Apple reportedly pays Google $1 billion a year for cloud infrastructure powering Siri's new AI model

Knowledge distillation is what Apple bought. Techtimes reports that Apple ran large volumes of queries through Google's Gemini, captured both the answers and the reasoning steps behind them, and used that output to train smaller Apple Foundation Models ^[5]; Gemini is reportedly trained on roughly 1.2 trillion parameters ^[6]. On a supported iPhone, the reply usually comes from a model running on the device's own Neural Engine ^[7]. Bloomberg reported the multi-year Google agreement is worth approximately $1 billion annually and is structured as a cloud computing contract ^[9]. Apple said it chose Google's technology after evaluating several options ^[10].

Queries the local model cannot handle go up to Private Cloud Compute ^[11]. Apple said some Siri AI and Apple Intelligence features that use its cloud servers will be subject to daily usage limits, and that expanded limits will "be available for a fee in the future" ^[13]. "Daily limits may vary by feature, request complexity, system demand, system policies, and other factors," Apple said ^[14]. Since local answers never touch those servers, the cap and the eventual charge apply to the overflow ^[15].

The two accounts of what those servers are do not agree. Techtimes says Private Cloud Compute runs on Apple Silicon, with personal identifiers stripped before processing ^[11]. CNBC says Private Cloud Compute uses Google Cloud infrastructure with Nvidia and Intel chips ^[12]. One version has Apple owning the escalation path; the other has it renting that path from the company that trained the teacher.

Five conditions sit between installing iOS 27 and holding a conversation with the new Siri: hardware from the iPhone 15 Pro up, a region outside the EU and China, English, an opt-in in the settings app, and a place off the waitlist ^[16]. iOS 27 installs on iPhone 11 and newer ^[2], so the 11, 12, 13 and 14 lines take the update without the feature ^[17]. The larger on-device model, which handles enhanced dictation and Siri customisation, runs only on the iPhone 17 Pro, iPhone 17 Pro Max and iPhone Air, the three current models with 12 GB of RAM ^[18]. Apple did not say what the waitlist is rationing.

Apple bought teacher output and kept the runtime. That is cheaper than licensing a frontier model, and it is still a dependency: under the agreement announced earlier this year, Gemini models and Google cloud infrastructure underpin the next generation of Apple Foundation Models ^[8]. Engadget's reviewers described Siri AI as markedly more capable of answering complex questions and carrying out multi-step prompts than its predecessor ^[19], and the limit they found was a delivery email Siri could not retrieve because the thread lived in Gmail rather than Apple Mail ^[20].

The next training cycle settles how much of the model layer Apple has handed over. If the generation after this one is trained without Gemini output, the roughly $1 billion a year was a purchase of curriculum; if the contract renews and grows, Apple is paying a rival, in perpetuity, for the layer beneath its own weights. "It's personal intelligence that's actually personal," Apple CEO John Ternus said at last week's iPhone launch event, his first since succeeding Tim Cook on Sept. 1 ^[21][22]. Apple first revealed this Siri in June, after a version announced in 2024 was delayed last year ^[23], and the software landed days before the iPhone 18 reaches stores ^[24].