Apple eyes AI chip start-ups to reduce reliance on NVIDIA
What's the story
Apple is on the lookout for chip start-ups to bolster its artificial intelligence (AI) server capabilities, according to The Information. The tech giant has been in talks with investment bankers and semiconductor start-ups as part of this exploration. The move comes as Apple looks to reduce its dependence on NVIDIA, a leading provider of AI server hardware.
Processing shift
Current AI processing setup
Apple currently handles some of its AI processing in-house, using custom-designed chips. However, for more demanding tasks, it still relies on NVIDIA hardware hosted in Google Cloud.
This arrangement is said to power the Google Gemini models used by the revamped Siri.
The company's own AI server chip, internally codenamed Baltra, was expected to ship this year but has reportedly been delayed.
Acquisition evolution
Shift in acquisition strategy
Apple has historically kept its acquisitions small, usually in the hundreds of millions. However, this strategy seems to be changing.
Earlier this year, Apple acquired Israeli start-up Q.ai for nearly $2 billion. The deal is second only to Apple's $3 billion acquisition of Beats Electronics in 2014.
During its last quarterly earnings call, CFO Kevan Parekh announced that the company would no longer target a "net cash neutral" position, possibly paving the way for larger acquisitions.
Team focus
Limitations of Apple's custom chips
Apple's chip design team has mostly focused on battery-powered mobile devices, not high-performance server processors.
This limitation was evident during the development of the revamped Siri, when engineers tried to run Google Gemini models on Apple's own server infrastructure.
The chips, designed for Mac workloads, reportedly couldn't handle AI models of that scale.
As a result, parts of Siri's AI workloads were processed using NVIDIA GPUs in Google's cloud infrastructure.
Future plans
In-House AI server chip development
Apple's in-house efforts are ongoing, with Bloomberg reporting that the company is developing an AI server chip based on the M5 Ultra.
A future M7 Ultra is expected to deliver a significant leap in AI performance, possibly bringing it closer to NVIDIA's Blackwell architecture.
The M7 Ultra is said to support up to 1.5TB of memory, twice as much as the M5 Ultra's capacity.
However, a server chip based on this technology may not be ready until 2029.