At its Arm Everywhere event in San Francisco, ARM (NASDAQ:ARM) outlined a shift in strategy that executives described as a response to rapidly changing demands in data centers, particularly as AI workloads become more “agentic.” The company used the event to highlight the scale of its existing footprint, discuss the evolution of its product packaging approach, and announce what it described as its first revenue-generating CPU sold as a finished chip: the Arm AGI CPU.
Arm highlights scale and its “ecosystem of ecosystems”
Arm executives opened by emphasizing the breadth of the company’s installed base, saying more than 350 billion Arm chips have shipped. The company also cited the development community supporting its architecture, describing an “ecosystem of ecosystems” spanning devices at the edge through cloud infrastructure and AI platforms.
From standalone IP to Compute Subsystems
Arm described an ongoing evolution in how it delivers technology to customers. Traditionally, it has offered standalone IP blocks such as CPUs, GPUs, and system IP. In response to rising chip complexity and longer development cycles at advanced process nodes, the company said it introduced Compute Subsystems (CSS) about three to four years ago to help customers integrate pre-verified IP blocks faster.
Arm said the approach can reduce time to production by up to a year or 18 months in some cases. It also provided an update on financial mix, stating that CSS already represents almost 20% of its royalties and is growing, while noting royalties typically lag licensing by two to three years.
Why agentic AI is driving CPU demand
Arm executives argued that while accelerators generate tokens, CPUs remain critical in AI infrastructure for orchestration and workflow execution. The company described a shift from conventional AI inference to “agents” that run multi-step workflows continuously, which Arm said increases tokens per human by about 15x or more.
Arm estimated that traditional AI data center configurations require roughly 30 million CPU cores per gigawatt, but that agentic workloads could increase that to about 120 million CPU cores per gigawatt—a fourfold increase—while staying within tight power envelopes.
Arm announces Arm AGI CPU as its first chip sold for revenue
Arm CEO Rene Haas said the company is entering “a new business” by supplying CPUs as chips, and introduced the Arm AGI CPU as the company’s first silicon chip that it is selling to customers for revenue. Haas said the decision was driven primarily by partner demand and by the need for a power-efficient CPU solution for agentic AI data centers.
Meta’s Head of Infrastructure, Santhosh Janarthanan, described Meta’s scale—citing about 3.5 billion daily users across its apps—and the rapid expansion of AI clusters from early deployments of 128 GPUs in 2023 to “tens of thousands” of GPUs in a single cluster today. He referenced large projects including “Prometheus,” which he said would surpass one gigawatt by the end of the year, and “Hyperion,” which he said is projected to reach five gigawatts in a few years. Janarthanan said Meta partnered with Arm after surveying the market and finding that options did not meet both performance and power requirements.
OpenAI’s Kevin Weil also emphasized the CPU’s role, calling it an orchestrator for agentic workloads that run tools in containers and execute scripts. Weil said power is a key constraint and argued that more efficient CPUs free up power for inference and other compute needs.
Product details: cores, I/O, memory, process, and rack density
Arm Executive Vice President Mohammed Awad said the Arm AGI CPU was designed around three principles: performance, scale, and efficiency. He described the product as starting with the Neoverse V3 Compute Subsystem and said the chip includes:
- 136 cores based on Neoverse V3
- 2 MB dedicated L2 cache per core
- Support for up to 3.7 GHz
- 96 lanes of PCIe Gen 6 with CXL 3.0
- DDR5 memory with up to 6 GB/s per core sustained bandwidth
- Targeting under 100 nanoseconds of memory latency via a dual chiplet design
- 300-watt TDP
- Built on TSMC 3-nanometer process technology
Awad said the company demonstrated high core density in standard Open Compute Project (OCP) racks, including an air-cooled 36-kilowatt rack configuration and a liquid-cooled 200-kilowatt rack configuration, and described contributions Arm is making to OCP such as Arm ServerReady and other tools intended to benefit the broader ecosystem.
Awad also said the Arm AGI CPU is available now and in customers’ hands for evaluation, with production by the end of the year. He added that Arm and partners have been investing in data center software for more than 15 years and said more than 1.25 billion Arm Neoverse cores have shipped into data centers.
In closing remarks, Haas said Arm will continue its IP and CSS roadmaps while committing to future generations of the chip, naming Arm AGI CPU 2 and Arm AGI CPU 3. Haas also discussed market sizing, saying Arm’s AI data center business represents about a $3 billion total addressable market (TAM) today in royalties, and that the Arm AGI CPU initiative could expand Arm’s opportunity to about $100 billion TAM in the future. He added that Arm believes it has an opportunity to address more than $1 trillion TAM by the end of the decade, spanning edge to cloud deployments.
About ARM (NASDAQ:ARM)
Arm Limited (NASDAQ: ARM) is a global semiconductor IP company best known for designing energy-efficient processor architectures and related technologies that underpin a wide range of computing devices. Founded in 1990 as a joint venture between Acorn Computers, Apple and VLSI Technology and headquartered in Cambridge, England, Arm develops the ARM instruction set architectures and core processor designs that chipmakers license and integrate into custom system-on-chip (SoC) products. The company operates a licensing and royalty business model rather than manufacturing chips itself.
Arm’s product portfolio includes CPU core families (such as Cortex and Neoverse lines), GPU and multimedia IP (Mali), neural processing units (Ethos) and a suite of system and physical IP blocks.
