Apple is paying Google roughly $1 billion per year for Gemini technology. Not Gemini models — Gemini technology. That distinction matters enormously, and most coverage gets it wrong. Google provides the architectural innovations, training techniques, and infrastructure. Apple trains its own models — the Apple Foundation Models (AFM) — using those techniques. Apple retains complete control over the software stack, the privacy guarantees, and how the models behave.

Thomas Kurian, Google Cloud’s CEO, confirmed the partnership at Google Cloud Next ‘26. Apple’s AFM Cloud Pro model is described as “comparable to Google’s Gemini frontier models.” Developers get access to Gemini-class intelligence through Apple’s APIs without any Google account, any Google SDK, or any data flowing to Google’s servers.

Let’s break down what this deal actually means for people building apps.

The Deal Structure

Here’s what we know about the arrangement:

What Google provides:

• Gemini Mixture of Experts (MoE) architecture and techniques
• Training infrastructure: Google Cloud + Nvidia Blackwell B200 GPUs
• Confidential computing capabilities for Private Cloud Compute
• Ongoing research collaboration on model architecture

What Apple controls:

• All model weights (AFM models are Apple’s property)
• The entire Private Cloud Compute (PCC) software stack
• Training data curation and filtering
• Model behavior, alignment, and safety tuning
• How models are deployed, served, and updated
• All user interactions and developer APIs

What the contract explicitly prevents:

• Google cannot train on Siri queries or any Apple user data
• Google has no access to PCC inference requests or responses
• Google cannot influence model behavior or outputs
• No Google branding or attribution in the user experience

This is fundamentally different from the Microsoft-OpenAI relationship, where Microsoft essentially resells OpenAI’s models through Azure. Apple isn’t reselling Gemini. Apple is using Gemini’s innovations to build its own competing models. Google gets paid for the R&D transfer, not for model inference.

Why Apple Chose Google Over Building From Scratch

Apple has thousands of ML engineers. They weren’t starting from zero. But building a frontier-class LLM from scratch takes years, and the market wasn’t going to wait. By licensing Gemini’s MoE architecture and training techniques, Apple compressed what would have been a 3-4 year development cycle into roughly 18 months.

The MoE architecture is particularly important. Siri’s model is 1.2 trillion parameters — impossibly expensive to run if every token activated every parameter. With MoE, only a fraction of experts activate per token, making inference tractable on Apple’s infrastructure. Google pioneered this approach with their Switch Transformer research and scaled it with Gemini. Apple didn’t want to reinvent that wheel.

The Nvidia Blackwell B200 infrastructure on Google Cloud is the other piece. Training a 1.2T model requires massive GPU clusters. Apple doesn’t operate AI training clusters at that scale — their data centers are optimized for services like iCloud, not GPU training. Rather than build that capability in-house (a multi-billion dollar, multi-year infrastructure project), they lease it from Google with confidential computing guarantees.

What Developers Actually Get

As a developer, you probably don’t care about corporate deal structures. You care about what lands in your Xcode. Here’s what the partnership delivers:

Free Cloud AI for Your Apps

When your app uses Apple’s Foundation Models framework with the PCC executor, users get cloud AI inference at no cost to you or them. Apple eats the compute cost as part of the Apple Intelligence offering. This means:

• No API keys to manage
• No per-token billing
• No rate limits on your app (Apple manages capacity)
• No usage tracking that could affect your business model

Compare this to calling Claude, GPT, or Gemini directly where you’re paying $3-15 per million tokens. For common tasks like summarization, entity extraction, or natural language understanding, the free PCC models are good enough — and they’re free.

Gemini-Class Model Quality

AFM Cloud Pro is described as comparable to Gemini frontier models. In practice, this means developers on Apple’s platform have access to reasoning capabilities that compete with the best available AI APIs — but bundled into the OS. Your app doesn’t need to ship an API key, manage billing, or explain to users why a feature costs money.

For AI app architecture decisions, this changes the calculus. The “should I call an external API or use on-device” question now has a third option: “use Apple’s cloud model for free with strong privacy guarantees.”

Privacy by Architecture

Every request to PCC is end-to-end encrypted. Apple’s security architecture ensures:

• Requests are processed in secure enclaves
• No data is persisted after inference completes
• Apple cannot read the requests or responses (cryptographic guarantee, not policy)
• Independent security researchers can audit PCC’s code

For developers building apps in regulated industries — healthcare, finance, legal — this is significant. You can use AI features that process sensitive data without the GDPR compliance headaches that come with sending data to third-party APIs. The data never leaves Apple’s encrypted pipeline, and there’s no data controller relationship with Google.

How This Compares to Microsoft-OpenAI

The Microsoft-OpenAI relationship is the obvious comparison, but the structures are radically different:

Aspect	Apple × Google	Microsoft × OpenAI
Model ownership	Apple owns AFM models	OpenAI owns GPT models
What’s licensed	Architecture & techniques	Model inference & hosting
Developer cost	Free (bundled in OS)	Pay-per-token via Azure
Data handling	Apple PCC (encrypted, no retention)	Azure data residency policies
Provider lock-in	Apple controls full stack	Dependent on OpenAI releases
Branding	No Google mention	Azure OpenAI branding

Microsoft resells OpenAI’s intelligence. Apple builds its own intelligence using Google’s research. The developer implications are profound:

• Microsoft developers are locked into OpenAI’s model releases, pricing changes, and deprecation schedules
• Apple developers get models that Apple controls end-to-end, with no dependency on Google’s roadmap
• If Google and Apple’s partnership ends tomorrow, Apple still has its trained models and can continue serving them

This is closer to how a car manufacturer licenses engine technology from another company but designs the entire vehicle themselves. The engine technology matters, but the driver experience is entirely the manufacturer’s.

The Privacy Question

“But wait — doesn’t Google get Apple’s data through the training infrastructure?”

No. This is the most misunderstood aspect of the deal. Here’s how confidential computing works in this context:

1. Apple’s training data lives encrypted on Google Cloud’s infrastructure
2. Training runs execute inside confidential computing VMs (Nvidia’s confidential computing on B200)
3. Google’s infrastructure operators cannot access the data inside these VMs — it’s encrypted in memory
4. Only Apple’s code, signed by Apple’s keys, can decrypt and process the training data
5. The resulting model weights are encrypted and transferred back to Apple’s infrastructure

Google provides the hardware and data center facilities. They cannot see what’s being trained, on what data, or what the resulting models do. It’s like renting a locked safety deposit box — the bank provides the vault but can’t open your box.

The contract explicitly bars Google from training on any Siri queries or interaction data. Even if Google could somehow access it (they can’t, technically), they’re legally prohibited from using it.

For developers thinking about privacy implications of AI, this architecture means: your users’ data processed through PCC never touches Google’s systems in any accessible form. The Google relationship is about hardware and research, not data.

What This Means for Multi-Model Strategies

If you’re building apps that use multiple AI models, the Apple-Google partnership creates an interesting dynamic:

Tier 1 (Free, Private): Apple’s on-device AFM models for simple tasks. Fast, private, no cost.

Tier 2 (Free, Cloud): Apple’s PCC models for complex reasoning. Gemini-class quality, still no cost, strong privacy.

Tier 3 (Paid, Specialized): Direct API calls to Claude, GPT, or specialized models for tasks where you need specific model behavior, longer context, or capabilities PCC doesn’t offer.

Before this partnership, most iOS developers went straight from “basic on-device ML” to “expensive cloud API calls.” Now there’s a powerful middle tier that’s free and private. This shifts the build vs buy decision because the “buy” option includes a free tier that’s actually good.

For agent-based architectures, you can route simple agent reasoning through PCC (free) and only escalate to paid APIs when the task requires specific model capabilities. This dramatically reduces costs for AI-heavy applications.

The Competitive Implications

This deal has consequences beyond Apple’s ecosystem:

For Google: They get $1B/year in guaranteed revenue and validate their AI research dominance. Even their biggest competitor pays for their technology. But they also enable a competitor — Apple’s models compete with Gemini for developer attention.

For OpenAI: The most concerning development. Apple’s platform — 2 billion devices — now ships with Gemini-class AI for free. Developers who previously would have integrated GPT now have a free alternative baked into the OS. OpenAI’s consumer moat weakens if Apple Intelligence becomes “good enough.”

For Anthropic: Less directly threatened. Claude’s differentiation (coding, safety, extended thinking) isn’t replicated by Apple’s general-purpose models. Developers who need Claude-specific capabilities will still pay for Claude. But casual AI feature integration? Apple’s free models eat that market.

For developers: More options, lower costs, better privacy defaults. The pricing comparison becomes less relevant when there’s a capable free option bundled with the platform.

Infrastructure Implications

The Nvidia Blackwell B200 detail is interesting for anyone thinking about AI infrastructure. Apple is training on the latest GPU architecture, which means:

• AFM models are optimized for B200’s capabilities (FP4 precision, 2x memory bandwidth vs H100)
• Apple can retrain and update models more frequently thanks to faster training times
• The cost-per-training-run is lower, enabling more experimentation

For developers running models locally on Apple Silicon, this doesn’t directly help — your Mac’s Neural Engine and GPU are what they are. But it means the cloud models you get for free through PCC are trained on better hardware and updated more frequently than if Apple were using older infrastructure.

What Could Go Wrong

This partnership isn’t without risks:

Dependency risk: If Google raises prices significantly, Apple’s AI costs balloon. Apple mitigated this by owning the models — they could theoretically move training to their own or another provider’s infrastructure.

Technology gap risk: If Google withholds breakthroughs from the licensing deal, Apple falls behind. The contract presumably covers ongoing technology transfer, but details aren’t public.

Regulatory risk: EU antitrust regulators may scrutinize a deal between two of the world’s largest tech companies. Both companies already face DMA obligations.

Quality ceiling risk: If AFM Cloud Pro is only “comparable” to Gemini (not better), developers still have reason to use Google’s API directly for cutting-edge tasks. Apple would need to invest in independent research to surpass what Google provides.

Developer Action Items

If you’re building for Apple’s platforms today:

1. Default to Apple’s Foundation Models for standard AI tasks (summarization, NLU, entity extraction). Free, private, and improving.

2. Use the Language Model Protocol to abstract your model dependency. If PCC improves enough, you can drop paid API calls entirely.

3. Keep paid APIs for specialized tasks where you need specific model behavior — Claude for coding, GPT for specific fine-tuned capabilities, specialized models for domain tasks.

4. Design for the privacy advantage. Apple’s architecture lets you market AI features as “private by default” — a genuine competitive differentiator against apps using standard cloud APIs.

5. Don’t over-index on one provider. The multi-model approach remains smart. Apple’s free tier handles the base case, paid APIs handle the edge cases.

The Bigger Picture

The Apple-Google partnership signals where the industry is heading: AI capabilities commoditizing at the infrastructure level while differentiation moves to integration, privacy, and user experience. Google’s economic model is selling AI technology (to Apple, Samsung, and others). Apple’s economic model is selling AI products (iPhones, Macs, services) that use that technology invisibly.

For developers, this is overwhelmingly positive. You get Gemini-class AI bundled free into the platform. You keep your privacy guarantees. You retain the option to use any other model through standardized protocols. The cost of adding AI features to your iOS app just dropped to zero for the most common use cases.

The era of “AI features cost money per request” isn’t over — specialized models and heavy usage still require paid APIs. But for the vast majority of AI-enhanced app features, Apple just made it free, private, and easy. That’s what a $1B/year partnership buys.

Frequently Asked Questions

Does Google have access to data from my app when it uses Apple’s Foundation Models?

No. Google provides training infrastructure and architecture technology. All inference happens on Apple’s Private Cloud Compute with end-to-end encryption. Google cannot access training data, inference requests, model outputs, or any user data. The contract explicitly prohibits Google from training on Siri queries or Apple user data.

Is AFM Cloud Pro actually as good as Gemini?

Apple describes it as “comparable to Google’s Gemini frontier models.” In practice, this means it’s competitive for general tasks but may lag on specific benchmarks where Gemini excels. For most app-level AI features (summarization, NLU, entity extraction), the quality difference is negligible. For cutting-edge reasoning or specialized tasks, you may still want to use Gemini or Claude directly.

Do I need to pay anything to use Apple’s cloud models in my app?

No. PCC inference is free for developers and users. Apple bundles the compute cost into the Apple Intelligence offering. You don’t need API keys, billing accounts, or usage tracking. Your app calls the Foundation Models framework, and Apple handles everything else.

What happens if Apple and Google end their partnership?

Apple owns its trained AFM models outright. If the partnership ends, Apple loses access to future Gemini architecture innovations and would need alternative training infrastructure. But existing models and deployed services continue working. Developers wouldn’t notice any disruption in the short term.

Can I use both Apple’s free models and paid APIs like Claude in the same app?

Yes, and Apple encourages this through the Language Model Protocol. You can route simple tasks to PCC (free) and complex tasks to Claude or GPT (paid) through the same API interface. This multi-tier approach gives you the best cost-performance ratio.

How does this affect the self-hosted vs API decision for Apple developers?

For apps targeting Apple platforms, PCC essentially gives you a “self-hosted quality” model without maintaining infrastructure, at zero cost. The self-hosted vs API tradeoff shifts because PCC offers cloud-API convenience with self-hosted-level privacy. You’d only self-host if you need models Apple doesn’t offer or need to avoid Apple’s infrastructure entirely.