Source: as part of its Google deal, Apple has full access to the Gemini model in its own data centers and can use distillation to produce smaller models
Before we get to today's column, we wanted to flag OpenAI CEO Sam Altman's major reorg, the company's new “Spud” model and its decision …
Context & Ripple Effects
Apple's Gemini arrangement had previously been described as a multibillion-dollar cloud contract, while separate reporting said Apple explored hosting Gemini-based Siri servers in its own data centers to meet its privacy standards. This report adds a more consequential implementation detail: access to the underlying model and the ability to derive smaller versions.
That matters because the partnership appears to extend beyond consuming a hosted model for Siri. It gives Apple a route to adapt a frontier model to its own infrastructure and product constraints while retaining its own customer-facing AI layer.
First-order effects
- Apple can reportedly run Gemini within its own data centers and use distillation to create smaller models, increasing its control over deployment, latency, and privacy-sensitive handling.
- Google becomes not only a cloud/model supplier for Apple features but the provider of a model Apple can adapt and operationalize internally; the report does not indicate Google branding on those features.
Second-order effects
- Apple's ability to distill could reduce the share of some inference workloads that must remain on large, centrally hosted models, even if Google remains the source of the frontier capability.
- The arrangement raises the bar for rival model providers seeking major platform partnerships: customers may want model access, fine-tuning, and infrastructure flexibility rather than a standard hosted-service contract.
Third-order effects
- If replicated, major device platforms may increasingly buy frontier-model capability while building proprietary layers of smaller, task-specific models around it—splitting model creation from product deployment.
- This points to AI competition being shaped as much by control over model access and deployment terms as by benchmark performance, though the durability of that shift depends on whether such access is offered broadly or only through bespoke deals.
The trend: Frontier-model vendors are evolving from API providers into strategic infrastructure partners whose commercial terms determine how much downstream platforms can own their AI stacks.