Google launches new Arm-based AI chip for data centers, challenging Amazon and Microsoft

On Tuesday, Google released new details about a new version of its data center artificial intelligence (AI) chips as well as an Arm-based core processor. Google’s tensor processing units (TPUs) are one of the only viable alternatives to Nvidia’s sophisticated AI processors (graphic processing units, or GPUs), but developers can only access them via Google’s Cloud Platform.


Google has unveiled Axion, its first Arm-based central processing unit (CPU) for data centers, which was developed using Arm’s Neoverse V2 architecture. Arm-based processors are typically a more cost-effective and energy-efficient alternative.

Axion is designed to deliver a 30% performance improvement over Google’s fastest Arm-based cloud tools and a 50% improvement compared to the latest, comparable x86-based chips produced by AMD and Intel. Moreover, Axion is claimed to be 60% more energy efficient than those same x86-based chips.
Google is using Axion in services like BigTable and Google Earth Engine. The Axion processor is built on open foundations and is designed to make it easy for customers to bring their existing workloads to Arm.

Why does it matter?

Rival cloud providers, like Amazon and Microsoft, have developed Arm CPUs to differentiate their computing offerings. Google had previously offered specialized chips for YouTube, AI, and its own devices, but not a CPU. Google’s entry into the Arm-based CPUs for data centers market would put it in direct competition with Amazon Web Services (AWS), which has been the leader in this domain since it released the Graviton processor in 2018. AWS has since launched a second and third iteration of its Graviton processors. In 2021, US giant chipmaker NVIDIA also introduced Grace, its first Arm-based CPU for data centers. Other companies, like Ampere, have also made strides in this area.

Following Amazon and Microsoft’s footsteps, Google will become less reliant on Nvidia and Intel while competing with them on AI chips and cloud workloads.