Siri’s AI overhaul runs partly on Nvidia chips sitting inside Google’s cloud. Apple tried running Gemini on its own server silicon first, and the chips couldn’t handle it.
They were built for Mac workloads, not a model this size, so Apple had to route around its own hardware to ship its own product.
That’s worth sitting with for a second. This is the company that built its entire identity around owning the stack from top to bottom.
The schedule makes it more uncomfortable. Baltra, Apple’s dedicated AI server chip, was supposed to ship this year. It’s late.
The M5 Ultra-based server chip is still being built. The chip, actually designed to go toe-to-toe with Nvidia’s Blackwell, the M7 Ultra, isn’t expected before 2029.
Apple is shopping for AI chip companies right now because its own answer to Nvidia is roughly three years out.
John Ternus takes over as CEO in September. Johny Srouji just picked up expanded control over Apple’s hardware engineering, chips included. Whatever gets acquired here, whatever gets built next, lands on their desks first.
They’re inheriting a chip roadmap that’s publicly behind on the one workload that matters most for the next decade of Apple’s products.
Apple has done this before. PA Semi cost $278 million in 2008 and quietly became the foundation for every custom chip Apple has shipped since.
Nobody wrote urgent headlines about that deal at the time. It just worked slowly in the background until it no longer looked optional.
Maybe this plays out the same way. Whatever Apple buys now could look like the obvious move in five years.
PA Semi happened while Apple had time. Nobody else was shipping a competing chip roadmap on a six-month cycle back then.
This time, Apple is buying its way into a race that’s already underway, against a company with a multi-year head start and a market that isn’t waiting around for anyone to catch up.
Kevan Parekh dropped the “net cash neutral” language on the last earnings call, which is Apple quietly saying it’s ready to spend more than it used to.
So, the money’s there, but the timeline isn’t any shorter for it, and 2029 is a long time to lean on a rival’s cloud to run your own AI assistant.
