Apple rebuilt Siri on Google and NVIDIA hardware
Two years after demoing a smarter Siri that never shipped, Apple tried again at WWDC 2026. This time they actually have something to show. The catch: they didn’t build it alone.
Siri AI is a ground-up rewrite that turns the voice assistant into a conversational AI. It reads your screen, pulls context from your messages and emails, and can take actions across the entire OS. It lives in a dedicated app now, not just a voice bubble. The Dynamic Island got a new animation for it.
Under the hood, this is not Apple silicon all the way down. Apple developed five new Apple Foundation Models (AFM) in collaboration with Google, building on Gemini technology. Four of them run on Apple hardware. The fifth and most capable one, AFM Cloud Pro, runs on NVIDIA GPUs inside Apple’s Private Cloud Compute infrastructure, which now extends into Google Cloud. Apple’s AI lead Amar Subramanya told CNBC that AFM Cloud Pro is comparable in quality to Google’s Gemini frontier models. No benchmarks have been released.
Craig Federighi was quick to clarify the Google relationship. “The amount of the Google Assistant we use is none.” Apple built its own World Knowledge Service for factual queries rather than relying on Google Search. The Gemini-derived models were “refined” for Apple’s needs, not simply plugged in.
A System Orchestrator decides whether each query runs locally or hits the cloud. Federighi positioned this router as central to the privacy story. All Private Cloud Compute binaries will be published for public inspection, including the new Google Cloud nodes.
The hardware cutoff is aggressive. The best on-device features require an iPhone 17 Pro or iPhone Air, an iPad with M4 and at least 12 GB RAM, or a Mac with M3 with 12 GB RAM. The standard iPhone 17 ships with 8 GB and does not qualify. Neither does the iPhone 16 Pro Max. Users with older hardware can still use Siri AI, but most requests will route to the cloud with higher latency.
iOS 27 itself supports every iPhone from the iPhone 11 onward, making this the broadest iOS release ever. The team clearly drew a line between the OS and the AI features layered on top of it.
The EU gets nothing. Siri AI will not launch on iOS or iPadOS in the European Union. Apple blames the Digital Markets Act, arguing it would have to grant any third-party assistant the same system-level access Siri AI has. Apple proposed a “Trusted System Agent” middleware layer with an 18-month phased rollout. The European Commission rejected it, along with every other proposal Apple put forward. macOS 27 and visionOS 27 will get Siri AI in the EU since the DMA only designates iOS and iPadOS as core platform services. Apple Watch is collateral damage: watchOS 27 requires a paired iPhone with Siri AI active, which EU users won’t have.
The economics are telling. Apple is spending roughly $14 billion in capex this year on AI. The rest of the industry has committed a cumulative $900 billion. Apple posted historic iPhone sales last quarter regardless. For developers with fewer than 2 million first-time downloads, Apple is offering free access to Foundation Models running in Private Cloud Compute, no cloud API costs. The move mirrors Apple’s existing Small Business Program for App Store commissions and directly targets the growing problem of AI experimentation costs. Meta and Amazon have discontinued internal AI token leaderboards. Uber burned through its 2026 AI budget in four months.
The developer story extends beyond free inference. The new Core AI library integrates with PyTorch, mapping ATen operators to Core AI operations so developers can bring existing models to Apple hardware. The Foundation Models Framework now supports image inputs, custom skills, and server-side model execution.
Simon Willison summarized the industry reaction well: given how badly burned everyone was by the 2024 Apple Intelligence announcements that never shipped, he’s holding to a strict “I’ll believe it when I see it” policy. The features do at least look feasible with today’s technology. Vision LLMs extracting information from the screen sidesteps the need for every app to ship custom integration code, which was the bottleneck in the original Apple Intelligence rollout.
Federighi took a clear shot at the competition: “Some appear to be racing forward, seemingly pursuing AI for the sake of AI, without clear regard for the people that it’s ultimately meant to serve.” That line works as both a jab at OpenAI’s chaotic product strategy and an acknowledgment that consumer sentiment around AI has shifted negative.
This was also Tim Cook’s last WWDC keynote as CEO. John Ternus, currently SVP of Hardware Engineering, takes over September 1. Cook closed with “It’s been the honor of a lifetime.” The transition adds weight to the entire event. Whether the Siri AI features actually ship on time this time will be one of Ternus’s first tests.
Sources: The Decoder, TechCrunch, Simon Willison, TechCrunch