Apple confirms frontier AI workload will run on Nvidia GPUs in Google Cloud

Apple is moving part of its most advanced AI work off its own compute systems and onto external hardware and cloud services, a move that indicates how demanding “agentic” reasoning can be even for a company known for vertical integration. Reporting that cites Apple software chief Craig Federighi says Apple’s frontier language model, FM Cloud Pro, will run on Nvidia GPUs hosted in Google Cloud.

In the account, Federighi’s confirmation came in the context of Apple scaling agentic workflows, which are AI tasks that break problems into multiple steps and decide what to do next as they go. Such workloads often require sustained compute at inference time and can push infrastructure requirements well beyond simple, single-turn responses.

The same reporting describes Apple as having engineered an M-series-based private cloud to keep AI compute in-house, reflecting Apple’s long-running privacy narrative for on-device and managed systems. But as the AI system shifted toward more complex, multi-step experiences, Apple reportedly found that its private setup could not sustain the needs at scale.

Instead of abandoning the broader architecture, Apple appears to be keeping the private cloud running while relocating specifically the “frontier” workload to Google Cloud. That division matters because it preserves the idea that Apple can still control much of its AI environment, while acknowledging that the most resource-intensive components may require the scale and specialization of major AI infrastructure providers.

The reporting also frames the decision as a competitive positioning move. It says FM Cloud Pro is designed to be on par with Google’s frontier models, a comparison that implicitly raises the bar for latency, throughput, and quality, since frontier models generally require large-scale training and heavy inference compute.

Apple did not issue a separate public statement in the materials reviewed here, and the details of what exactly will run where, such as which Nvidia GPU configurations are involved and whether data handling and retention rules differ from Apple’s private cloud, were not disclosed in the cited post. As a result, it remains unclear how Apple’s privacy controls will be implemented for the external-GPU portion of the workload.

The episode lands in a broader pattern across the tech sector, where even companies that build their own chips and infrastructure increasingly rely on hyperscalers for peak capacity. For Nvidia and Google, the development underscores how demand for frontier AI expands beyond training into scaled deployments, and how Nvidia’s GPU ecosystem continues to be a common building block across cloud providers.

For watchers of Apple’s AI strategy, the next question is whether the company will expand FM Cloud Pro’s cloud usage as capabilities grow, or whether it will attempt to bring more of the workload back in-house. Investors and engineers will also likely look for indicates about performance, cost structure, and the operational boundaries of Apple’s agentic systems.

In the absence of additional primary disclosure from Apple, the immediate takeaway is directional: the most advanced, agentic portion of Apple’s frontier AI is being run using Nvidia GPUs in Google Cloud, even as Apple retains a private infrastructure for other components.