The same PCC privacy promises — no data storage, no logging, no training on user data — now apply to Apple Intelligence requests processed on Google Cloud infrastructure . Apple's security research blog and Craig Federighi explicitly stated that "we use none of the models that Google deploys to its customers" . Apple licenses the Gemini model weights and runs them inside its own PCC environment, not on Google's shared AI infrastructure . Google is contractually prohibited from training on any data sent through Apple's PCC pipeline .

What Data Actually Leaves Your Device

Not everything you say to Siri goes to Google. Apple uses a three-tier processing strategy:

• Tier 1 — On-device (most routine requests): Simple tasks — setting alarms, controlling music, opening apps — are handled entirely by Apple's own neural models running locally on your iPhone or Mac. No data is transmitted anywhere .
• Tier 2 — Apple Private Cloud Compute (moderate complexity): Queries that need more power but fall short of the heaviest workloads go to Apple's own PCC servers on Apple silicon .
• Tier 3 — PCC on Google Cloud (heaviest queries): Only the most demanding requests — image understanding, generation, and advanced reasoning — are routed to Google Cloud .

The data that leaves the device is the query (prompt) and the minimal context needed to generate a response. Metadata like IP addresses and request logs are not retained . Apple's latest third-generation Foundation Models include three server-based models running on PCC, which the company says ensures user data is "never stored or shared with anyone, including Apple" .

Apple's Broader AI Strategy: Foundation Models and the Google Partnership

The January 2026 multi-year deal between Apple and Google places Gemini at the core of Apple's next-generation Foundation Models . Apple launched five new Foundation Models at WWDC 2026 — a mix of two on-device models and three that run on PCC . The most demanding model, AFM 3 Cloud Pro, runs on NVIDIA GPUs inside Google Cloud .

Apple chose Google Cloud as its first third-party PCC partner because of Google's confidential computing capabilities, stacking NVIDIA Confidential Computing on Blackwell GPUs, Intel TDX on CPUs, and Google's Titan chip anchoring the root of trust . Apple describes the result as "the first time these primitives have been integrated into a comprehensive, end-to-end confidential inference pipeline capable of operating at a global scale" .

This architecture allows Apple to offer state-of-the-art AI — image understanding, generation, reasoning — without abandoning its privacy-first brand . WWDC 2026 also introduced tools that let developers swap AI providers (Apple on-device, Google Gemini, Anthropic Claude) via Swift Package Manager without code changes .

Usage Limits for Third-Party AI in Apple Creator Studio

Apple Creator Studio (a $129/year subscription) includes AI features in Keynote, Pages, Numbers, Freeform, Final Cut Pro, and Pixelmator Pro . Apple's official support page states that "Some artificial intelligence features of Apple Creator Studio utilize third-party models and may have usage limits and restrictions" .

Apple does not publicly list exact per-feature caps for each third-party model. The company says users can generate at least 50 images and 50 Keynote presentations (with approximately 8–10 slides each) per month, plus presenter notes for 700 slides . However, developer and security researcher Steve Troughton-Smith reported that actual caps appeared dramatically lower in practice — a single Keynote slideshow used half of his monthly ACS limit, suggesting a limit of just two presentations .

Image and Shape Generation in Pixelmator Pro use third-party models and require an Apple Intelligence-capable device . Apple does not publish precise, per-feature usage tiers for third-party AI in Creator Studio. The advertised floor is 50 AI presentations/images per month, but some users have encountered lower effective limits .

The same PCC privacy promises — no data storage, no logging, no training on user data — now apply to Apple Intelligence requests processed on Google Cloud infrastructure . Apple's security research blog and Craig Federighi explicitly stated that "we use none of the models that Google deploys to its customers" . Apple licenses the Gemini model weights and runs them inside its own PCC environment, not on Google's shared AI infrastructure . Google is contractually prohibited from training on any data sent through Apple's PCC pipeline .

What Data Actually Leaves Your Device

Not everything you say to Siri goes to Google. Apple uses a three-tier processing strategy:

• Tier 1 — On-device (most routine requests): Simple tasks — setting alarms, controlling music, opening apps — are handled entirely by Apple's own neural models running locally on your iPhone or Mac. No data is transmitted anywhere .
• Tier 2 — Apple Private Cloud Compute (moderate complexity): Queries that need more power but fall short of the heaviest workloads go to Apple's own PCC servers on Apple silicon .
• Tier 3 — PCC on Google Cloud (heaviest queries): Only the most demanding requests — image understanding, generation, and advanced reasoning — are routed to Google Cloud .

The data that leaves the device is the query (prompt) and the minimal context needed to generate a response. Metadata like IP addresses and request logs are not retained . Apple's latest third-generation Foundation Models include three server-based models running on PCC, which the company says ensures user data is "never stored or shared with anyone, including Apple" .

Apple's Broader AI Strategy: Foundation Models and the Google Partnership

The January 2026 multi-year deal between Apple and Google places Gemini at the core of Apple's next-generation Foundation Models . Apple launched five new Foundation Models at WWDC 2026 — a mix of two on-device models and three that run on PCC . The most demanding model, AFM 3 Cloud Pro, runs on NVIDIA GPUs inside Google Cloud .

Apple chose Google Cloud as its first third-party PCC partner because of Google's confidential computing capabilities, stacking NVIDIA Confidential Computing on Blackwell GPUs, Intel TDX on CPUs, and Google's Titan chip anchoring the root of trust . Apple describes the result as "the first time these primitives have been integrated into a comprehensive, end-to-end confidential inference pipeline capable of operating at a global scale" .

This architecture allows Apple to offer state-of-the-art AI — image understanding, generation, reasoning — without abandoning its privacy-first brand . WWDC 2026 also introduced tools that let developers swap AI providers (Apple on-device, Google Gemini, Anthropic Claude) via Swift Package Manager without code changes .

Usage Limits for Third-Party AI in Apple Creator Studio

Apple Creator Studio (a $129/year subscription) includes AI features in Keynote, Pages, Numbers, Freeform, Final Cut Pro, and Pixelmator Pro . Apple's official support page states that "Some artificial intelligence features of Apple Creator Studio utilize third-party models and may have usage limits and restrictions" .

Apple does not publicly list exact per-feature caps for each third-party model. The company says users can generate at least 50 images and 50 Keynote presentations (with approximately 8–10 slides each) per month, plus presenter notes for 700 slides . However, developer and security researcher Steve Troughton-Smith reported that actual caps appeared dramatically lower in practice — a single Keynote slideshow used half of his monthly ACS limit, suggesting a limit of just two presentations .

Image and Shape Generation in Pixelmator Pro use third-party models and require an Apple Intelligence-capable device . Apple does not publish precise, per-feature usage tiers for third-party AI in Creator Studio. The advertised floor is 50 AI presentations/images per month, but some users have encountered lower effective limits .