What Apple’s Siri U-Turn Means
Apple’s shift from proprietary Apple Intelligence infrastructure to a redesigned Siri running on Google Cloud with Nvidia Blackwell chips is a strategic reversal that reshapes how its voice assistant works, how user data travels, and how much control Apple keeps over its own AI stack. The new Apple Siri redesign moves Siri from a narrow assistant into a modern AI system that can summarize documents, control apps in sequence, and hold longer conversations. Apple will keep simple requests on-device using Apple Silicon, while complex prompts are routed to large cloud models. Earlier, Apple pledged that Apple Intelligence would run only on Apple Silicon and on its own servers, but performance limits and latency pushed the company to consider alternatives. The result is a hybrid design that trades some infrastructure purity for speed, scale, and competitive parity with other AI platforms.
From Apple-Only Servers to Google Cloud Nvidia
Apple originally tried to keep Siri’s advanced processing inside its Private Cloud Compute stack, running on Apple Silicon in Apple-controlled data centers. According to AppleInsider, Apple even attempted to run a version of Google Gemini under Private Cloud Compute, but the system was too slow to be usable. The Information reports that Apple’s internal cloud AI models are around 150 billion parameters, while the customized Gemini model it is now licensing sits near 1.2 trillion parameters and can respond much faster. Faced with this gap, Apple abandoned plans for fully proprietary AI servers for Siri and moved to Google Cloud Nvidia infrastructure. The upgraded setup uses Google’s data centers with Nvidia Blackwell B200 GPUs, hardware built for large-scale AI workloads that can deliver lower latency and higher throughput than Apple’s first-party alternative.
Inside the New Apple Siri Redesign
The Apple Siri redesign follows a hybrid architecture that splits work between the device and the cloud. Routine tasks—timers, quick settings changes, simple questions—stay on iPhone, iPad, or Mac, powered by on-device Apple Intelligence models. Heavier workloads, such as summarizing long emails or coordinating actions across multiple apps, are offloaded to cloud-hosted Gemini running on Blackwell chips. The Tech Portal notes that Apple is expected to use model distillation so its smaller in-house models can learn from Gemini’s responses and improve over time. Siri will also become more context-aware, drawing on Mail, Messages, Calendar, Photos, and Notes to answer in a more personal way. Conversations will feel closer to chatbots like Gemini or ChatGPT, with follow-up questions and persistent context rather than isolated commands, and this full rebuild is expected to debut around September 2026 alongside iOS 27.
Why Apple Chose Gemini and Blackwell Chips
Several factors pushed Apple toward Google Cloud Nvidia and away from an Apple-only stack. Gemini’s size and maturity offered an immediate route to match rivals’ AI capabilities, rather than waiting for Apple’s 150-billion-parameter models to catch up. Nvidia’s Blackwell B200 GPUs add another advantage: they are tuned for very large models, reducing latency for complex Siri queries. AppleInsider reports that Apple will enable confidential compute features on these GPUs, which keep data encrypted while it is processed. That helps preserve Apple’s privacy guarantees even when computation leaves the device. The Information also indicates Apple is paying for a customized Gemini variant, giving it more control over behavior and integration. Strategically, Apple is trading some vertical control for speed to market, aiming to deliver a competitive assistant while its own Apple Intelligence infrastructure matures.
Privacy, Control, and What Comes Next
Federighi previously said Apple Intelligence cloud processing would run only on Apple servers for privacy and security, a promise now complicated by the move to Google Cloud Nvidia infrastructure. In the updated model, requests that exceed on-device limits are sent to Google’s data centers, where confidential computing on Blackwell chips keeps them encrypted during processing. Apple’s existing rules—such as not retaining prompts for training—are expected to carry over, but AppleInsider notes it is not yet clear how far these protections extend beyond encryption. This reliance on third-party infrastructure raises questions about long-term control and trust, even as Apple keeps its privacy branding. The redesigned Siri is expected to launch around September 2026 with iOS 27, and its success will hinge on whether users accept this mix of Apple Intelligence infrastructure, Google Cloud, and Nvidia hardware as both fast and safe.
