Hartley Charlton, reporting for MacRumors:

Apple used Tensor Processing Units (TPUs) developed by Google instead of Nvidia’s widely-used graphics processing units (GPUs) to construct two critical components of Apple Intelligence.

The decision is detailed in a new research paper published by Apple that highlights its reliance on Google’s cloud hardware (via CNBC). The paper reveals that Apple utilized 2,048 of Google’s TPUv5p chips to build AI models and 8,192 TPUv4 processors for server AI models. The research paper does not mention Nvidia explicitly, but the absence of any reference to Nvidia’s hardware in the description of Apple’s AI infrastructure is telling and this omission suggests a deliberate choice to favor Google’s technology.

Nvidia and Apple’s kerfuffle runs back to between 2007 and 2008 when Apple shipped Nvidia graphics processors, specifically the GeForce 8600M GT, in MacBook Pro models. Those graphics cards were defective and would stop functioning after a few months of normal usage, which led to a class-action lawsuit against Apple for shipping faulty products to buyers. Apple apologized and set up a repair program for affected customers to receive a repaired computer free of charge, but it wanted Nvidia to finance it, since, at the end of the day, it was Nvidia’s fault the graphics cards were defective. Nvidia refused to pay Apple back, and so, in 2012, Apple stopped shipping Nvidia cards in any of its products. That was the end of that relationship — it has never been repaired since.

One complication in this otherwise severed relationship was that Nvidia launched Omniverse Cloud application programming interfaces on Apple Vision Pro in March, which was the first time the two companies ever worked with each other in more than a decade. Still, though, Apple and Nvidia arguably hate each other and aren’t on speaking terms after this (relatively minor) disagreement from a while ago. It’s just like Apple and Intel’s once-great relationship that turned sour after the launch of Apple silicon, but that one is understandable since Intel lost one of its most valuable clients, if not the most valuable.

Apple makes the best computers on the market, but before it switched to Apple silicon, it used GPUs from Advanced Micro Devices, Nvidia’s biggest competitor. This made gaming performance on the Mac suffer immensely, but it wasn’t that big of a deal for Apple, since game developers had already deprioritized the Mac since its user base is less gaming-inclined. But now, gaming aside, Nvidia makes the best artificial intelligence processors, and every AI firm is buying up its entire stock of H100 processors — more than it can even make. Microsoft and Google know this, which is why they’re building their own processors to try and compete, but the mix of proprietary software that runs on Nvidia’s AI chips and the sheer grunt of the processors still makes them the best. Still, though, interested firms can rent out Azure or Google Cloud neural processing units, as they’re called, directly made by one of the two companies without involving Nvidia.

Apple entered the AI arena later than most, but a few months ago, it found itself needing to train its own set of models for Apple Intelligence — and it could choose any processors it wanted. And, in the end, it opted for Google’s processors, hosted in the cloud, with no help from Nvidia. Google sells access to its NPUs — called “Cloud Tensor Processing Units,” the same ones it uses to train Gemini, its AI product — to anyone via Google Cloud, but I assume it cut Apple a deal since the two companies already have a contract to share search revenue on the iPhone. Google and Apple technically aren’t enemies, but they’re also not friends, and now they’re competing in the hottest market of the year: AI. Google has a vested interest in making Gemini better than Apple Intelligence because it has the power to sway markets and put Google back at the top financially again, but it decided to lend Apple a hand in training its models, for some reason — probably monetary.

Obviously, the most shocking deal would be if Apple hosted the end-user models on Google’s servers, which I assume Google would object to, even for an enormous sum of money. But that wouldn’t be favorable for Apple, either, since one of its biggest selling points is privacy via Private Cloud Compute, only possible with Apple silicon. Why Apple didn’t train Apple Intelligence’s foundation models, as it calls them, on Apple silicon from the get-go is unclear, but it’s most likely because it isn’t powerful enough. The more powerful an NPU is, the more complex and accurate a large language model can be, which affects how precise inference — the process of predicting the next token in a sequence — is. Thus, if Apple trained Apple Intelligence with less performant NPUs, it would negatively affect the performance of the models on the end-user side. It could choose to do so just to satiate its own ego, but that’s a bad trade-off.

So, to recap: Nvidia makes the best NPUs, but Apple hates Nvidia, so it was between Microsoft and Google — and since it was already on good terms with the latter, it trained its LLMs on Google’s servers for whatever sum of money the two corporations agreed on. It’s not that unusual once the chain of events is broken down, but from afar, it really does look peculiar. Why would Google give its computing power to its direct competitor? But it actually isn’t that odd upon close examination because companies do this all the time; Apple buys displays from Samsung, even though that same technology could be used in Samsung Galaxy phones. (In some cases, the same screens are used in computing products, like the Google Pixel.) It’s unusual, but not unheard of. Samsung makes the best displays, and Google makes the best NPUs — aside from Nvidia, of course.