Ask most people what’s powering the AI boom and they’ll name a GPU maker. Fair enough, GPUs do the flashy work of training and running models, but every one of those GPU clusters sits behind a CPU that schedules the work, manages memory, and moves data around, and increasingly, that CPU carries an architecture most people associate with phones, not data centers.
That’s Arm’s story right now, and it’s worth understanding on its own terms, not just as another name in the AI chip rally, but as a genuine structural shift in who builds the infrastructure underneath AI.
The job nobody talks about
Training runs and inference queries get the headlines, but neither happens without an orchestration layer; the CPU work of scheduling, memory management, and increasingly, coordinating fleets of autonomous AI agents rather than single prompts. Arm’s own framing is that agentic AI is driving something like a four-fold increase in CPU core demand per gigawatt of data center capacity, since agents don’t just answer once – they run continuously, spawn subtasks, and keep hammering the orchestration layer in ways a single chatbot response never did.
Whether that specific multiplier holds up under real-world load is something only time and independent benchmarking will settle. But the underlying logic that agentic workloads are CPU-hungrier than the chatbot era, is a claim showing up consistently across hyperscaler earnings calls and product roadmaps, not just Arm’s own materials.
Arm are already inside the infrastructure – not just pitching it
Here’s the part that separates Arm from a typical AI hype story: it doesn’t need the future to arrive for this to matter, because it’s already embedded in production infrastructure today. Arm’s Neoverse compute platform underpins AWS Graviton, Google Axion, Microsoft Azure Cobalt, and Nvidia’s own Vera CPU – confirmed directly by each company, not just Arm’s telling of it. Google’s own engineering blog describes Axion as built “on our high-performance Arm Neoverse V2 platform.” That’s about as unambiguous as corporate confirmation gets.
As of Arm’s most recent public disclosure the company said it had shipped 1.5 billion Neoverse cores into data centers, with 500 million of those landing in just the prior nine months. That’s a meaningfully steeper curve than Arm’s own figures from earlier in the year, and it’s worth citing the current number rather than a stale one, since this is a business moving fast enough that six-month-old stats undersell it.
From licensor to chipmaker is a strategic pivot
For 35 years, Arm’s business model was licensing: design an architecture, let others build the silicon, collect royalties. That changed in March 2026 with the launch of the AGI CPU, Arm’s first-ever in-house production chip, co-developed with Meta as lead partner and purpose-built for agentic AI workloads. CEO Rene Haas called it “the next phase of the Arm compute platform and a defining moment” for the company, a framing Arm has stuck with consistently across its own newsroom announcements.
This is not a small pivot. It puts Arm in direct competition with Nvidia’s Vera, AMD’s EPYC Venice, and Intel’s Clearwater Forest – all companies Arm previously either licensed to or coexisted alongside. Early commercial traction has been real: what started as a roughly $1 billion revenue opportunity for fiscal 2027–2028 combined grew to more than $2 billion in customer demand within months, with Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP and SK Telecom joining Meta as launch partners, as per Arm’s own disclosures.
The Bottom Line
Arm’s relevance to AI data centers isn’t a story about a chip company riding a rally, it’s a structural one. It’s architecture already sits inside the custom CPUs of every major hyperscaler, it’s own first-party silicon is gaining real commercial traction, and the industry’s move toward agentic AI plays directly to arguments Arm has been making about CPU efficiency for years.
