For most of its life, Arm sold the recipe rather than the cake. Chip designers licensed its instruction set, wrapped their own silicon around it and paid a royalty on every unit that shipped. Arm has now announced the Arm AGI CPU, a data center processor it designed itself and will sell as a finished part, which the company calls the first production silicon in its history.
Blame agents, or credit them. An AI agent is a model that does not answer once and stop. It plans, calls tools, waits for results and tries again, and something has to keep that loop organised. Accelerators generate the tokens. CPUs run the choreography, and choreography turns out to be core-hungry work.
Arm's release says data centers should expect to need more than four times their current CPU capacity per gigawatt as agent workloads scale. Data Center Knowledge, reporting from the Arm Everywhere event in San Francisco, puts a sharper figure on it from chief executive Rene Haas: roughly 30 million cores per gigawatt now, heading toward 120 million.
"AI has fundamentally redefined how computing is built and deployed. Agentic computing is accelerating that change," Haas said in the announcement. Strip the ceremony out of that sentence and it is a company explaining why licensing alone stopped being enough.
Specifications first, because they are the honest part. The AGI CPU packs up to 136 Arm Neoverse V3 cores, each fed at 6GB/s of memory bandwidth with sub-100ns latency, inside a 300-watt thermal budget. Every program thread gets a dedicated core, which is the silicon equivalent of refusing to seat two parties at one table, and Arm says it is there to stop the throttling and idle threads that make performance unpredictable under sustained load.
Density is where the pitch actually lands. A high-density 1U chassis reaches 8,160 cores per rack on air cooling alone, and liquid-cooled systems go past 45,000. TSMC is building the part on its 3nm process, which TSMC states in Arm's own release and Data Center Knowledge repeats.
Two larger claims deserve a harder look. Arm says the AGI CPU delivers more than twice the performance per rack of x86 CPUs and enables up to $10 billion in capital savings per gigawatt of AI data center capacity. Both carry an asterisk. At the foot of the release, that asterisk resolves to three words: "Based on estimates." No benchmark, no methodology, no third party named.
Meta is the lead partner and co-developer, pairing the AGI CPU with its own MTIA accelerators, and Arm says the pair are committed across multiple generations of the roadmap. Named customers include Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP and SK Telecom. ASRock Rack, Lenovo, Quanta Computer and Supermicro are building systems, with early machines available now and broader availability expected in the second half of the year.
Awkwardly, selling chips lines Arm up against customers who license its architecture to build their own, including AWS with Graviton, Microsoft with Cobalt and Google with Axion. All three sit on Arm's list of supporters anyway, alongside Broadcom, Marvell, NVIDIA, Micron, Samsung and SK hynix.
Haas told the keynote audience the data center CPU market could be worth more than $1 trillion to Arm through 2030, according to Data Center Knowledge, which reads as ambition rather than forecast. The nearer test is duller than that. An architecture licensor now has to ship firmware, drivers and support contracts on a schedule, and AGI CPU 2 is already slated for 2027.
Leave a Reply