Could the man who built Google’s Gemini Assistant be the one to save Siri? That’s the bet Apple is making in replacing John Giannandrea, its AI chief since 2018, with Amar Subramanya-a veteran of both Google and Microsoft-in what insiders see as the most consequential leadership change in its AI division since the launch of Apple Intelligence.
The timing is telling. Apple Intelligence, positioned as the generative AI showpiece for the company when it was announced last June 2024, has managed a string of public misfires since its rollout last October. Where the notification summary feature was supposed to summarize multiple alerts into a tidy update, it generated a string of glaring factual errors. It falsely claimed in late 2024 and early 2025 that Luigi Mangione, accused in the killing of UnitedHealthcare CEO Brian Thompson, had taken his own life, and that darts player Luke Littler had won a championship before the final match even started. The BBC lodged formal complaints over the two incidents, underlining the reputational risk of flawed AI outputs.
The more damaging blow came from Siri’s much-hyped overhaul. Internally, the project had been meant to showcase Apple’s leap into conversational, context-sensitive assistants-but weeks before its planned April 2025 debut, software chief Craig Federighi found that many of the flagship features simply didn’t work on his own iPhone. The launch was shelved indefinitely, prompting class-action lawsuits from iPhone 16 buyers who’d been promised an AI-enhanced Siri. By March, Tim Cook had already stripped Siri oversight from Giannandrea, assigning it to Vision Pro creator Mike Rockwell, and removed the company’s secretive robotics division from his control.
Organizational fractures added to the leadership turmoil. A Bloomberg investigation detailed a communications breakdown between AI and marketing teams, budget misalignments, and morale collapse so severe that some employees derisively referred to the group as “AI/MLess.” Talent losses mounted, with researchers defecting to OpenAI, Google, and Meta.
Now, in a move that would have been unthinkable a decade ago, Apple is reportedly integrating Google’s Gemini models into the next version of Siri. For a company that has spent more than 15 years in fierce competition with Google across mobile OS, app stores, browsers, maps, and cloud services, the decision is as much a technical concession as a strategic one. Gemini’s architecture is optimized for multi-modal reasoning and large-context processing-capabilities far beyond what Apple’s current on-device models can deliver.
That gap, in essence, reflects Apple’s privacy-first AI philosophy: While competitors like Microsoft and Google have invested billions in hyperscale AI data centers and frontier models, Apple has focused on running AI workloads directly on its devices using Apple Silicon. The company’s neural engines can deliver impressive low-latency inference for compact models, and when tasks require cloud processing, Apple deploys Private Cloud Compute-servers designed to process data ephemerally and delete it immediately. This minimizes user data collection, but sets hard limits: on-device models are smaller, less capable, and trained on licensed or synthetic datasets rather than the massive real-world corpora powering competitors’ systems.
The trade-off is stark. Large language models like Gemini or GPT-4o can avail themselves of hundreds of billions of parameters, trained on diverse, high-volume datasets, affording rich reasoning and flexible dialogue. By contrast, Apple’s models must hew to the memory and compute envelope of iPhones, iPads, and Macs-less than 10 billion parameters-which sharply limits their ability to generalize across complex queries. That’s why Apple’s Siri upgrade, despite years of engineering, still trailed competitors in contextual understanding and multi-step task execution.
Subramanya’s arrival suggests a potential recalibration. At Google, he led engineering for Gemini, a project that integrated large-scale transformer architectures into consumer-facing assistants. At Microsoft, he oversaw AI work that bridged cloud-scale models with enterprise-grade privacy and compliance-a skill set directly applicable to Apple’s tightrope walk between capability and privacy. Now, he inherits the reins for Apple’s foundation models, AI research, and safety, reporting to Federighi, who has taken a more hands-on role in AI development.
Apple has already increased its AI spending and struck deals to integrate external models, such as ChatGPT, into Siri, but it remains to be seen if the company can square its privacy constraints with the performance benchmarks set by cloud-first competitors. The next version of Siri-powered in part by Gemini-will be the first real test of whether Apple can close that gap without abandoning the principles that have defined its AI strategy.
