Apple's WWDC 2025 keynote revealed a company making a deliberate choice: rather than chase headline-grabbing AI breakthroughs, it is doubling down on design language and embedding small, practical AI features into existing products. The announcement of Liquid Glass—a transparent, glass-morphic redesign across iOS, iPadOS, macOS, and visionOS—landed awkwardly in a moment when the industry is obsessed with large language models. Yet this move makes strategic sense. Apple unified all its operating systems under version 26 and reshaped their visual hierarchy around translucency and layering, borrowing heavily from the Vision Pro aesthetic. The result is coherent, but the timing raised eyebrows: why redesign when everyone else is racing to build better AI?
The answer lies in Apple's fundamental constraint and advantage. Apple is not an AI company. It has no DeepMind, no internal research equivalent to OpenAI or Anthropic. Its strength is hardware and software integration for consumers already locked into its ecosystem. Rather than promise a Siri that works like a true assistant—something neither Google, Amazon, nor OpenAI has actually delivered—Apple announced dozens of narrow, on-device AI features: live translation in FaceTime, visual search from the camera, predictive text in Messages. These are not revolutionary. They are competent, private, and integrated where users already are. The company explicitly did not mention Siri during the keynote, a telling omission that signals internal acknowledgment that the ambitions of last year went unfulfilled.
The deeper issue is data. Apple's privacy-first model means it collects almost no user behavior to train its models. Its largest foundation model has 15 billion parameters—a fraction of GPT-4O's estimated 200–300 billion. Without proprietary training data or public scraping rights, Apple cannot build the knowledge base that separates capable models from mediocre ones. This is where the episode's second major theme emerged: the economics of training data are shifting. The New York Times signed a major deal with Amazon (not OpenAI) to license its recipes and editorial content for Alexa training. Reddit revealed that Anthropic, despite claiming ethical high ground, had scraped hundreds of thousands of visits worth of its data. As foundation models converge in raw capability, ownership of high-quality, fresh, verified training data becomes the differentiator. Publishers now understand they are selling gold, not raw materials. For Apple to compete here, it would need to either build or buy premium data sources—a path that contradicts its stated values and appears beyond its current strategy.
Meanwhile, the AI market itself is consolidating rapidly. OpenAI hit $10 billion in annualized recurring revenue by late 2024, doubling from $5.5 billion six months earlier. Anthropic reached $3 billion. Yet the real money is flowing to the wrapper layer: Cursor, an IDE that simply integrates OpenAI and Anthropic models with code repositories, raised funding at a $10 billion valuation after hitting $100 million ARR in a single year. This pattern—best user experience around commodity models—is becoming the playbook. Apple understands this. Its redesign and careful feature integration follow that logic. The risk is whether users will tolerate incremental product updates when the rest of the world is shipping agents and reasoning systems.
The Nokia parallel is unavoidable. Apple is not ignoring the shift to AI; it is absorbing it slowly, through design and UX rather than through bold claims or proprietary breakthroughs. The question is whether slow absorption is fast enough.