At WWDC on Monday, Apple peeled back the curtain on its AI infrastructure, revealing a system design that prioritizes privacy over raw scale—but still leans on Nvidia GPUs and Google’s frontier models for heavy lifting.

The headline: Apple Foundation Model Cloud Pro (AFM Cloud Pro), its most capable model, will run on Nvidia GPUs hosted in Google Cloud, integrated into Apple’s Private Cloud Compute (PCC) infrastructure. This is the first official confirmation that Nvidia hardware powers any Apple Intelligence feature, despite the Google partnership announced in January.

Apple AI executive Amar Subramanya confirmed AFM Cloud Pro is comparable to Google’s Gemini frontier models. But the architecture is anything but off-the-shelf. VP of software Sebastian Marineau-Mes explained Apple required Nvidia’s latest chips to be configured with “ambiguous confidential compute”—a technology that prevents the GPU from reading server memory—to meet Apple’s privacy guarantees.

On the device side, a system orchestrator routes each AI query to the appropriate model based on compute needs and data sensitivity. Craig Federighi called this orchestrator “key to the privacy architecture of our entire system.”

Perhaps the most revealing detail: Apple’s on-device and cloud models—AFM Core, Core Advanced Cloud, and Cloud Image—are custom-built for Apple Silicon. But they were refined using outputs from Google’s Gemini frontier models. Federighi clarified Apple is not using Gemini directly or Google’s standard cloud infrastructure; instead, Google’s technology was used in the training pipeline for Apple’s own models.

The trade-off is clear: Apple avoids the massive infrastructure spend of rivals, but still relies on third-party silicon and frontier models to close the capability gap—while betting that local data access and strict privacy boundaries will win users over.

Source: CNBC