Apple Unveils Conversational Siri at WWDC but Will Run Its Most Advanced AI Model on Nvidia GPUs in Google Cloud
Apple unveiled a conversational Siri and new cloud models at its Monday developers conference in Cupertino. Executives said the features run on Nvidia GPUs inside a privacy-focused extension of Private Cloud Compute.
CnbcApple demonstrated a redesigned Siri that can hold back-and-forth conversations at its annual Worldwide Developers Conference on Monday in Cupertino, Calif. In one demo the assistant checked concert dates, set a reminder to buy tickets, and provided directions to pick up a friend en route to the venue.
The company also confirmed that its most advanced model, Apple Foundation Model Cloud Pro, will run on Nvidia GPUs hosted in Google’s cloud as part of Apple’s Private Cloud Compute infrastructure.
This marks the first official acknowledgment that some Apple Intelligence features will operate on Nvidia chips. ” Apple is instead highlighting privacy advantages and on-device convenience, he said. Amar Subramanya, an Apple AI executive, stated that AFM Cloud Pro is comparable to Google’s Gemini frontier models.
Sebastian Marineau-Mes, Apple’s VP of software, said recent Nvidia technology called “ambiguous confidential compute” allowed the partners to configure the chips so they cannot read data on the servers. “We wanted to avail ourselves of the latest technology from Nvidia, and so we set out to extend private cloud compute to third-party cloud,” Marineau-Mes said.
Apple executives described a system orchestrator inside the company’s operating systems that routes each AI query to the appropriate model—on-device or in the cloud—based on required computing power and personal data.
The company is not using Google’s public Gemini service or its standard cloud infrastructure. Instead, Google’s technology helped train Apple’s own third-generation AFM models, which are custom-built for Apple Silicon and refined with outputs from Gemini frontier models, Subramanya said. The four models discussed are AFM Core, Core Advanced Cloud, and Cloud Image.
All are trained on proprietary data with reinforcement learning and designed to run on Apple chips, according to Subramanya. Apple said it collects less data than web-based services such as OpenAI’s ChatGPT or Anthropic’s Claude and uses locally stored information such as calendars and text messages to personalize responses.
