Apple may be shopping for AI chip companies; Here's why

Apple is reportedly exploring plans to acquire companies that design AI chips to manage demanding AI tasks and reduce its dependence on Nvidia. Nvidia’s GPUs are widely used to train and run advanced AI models, and power AI infrastructure for major tech companies.

Apple already performs some AI processing in its own data centres, although for heavy-duty AI workloads, it heavily depends on Nvidia GPUs, which are hosted on Google Cloud. Apple is also rumoured to develop its own AI server chip, internally known as Baltra, which was expected to launch this year. However, the launch has been delayed.

Must read: Warren Buffett changed his mind twice on Big Tech: First Apple, now Alphabet

Apple AI startup acquisition plans

According to The Information report, Apple has started to explore several semiconductor startups for potential acquisition. Over the years, Apple has acquired smaller startups; recently, it acquired Israeli AI startup Q.ai for nearly $2 billion earlier this year, and it is also reported to be in talks with PrismML. 

Now, the tech giant is reportedly planning larger acquisitions to strengthen its AI capabilities. Currently, Apple is facing performance limitations with its own AI server infrastructure. It uses the in-house M2 Ultra chips, but it not powerful enough for demanding AI workloads compared with specialised AI accelerators.

Must read: OpenAI rejects Apple's trade secret theft claims, says 'no evidence' supports lawsuit

For this reason, Apple depends on external infrastructure for its AI tasks. It is also reported that Gemini-powered Siri handles demanding AI requests using Nvidia GPUs hosted on Google Cloud.

Furthermore, Apple is also targeting AI startups that specialise in model compression, a technology that enables advanced AI models to run more efficiently on devices such as the iPhone. 

Bloomberg also reports that Apple is building an AI server chip based on the M5 Ultra, while an upcoming M7 Ultra chip is expected to offer major AI performance gains. The new generation is said to support up to 1.5TB of memory, though a server variant is not expected until 2029.