M.G. Siegler • #ai • Jul 14, 2026

Apple Going All-In on AI Chips

Might they too challenge NVIDIA one day, at least on inference?

While the high-level itself is interesting – that Apple is seemingly doubling and tripling down on the AI aspects of their Apple Silicon efforts – one part of Gurman's newsletter this week stood out:

Apple had been planning major neural-processing upgrades for the M7 family and ultimately decided those improvements were important enough to justify accelerating the next generation rather than completing the M6 lineup. Those changes go into high gear with the M7 Ultra. I’m told the processor dramatically upgrades AI performance, bringing it closer to the class of dedicated AI accelerators such as Nvidia Corp.’s Blackwell.

Uh, is this suggesting that Apple is about to make a chip that's comparable to the chips that are fueling the Age of AI and have turned NVIDIA into the most valuable (and profitable and perhaps important) company in the world? Well, no, not exactly.

If I'm reading Gurman's (too-simplistic) line correctly, I believe it just suggests that the M7 Ultra may have performance that's comparable in some ways with NVIDIA's Blackwell. Which chip? Unclear, there's the B100 and B200 model, as well as other variants like the 'GB 200 Superchip' which is a GPU paired with a 'Grace' CPU. It's undoubtedly not about to touch the actual Blackwell racks that NVIDIA sells – which run from hundreds of thousands to millions of dollars – but still, just matching the chips would be impressive for AI workflows.

Specifically, he also undoubtedly means inference (seemingly hence "accelerators") and not training of AI models, as again, there's almost no way for a "single" chip (even two or more fused together as Apple does with their 'Ultra' models) to match the processing power of what NVIDIA's systems can offer. The power required alone (see also: every headline about data center power consumption) would make this a completely unfair comparison. As would the current state of memory bandwidth between the two types of chips – which stands at about an order of magnitude difference, which also obviously matters for token generation, I might add!

That said, because Apple's chips use unified memory, they actually do have a potential inherent advantage here for running AI models. The problem has been that the amount of memory they utilize is just so much smaller than what NVIDIA's overall systems call to, so it has historically only been good for smaller, local models (see also: Apple's Siri AI hybrid strategy). But if the M7 Ultra is able to jack up available memory to 1.5TB as the story mentions, that could very well compete with a Blackwell when in comes to inference for LLMs.

But even with the current memory crunch issues aside – and those may very well not even be alleviated by the time such a chip is ready – the huge caveat here is that this M7 Ultra chip is slated to launch in 2028. Blackwell is doing all of this in 2026 and it's not like NVIDIA is sitting still. They already have their next generation of chips seemingly ready to roll and then will presumably have a generation after that ready to roll too. Further, the AI labs are also not sitting still and so today's state of the art models are likely to be quite a bit larger in 2028 too...

Anyway, my point is that this almost throwaway sentence sounds insanely impressive – and it would be in many ways – but there's also likely quite a bit of nuance to it.

Still, if Apple is this focused on ramping up AI capabilities for their chips, it is fair to wonder if this won't be yet another huge advantage the company has in the future thanks to Apple Silicon. Again, Apple is clearly already focused on the hybrid approach of doing some AI work on-device while offloading the bigger stuff to the cloud. If they can bring more and more of that work on device...

And that mixed with the current trend of companies balking at the AI compute costs and mixed with the notion that perhaps some of their frontier models are increasingly going to be paired (or distilled) down because they're overkill for a lot of AI workloads... and yeah, Apple could be sitting pretty in a few years.

Perhaps also why NVIDIA is building their own AI PC chips in the form of 'RTX Spark' alongside Microsoft. The battleground changes but the combatants remain the same...

One more thing: the above is obviously all about computer-grade chips, but what about servers?

The chip may also ultimately be the basis of a coming overhaul to the AI server strategy. Apple plans to soon deploy a more powerful server based on the M5 Ultra under the internal code name J246, but engineers are already developing another new server chip for launch by 2029 that is built around the M7 Ultra’s capabilities.

This, of course, points to a world in which Apple doesn't need to rely on Google Cloud (running NVIDIA chips, no less) for that offloaded work. But might it also point to another world in which Apple, like Google and Amazon before them, start selling their chips to others to use? Or what about their own actual 'Apple Cloud' business? Let's not get ahead of ourselves, but everyone is doing it these days...

👇

Previously, on Spyglass...