Apple Outsources Siri Cloud Workloads to Nvidia Blackwell GPUs
In a significant departure from its historical strategy of vertical integration, Apple will rely on Google Cloud's fleet of Nvidia Blackwell B200 data center chips to power its major Siri overhaul launching in September 2026.
Apple had previously attempted to run a modified version of Google's Gemini model on its in-house "Private Cloud Compute" server system, which is powered by custom Apple Mac-series silicon. However, Apple's internal hardware proved too slow for the demands of the revamped Siri. To solve this bottleneck, Apple turned to Google Cloud's infrastructure, utilizing Nvidia's advanced Blackwell B200 GPUs. This infrastructure represents a major component of the massive hyperscaler capex boom (see Hyperscaler Capex Surge: The $690 Billion AI Infrastructure Sprint). To address privacy concerns, Apple will utilize Nvidia's hardware-based confidential compute feature to encrypt user data during processing.
This transition highlights the massive, parabolic demand for Nvidia's chips as the industry shifts toward real-world, consumer-facing agentic AI (see Nvidia's Record Q1 FY2027: Parabolic Demand Driven by Agentic AI).
Key Quotes
"Apple will rely on Google's fleet of Nvidia chips to power its overhauled version of Siri when it launches in September, according to a new report from The Information." — Tim Hardwick, MacRumors
"Apple reportedly tried to get a modified version of Gemini working on its in-house server system, but found that it ran too slowly." — Tim Hardwick, MacRumors
Interpretation
This development is a massive validation of Nvidia's hardware moat. Apple is a company famously obsessed with controlling its entire hardware and software stack, and it has spent billions developing its own silicon. Yet, when faced with the real-world performance requirements of consumer-scale, low-latency AI inference, Apple had no choice but to bypass its own silicon and rent Nvidia's Blackwell chips through Google Cloud. This highlights that Nvidia's technology is essential not just for training, but for the most prominent consumer-facing inference workloads in the world.