. Apple's engineers know that. It's why they made the Google deal. Pride is expensive, but a bad product is more expensive.