Apple has effectively given Siri a second architecture.

The company’s new Siri AI rollout, described by The Decoder as a rebuilt virtual assistant, is not being framed as a pure on-device rewrite or a simple cloud upgrade. Instead, Apple is pairing Apple Foundation Models with a System Orchestrator that decides where a request should run, moving work between on-device processing and Nvidia-powered cloud compute. That matters because the technical center of gravity is no longer one model or one device class. It is the routing layer.

That shift also changes the product story. Siri is being positioned as a system-wide assistant capable of taking actions across the OS, but the feature set is not evenly available everywhere. Apple says the strongest on-device capabilities require newer hardware with at least 12 GB of RAM, which immediately narrows the install base. And in the European Union, the Digital Markets Act adds another constraint: Apple says Siri AI will not launch on iPhones and iPads there, even as it remains available on macOS and visionOS.

System Orchestrator: the hybrid backbone

The most important architectural detail in Apple’s reboot is not the model branding. It is the orchestration.

A System Orchestrator sits between the user request and the execution environment, deciding whether a task should be handled locally or offloaded to the cloud. In practical terms, that makes Siri AI a hybrid stack rather than a single inference path. Simple or privacy-sensitive tasks can stay on device. Heavier or less latency-bound requests can be sent to Nvidia-powered cloud compute.

That design has three immediate implications.

First, latency becomes a routing problem as much as a model-quality problem. Apple is no longer promising that everything happens instantly on the handset. It is promising that the system can choose the fastest acceptable route.

Second, privacy is being governed by architecture rather than by slogan. On-device execution is the default story Apple has always preferred to tell, but the new assistant is explicit about cloud fallback when needed. That makes the Orchestrator the policy layer as much as the performance layer.

Third, hardware fragmentation becomes part of the feature definition. The Decoder reports that the best on-device features require at least 12 GB of RAM, which raises the bar well above many existing Apple devices. For developers and product teams, that means the feature envelope will depend on whether a device can actually host the local path the Orchestrator wants to use.

Foundation Models and lineage: Google, Gemini, and Apple’s own stack

Apple’s model story is more nuanced than a binary choice between “Apple AI” and “Google AI.” According to The Decoder, Apple Foundation Models were developed in close collaboration with Google and build on Gemini technology. That is a significant lineage claim, but it is also one Apple is trying to contain carefully.

During a Tech Talk after the keynote, Craig Federighi drew a boundary around the partnership: “The amount of the Google Assistant we use is none.” Apple’s point is that Siri AI is not the Gemini app, not Google’s consumer model bundle, and not Google Search dressed up as assistant intelligence. That distinction matters because it defines where Apple wants credit, and where it wants distance.

Apple also says it relies on its own World Knowledge Service for world knowledge, rather than using Google as a knowledge base. That is an important technical and strategic line. If the system can query a proprietary knowledge service while still leaning on Google-derived model lineage, Apple can argue it has separated model provenance from retrieval infrastructure.

For observers, that means the interesting question is not whether Google is involved. It plainly is. The question is how deep the dependence runs, and which parts of the stack Apple is trying to own outright.

Regulatory and market positioning: EU constraints and global rollout

The EU rollout story is as important as the model story.

Apple says Siri AI will not launch on iPhones and iPads in the European Union because of Digital Markets Act constraints, while macOS and visionOS remain in scope. That creates a very different go-to-market map from the one Apple can execute elsewhere. It also suggests that regulatory interpretation is shaping product packaging at the platform level, not just the feature level.

This matters for two reasons.

One is user experience. If Siri AI is available on desktop and spatial-computing devices but blocked on mainstream mobile hardware in the EU, Apple risks creating a fragmented assistant experience across regions and device categories.

The other is strategic leverage. Apple can use macOS and visionOS to ship the assistant in constrained markets while preserving a more aggressive mobile rollout elsewhere. In effect, the EU becomes a test case for how much of Apple’s hybrid AI strategy can be decoupled from the iPhone.

That is especially notable because Siri has historically been thought of as a phone-first assistant. A DMA-shaped rollout would push it into a more explicit platform hierarchy, where availability depends as much on regulation as on silicon.

Implications for developers, tooling, and performance

For developers, the key lesson is that Apple is treating assistant behavior as an orchestration problem across heterogeneous execution targets.

That has implications for tooling. Apps and system integrations will need to tolerate different response paths depending on whether a query is resolved on-device or pushed into the cloud. Performance expectations will also need to be calibrated around model routing rather than a single benchmark number. A request that stays local on a 12 GB RAM-capable device may behave differently from the same request on older hardware or in a region where the assistant is not fully enabled.

The World Knowledge Service is another signal that Apple wants a tighter retrieval story than generic web search. If that service is what feeds world knowledge into the assistant, then the quality of Siri AI will depend not just on model size or inference speed, but on how well Apple curates and updates its own knowledge layer.

For tooling teams, that suggests three priorities:

  • support graceful degradation across hardware tiers
  • design for routing-aware assistant behavior rather than fixed execution assumptions
  • treat privacy controls and retrieval boundaries as first-class product requirements

That is a more complex developer surface than a conventional voice assistant, but it is also closer to how modern AI systems are actually deployed.

What to watch next: 2026–2027

The next 12 to 24 months will show whether Apple’s hybrid approach is robust or merely pragmatic.

The most important signals to watch are latency thresholds, because the Orchestrator only works if local and cloud paths feel seamless. Hardware-enablement requirements will also matter, especially if the 12 GB RAM threshold becomes the dividing line for premium assistant behavior across the Apple portfolio. And regulatory compliance will remain a moving target, particularly if DMA interpretation continues to shape what can ship on iPhone and iPad in the EU.

If Apple can keep the assistant responsive while preserving credible privacy boundaries, the System Orchestrator could become the defining abstraction of this Siri generation. If it cannot, the hybrid stack may end up looking less like a breakthrough than a compromise.

Either way, Apple’s second shot at Siri is no longer about asking whether the assistant is “smart enough.” It is about whether Apple can make a routed AI system feel simple to use while the underlying stack gets more complicated behind the scenes.