Arm has announced a trio of products across mobile, robotics, and data centre chips. And wherever AI compute goes next, Arm is positioning itself to be the default layer underneath it.
The long game
Until recently, when you asked an AI assistant something, whether that was via your phone or laptop – the device itself was essentially acting as a messenger, taking your question and sending it to a giant data centre somewhere. A huge model then found the answer. The answer then comes back to the device. Your device, however, did practically none of the actual thinking.
But things are changing.
Agentic AI is taking devices from simple prompt fetchers and turning them into something wholly more … alive. These are autonomous, multistep agents. Compute is coming out of the Cloud and into phones, and eventually, into the machines that act in the physical world.
Already, AI agents can complete multi-step tasks on your behalf. For example, booking a restaurant, checking your calendar, checking flights, coordinating with other apps. But that takes a lot more back-and-forth, a lot more decision-making, and a lot more “housekeeping”, which is the legwork of fetching data, checking results, calling other tools, or trying again if something fails. Housekeeping work is expensive, and it can be slow if it needs to go on a round-trip to a data centre somewhere, and then come back again.
When you add all that up, does it make sense to send everything to the Cloud? Because cost, speed, and privacy are the three factors that determine where AI belongs.
“Intelligence is created in the Cloud. It’s becoming personal at the Edge and increasingly physical. And Arm spans all three,” says Arm CMO, Ami Badani.
Arm’s chips already sit inside phones (the Edge), where they are the dominant player. They’re quickly gaining traction inside Cloud data centres, and their low-power designs are already inside billions of sensors, emerging robotic and autonomous systems, and industrial devices (the physical world).
Fortunate then, that as computing work spreads out from the Cloud and onto the devices themselves, Arm doesn’t need to break into new markets; the market is moving towards Arm.
The announcements
Phones, tablets, and laptops at the Edge – an “AI-native” upgrade
Arm has released CSS for Mobile 2, a new AI-native compute platform for phones, tablets, and laptops built around agentic AI and cinematic graphics from the ground up. It includes a new graphics chip (Mali G2-Ultra NX) which uses AI to render game visuals more efficiently, alongside a new C2 CPU cluster tuned for running AI agents on the device itself. Arm says the new GPU delivers up to a 4x improvement in graphics performance-per-watt generation-on-generation, plus a big reduction in DRAM power use. By handling more AI locally, phones can also gain speed without draining the battery.
“The costs, both economics, but also the latency and the personalness, drives an interest in having that compute as local as possible,” says Chris Bergey, EVP, Edge AI. “It just absolutely pushes the envelope [on] things that have never been possible to do in a mobile handset before.”
Preparing for physical AI as an industry standard
Arm is looking to do for robotics what it has already done for phones, and, more recently, for Cloud servers, which is to get in early, help define the shared language and building blocks that industry uses, so that when the market matures, Arm’s chips are already the default choice.
During the briefing, the company announced ‘Arm Total Design for Physical AI’, which is an ecosystem programme of more than 80 partners, plus a manifesto in the pipeline that proposes a shared framework for describing robot capability levels.
“We believe [physical AI] is going to be one of the biggest markets in the history of computing, if not the biggest market in the history of computing,” says Drew Henry, EVP, Physical AI. According to Henry, Arm estimates that the current physical-AI compute market is roughly $25 billion a year, and it is projected to grow to more than $200 billion a year sometime in the 2030s.
Talking about the standards push, Henry says: “We don’t have a similar thing in the robotics space … we are now going to do a call-out to the industry to help develop a common language.”
The Cloud
Even though attention is moving toward the Edge and robotics, AI agents still run in the Cloud. It’s the CPU doing the coordination work of managing agents. And so Arm has announced Neoverse CSS N4, a new, more customisable compute subsystem for data centre chips. Arm says it delivers more than twice the socket-level performance of its predecessor, CSSN3 along with 25% better performance-per-watt and 75% more memory bandwidth.
“Agentic AI is … changing the role of CPUs more broadly … it’s actually stressing multiple different aspects of the system,” says Mohamed Awad, EVP, Cloud AI.
According to Awad, Arm shipped 500 million Neoverse chip cores in the last nine months alone, versus 1.5 billion cumulatively since the platform launched in 2018 – a clear indication of a sharp acceleration.
When these three announcements are taken together it indicates that Arm is positioning itself to become the default layer everywhere that AI goes next. Not just in servers, but in your pocket, and in the machines that move, physically, through the world.
