Six weeks into iOS 18 and Apple Intelligence on my iPhone 15 Pro, and Snapdragon Summit just gave me the comparison point I was waiting for. The Snapdragon 8 Elite landed last week with a 45 TOPS Hexagon NPU and Qualcomm’s clearest statement yet that the entire inference stack should run on the device. Now I can put both architectures side by side and ask the question Apple has not answered: what happens when the cloud call cannot be made?
What Apple actually built with Apple Intelligence
Apple Intelligence in iOS 18 runs two inference paths. Simple queries (message summaries, basic Siri requests, notification prioritization) process locally on the A17 Pro’s Neural Engine in the iPhone 15 Pro, or on the A18 and A18 Pro in the iPhone 16 lineup. Complex queries route to Apple’s Private Cloud Compute infrastructure. Apple claims the server-side processing is cryptographically verified and does not store data. That is a more serious privacy guarantee than what you get from a raw OpenAI API call, but it is still categorically different from local weights you own.
The Messages summarization feature is the test case. Conversation summaries under 500 tokens run on-device. Longer threads or requests involving external data (calendar integration, location context) go to the cloud. The user does not choose. The system decides based on complexity and context window size. This is not a privacy setting you toggle. It is an architectural constraint baked into how Apple Intelligence was designed.
Apple’s pitch is that Private Cloud Compute makes the cloud hop acceptable. The cryptographic guarantees are real. My question is what happens when those servers are unavailable, rate-limited, or simply outside the regulatory perimeter you need to operate within.
What the Snapdragon 8 Elite comparison reveals
Snapdragon Insider bias on the table: I covered the Snapdragon 8 Elite platform play in detail after Summit last week, and the architectural difference from Apple’s approach is not subtle. The Hexagon NPU at 45 TOPS handles the full inference stack locally. Llama 3.1 8B, Mistral 7B, and Phi-3 run on-device without any cloud round-trip. The trade-off is capability ceiling: you cannot run GPT-4 class models on a phone chip today. But the models you can run are yours, locally, with no server dependency.
Apple’s A17 Pro Neural Engine runs at 35 TOPS, slightly behind the Snapdragon 8 Elite but in the same performance tier. The silicon gap is not the story. The software architecture is. Apple routes complex queries to the cloud because iOS is built around Apple’s services ecosystem and Apple’s control of the experience. Qualcomm ships the NPU and steps back: OPPO, Xiaomi, and OnePlus can configure a fully local inference stack on 8 Elite hardware if that is what their customers need. Apple does not give you that choice, because Apple does not want you to have it.
Why this matters beyond the benchmark
The Apple-EU standoff over Siri AI in July showed exactly where hybrid cloud architectures break down for regulated markets. The European Commission blocked Apple’s next-gen Siri features from launching in the EU because of concerns about how Apple’s AI infrastructure integrates with the Digital Markets Act. Apple’s response was legal challenges and delay. The EU’s response was to make clear that cloud-routed AI from a US hyperscaler is not the same as sovereign AI infrastructure, regardless of how good the cryptography is.
From my trade perspective, European enterprises are asking this question before they commit to any AI deployment: can we run this without depending on US cloud infrastructure? Apple Intelligence running complex queries through Private Cloud Compute, which is US-based infrastructure Apple controls, does not answer that question the way regulated industries need it answered.
Meta’s Llama 3.2, which I covered in September, changes the calculation on the Snapdragon side. The 11B parameter multimodal variant runs on hardware like the Snapdragon 8 Elite without any licensing fees, usage restrictions, or cloud dependency. Nobody can revoke your local weights. No regulatory body can block a model you are running on your own silicon. That is a sovereignty argument Apple’s hybrid model cannot match, no matter how good the Private Cloud Compute security documentation is.
iOS 18 is a well-executed update. The Messages redesign is clean, the Siri improvements are noticeable in daily use, and the on-device processing for simple tasks is fast enough that you mostly do not notice the handoff. But the strategic choice to route complex inference to the cloud rather than extending local model capability tells you exactly what Apple is optimizing for. They own the experience end to end. You use what they decide to give you, processed where they decide to process it. Whether that trade-off is acceptable depends on whether you need to own the stack or whether you are comfortable renting the intelligence from a company that can change the terms.
Related reading: the Qualcomm-Xiaomi silicon pipeline behind the Snapdragon side of this comparison.