The Most Important AI Launch This Week Is Not in a Data Center

iOS 27 and iPadOS 27 started rolling out today. The headline is a rebuilt Siri. The useful fact is where it runs.

For three years, most “AI strategy” has meant a cloud API, a token budget, and a hope that the status page stays green. That model is fast. It is also fragile. When the model lives somewhere else, work stops when the network, the vendor, or the region fails.

On-device intelligence is a different design. The phone already has the context: the screen, the calendar, the draft, the photo. A local assistant that can see the work and act without a round trip is not a gadget feature. It is optionality. The same week, Perplexity put a portable computer on Windows machines with NVIDIA RTX chips—another bet that serious work should be able to run next to the person doing it.

Fragile companies will treat this as a consumer story. Anti-fragile companies will treat it as a design brief.

The stack that holds up under stress is three tiers, not one.

Cloud frontier models. Use them for the hard, infrequent work: novel analysis, long-context synthesis, messy judgment. They are an option you exercise when the payoff is high. They should not be the floor every daily workflow stands on.

Local or private data center. This is the tier most companies skip. On-prem GPUs, colo, or a regional private cloud hold proprietary data, regulated work, and agents that run all day. When the public API is dark or rate-limited, this layer still produces output you control.

Local device. Phone, laptop, RTX box. This is last-mile continuity. Drafts, search, summarize, meeting notes, screen-aware help. It should keep a knowledge worker useful in an airport, a blackout, or an outage.

The rule is simple. Default work runs on the device or the private floor. The frontier model is a spike, not a dependency. If any one tier fails, the other two still produce a worse-but-alive result.

Do not buy three identical copies of the same stack. That is expensive fragility. Different jobs belong on different tiers. Frontier for spikes. Private compute for secrets and volume. Device for continuity.

The next outage will not invent that architecture for you. The devices arriving this week already contain the cheaper version of the lesson: intelligence that lives closer to the work is harder to knock over. Companies that collapse those three tiers into one API have a vendor. Companies that keep all three have a system that gets stronger when the first one fails.