Apple is scaling up the pre-training of its foundation models — the general-purpose AI systems that sit at the base of modern AI stacks — on Google's Tensor Processing Units (TPUs), placing some of the most compute-intensive work in its AI strategy on chips designed by one of its biggest rivals.

What Apple is actually doing

Pre-training is the heavy-lifting phase of building a large language model. It consumes vast amounts of compute, which makes the choice of hardware a strategic decision rather than a procurement detail.

Apple's hardware of choice is Google's TPUs — custom chips Google built specifically for machine learning workloads, rather than general-purpose graphics processors adapted for AI.

Apple's first AFM server model was trained on a cluster of 8,192 TPU v4 chips. The on-device model, the smaller version designed to run locally on Apple hardware, was built on a 2,048-chip slice of TPU v5p chips, a newer generation in Google's TPU lineup.

The performance gains were substantial. The move gave Apple more than double the floating point operations per second (FLOPS), along with triple the high bandwidth memory — allowing the chips to perform far more math per second while holding far more data close at hand. The result was training time for large language models cut by up to 2.8 times.

How Apple ended up here

The story starts in July 2024, when Apple revealed it had trained its initial Apple Foundation Models, known as AFM, on Google Cloud TPUs instead of the NVIDIA GPUs most of the industry was using.

Availability was part of the reasoning. The shift came amid NVIDIA H100 supply issues, combined with what the research describes as the superior performance and cost-effectiveness of Google's TPUs for Apple's needs.

In January 2026, Apple entered a multiyear agreement with Google for access to Gemini technology, a deal reportedly worth $1 billion annually. Then, in June 2026, Apple announced its third-generation AFM models, a release that marked custom development work done with Google.

Not everything runs on TPUs, however. One of the third-generation models, AFM 3 Cloud Pro, operates on NVIDIA Blackwell GPUs inside Google Cloud.

Apple has also been consistent on one point throughout: the company maintains that no private user data was used to train its foundation models, and it says it honors publisher opt-out preferences.

What this means

The most immediate takeaway is that Google has become a load-bearing pillar of Apple's AI effort. Between TPU-based training, Google Cloud infrastructure, and the Gemini licensing deal, Apple's AI roadmap now runs through Mountain View in several places at once.

AFM 3 Cloud Pro running on Blackwell GPUs also shows Apple remains invested in NVIDIA's ecosystem — just through Google Cloud, rather than as a direct anchor customer for its training runs.

From here, the details worth watching are concrete: how the multiyear Gemini agreement shows up in Apple's releases, and which chips — Google's or NVIDIA's — each successive AFM generation is disclosed to run on, a cadence that so far spans the July 2024 TPU disclosure and the June 2026 third-generation launch.