Apple disclosed that the infrastructure behind its first Apple Intelligence models relied on Google-designed tensor processing units rather than Nvidia’s dominant AI accelerators. The details appeared in a research paper describing the hardware and software used to train models for features planned across the iPhone maker’s products.
The paper did not say that Apple had formally rejected Nvidia. Apple also declined to comment on the absence of Nvidia hardware from its description. Even so, the scale of the Google TPU deployment showed that Apple had built a significant part of its AI programme on an alternative chip and cloud ecosystem.
That choice mattered because Nvidia had become the default supplier for much of the AI industry. Apple’s use of Google infrastructure demonstrated that a large technology company could pursue advanced model development without making Nvidia GPUs the foundation of every workload.
It also made the infrastructure behind Apple’s consumer-facing AI strategy visible.
Apple Used Two Large Google TPU Clusters
Apple’s researchers described using two types of Google TPU in large clusters. A system intended to support a model operating on iPhones and other Apple devices was trained with at least 2,048 TPUv5p chips. Work on the larger server-side model used at least 8,192 TPUv4 processors.
Those numbers revealed a substantial infrastructure commitment. They also showed that Apple was developing a split AI architecture: some capabilities could be associated with devices, while more demanding processing could be served from larger systems. Both parts depended on access to specialised training hardware.
The research indicated that Apple’s engineers believed the same Google chips could support even larger and more complex models. The disclosed clusters were therefore not necessarily a one-off experiment. They could provide a foundation for later generations of Apple Intelligence as the company expanded model capability and product coverage.
Google’s Cloud Became Part of Apple’s AI Stack
Google’s TPU strategy differed from Nvidia’s approach in an important way. Nvidia sold GPUs that customers could deploy through multiple cloud providers or in their own infrastructure. Google offered TPUs through Google Cloud, meaning customers had to use Google’s platform to gain access to the processors.
Apple’s hardware decision consequently carried a cloud decision with it. Training on TPUs tied part of the model-development process to Google’s infrastructure and software environment. The arrangement gave Google a role in the underlying stack even as Apple presented Apple Intelligence as a distinctive layer within its own devices and operating systems.
Using a second source could also reduce dependence on Nvidia at a time when demand for AI accelerators was exceptionally high. The source characterised the choice as a strategic alternative, although Apple itself did not say that avoiding Nvidia was its objective. The evidence supported a narrower conclusion: Apple had demonstrated that Google TPUs could handle training at the scale its initial AI services required.
Apple Intelligence Was Already Reaching Developers
The infrastructure disclosure arrived as Apple began distributing Apple Intelligence capabilities through a developer beta. Early features included Writing Tools, a version of Siri able to understand typed as well as spoken requests, email summaries and Smart Replies in Mail, and AI-assisted transcription.
The beta programme gave app developers an early view of forthcoming software and allowed Apple to collect real-world feedback before a broader release planned for the autumn. That process could help the company refine individual features, improve its ChatGPT integration and prepare Siri to interact with third-party applications.
Apple had announced a wider set of AI plans at its annual developer conference, including the integration of OpenAI’s ChatGPT into its software ecosystem. The combination illustrated that Apple’s strategy involved several partners: Google supplied training infrastructure, OpenAI contributed an external model service, and Apple controlled how the resulting capabilities appeared to users.
The Move Challenged Assumptions About Nvidia’s Hold
Nvidia remained the clear leader in AI chips, and the source did not suggest that Apple’s choice immediately threatened that position. Nvidia hardware powered much of the cloud AI market, while Google, Amazon and other providers together represented a smaller alternative segment.
Nevertheless, Apple’s decision was strategically meaningful. A customer of Apple’s scale could validate a competing accelerator, encourage more investment in alternative hardware and show other buyers that advanced model training was not restricted to one supplier. For Google, leasing TPUs through its cloud created a route to participate in AI infrastructure without selling the chips independently.
Investors showed little immediate reaction. Apple’s shares fell only 0.1% to $218.24 during the trading session cited by the source. That muted movement suggested that the market was waiting to see how Apple’s deliberate approach translated into polished products and user adoption.
The lasting significance was not that Apple had definitively abandoned Nvidia, because the company had made no such declaration. It was that Apple built consequential AI models on Google silicon at a scale large enough to support consumer features. In a market often described as dependent on Nvidia, that was practical evidence of another route.
