Apple is emerging as a potentially strong player in Artificial Intelligence not because it leads the frontier model race, but because the value of raw intelligence is falling as models improve across the board. As open and commercial systems become more capable on lower-end hardware, the gap between the best model and the rest is narrowing. That shift weakens the assumption that the biggest moat comes from owning the most advanced model and the largest compute footprint. It also favors companies that preserved flexibility rather than committing heavily to expensive infrastructure and subsidized usage.
Several rivals are portrayed as exposed to the financial strain of chasing scale. OpenAI raised at a $300B valuation and then shut down Sora, the video product they’d been positioning as a creative industry flagship, because it was running at roughly $15M a day in costs against $2.1M in daily revenue. Disney had already signed a three-year licensing deal for Sora to generate content from Marvel, Pixar, and Star Wars characters. They were finalising a $1B equity stake in OpenAI. On the model side, Gemma 4 is presented as evidence of fast-moving commoditization: it scores 85.2% on MMLU Pro and matches Claude Sonnet 4.5 Thinking on the Arena leaderboard. 2 million downloads in its first week. Anthropic is taking a different path by building workflow products around its models, but even that strategy comes with steep subsidies. One analysis found a max-plan subscriber consuming $27,000 worth of compute with their 200$ Max subscription.
Apple’s main advantage is framed as context. The company already sits on personal and environmental data across 2.5 billion active devices, including health signals, photos, notes, messages, location history, and app behavior. That creates the foundation for useful on-device Artificial Intelligence without sending sensitive information to external services. The Gemini deal, where Apple signed a $1B to license Google’s frontier model for the queries that need cloud-scale reasoning, fits that approach. Rather than owning the most expensive model stack, Apple keeps control of the operating system, the local context, and the user relationship.
Its chip architecture strengthens that position. Apple’s unified memory design puts CPU, GPU, and Neural Engine on the same die with shared high-bandwidth memory, which suits LLM inference because the workload is heavily constrained by memory movement rather than raw compute. That enables local execution techniques that are difficult on conventional systems. Someone recently ran Qwen 397B, a 209GB model, on an M3 Max Mac at ~5.7 tokens per second, using only 5.5GB of active RAM. The weights live on the SSD and stream in at ~17.5 GB/s as needed. The broader implication is platform control: Apple may not need to win the model race if it becomes the preferred place where models and agents run efficiently.
