On March 24 (U.S. time) in San Francisco, Arm Holdings plc, the global chip IP leader, did something it had never done in the past 35 years—officially launched and began selling its first self-developed mass-production chip: the Arm AGI CPU.

Arm, known for its strict rule of "selling blueprints, not building chips" and dubbed the "Switzerland of the chip world," has torn off its neutral label, stepping from behind the scenes onto center stage, officially entering the trillion-dollar AI data center battlefield.

The "Rebellion" Has Been Brewing for a Long Time

For years, Arm has licensed chip architecture designs to hundreds of clients, including Apple, Qualcomm, and NVIDIA, without ever producing a single physical chip itself. Yet, it has supported the shipment of over 350 billion chips globally.

However, the impact of the AI wave, especially the explosion of "Agentic AI," is rewriting the rules. Arm clearly doesn't want to miss this opportunity.

"Agents are changing everything," Arm CEO Rene Haas stated bluntly at the launch event. As AI applications shift from model training to large-scale inference and agent interaction, expensive GPUs/accelerators in data centers handle content generation (Tokens), while the "dirty work" of massive task scheduling, data movement, and logical control falls entirely on the CPU's shoulders.

Arm calculates that under the same power budget, future data center demand for CPU cores will surge from around 30 million to 120 million, a fourfold increase. Traditional x86 architectures, burdened by historical baggage, have hit bottlenecks in energy efficiency. Clients need a CPU built for AI-native, ultra-efficient performance.

Thus, Arm's "rebellion" became inevitable. From pure IP licensing to providing compute subsystems (CSS) to now directly selling finished chips, Arm has completed a "triple jump" in its computing platform strategy.

The 136-Core Beast

The Arm AGI CPU is designed from the ground up to be the "super scheduler" and "data mover" for data centers in the Agentic AI era.

Let's examine its technical specifications—they're simply luxurious.

? In terms of process and architecture, it adopts TSMC's top-tier 3nm process, based on a dual-Chiplet design, integrating up to 136 Arm Neoverse V3 high-performance cores on a single chip.

Its core frequency reaches 3.7GHz, with each core enjoying 2MB of dedicated L2 cache and a thermal design power (TDP) controlled at 300 watts.

Additionally, it supports DDR5-8800 high-speed memory, providing each core with up to 6GB/s of dedicated memory bandwidth, reducing memory access latency to below 100 nanoseconds. Meanwhile, it offers 96 PCIe Gen 6 lanes and supports the CXL 3.0 protocol, ensuring data throughput capabilities.

Notably, Arm has boldly abandoned traditional simultaneous multithreading (SMT) technology. In Agentic AI scenarios, each thread needs to give its all, and the worst scenario is core resources being divided and contested. Therefore, they believe that equipping each core with sufficient and exclusive bandwidth is the key to delivering sustained, stable "full-throttle" performance.

Arm officially claims that, under standard reference server designs, the Arm AGI CPU can deliver over twice the per-rack performance of the latest x86 systems.

With its high-density design, it can deploy 8,160 cores in a standard air-cooled rack; if using a liquid cooling solution in collaboration with Supermicro, the number of cores per rack can soar to over 45,000.

According to Arm's calculations, for clients operating large-scale AI data centers, this means potential capital expenditure (CAPEX) savings of up to $10 billion per gigawatt (GW) of computing power.

A Luxurious "Friend Circle"

The success of any chip hinges on ecosystem support. The debut of the Arm AGI CPU is backed by a dream team of clients and partners, perhaps its greatest source of confidence in this "rebellion."

Meta is the inaugural customer and a deep collaborative developer. Santosh Janardhan, head of Meta's infrastructure, revealed that about two and a half years ago, they couldn't find a CPU on the market that met their stringent performance and power requirements, ultimately partnering with Arm to define the product from scratch.

OpenAI has also endorsed it. Sachin Katti, head of industrial computing at OpenAI, pointed out that more efficient CPUs mean more power can be reserved for AI inference, fostering stronger intelligence. Additionally, companies like Cerebras, Cloudflare, SAP, and SK Telecom have plans to deploy it.

Even more remarkable is the collective endorsement from the entire industry chain. In a launch event video, executives from over 50 industry giants, including NVIDIA, Google, Microsoft, Amazon Web Services, Broadcom, Marvell, TSMC, Samsung, and Micron, appeared to express their support. NVIDIA CEO Jensen Huang stated in the video, "Accelerated computing makes the CPU an indispensable partner."

Arm has also placed high commercial expectations on this. Rene Haas revealed that Arm anticipates its annual total revenue will reach $25 billion in five years, roughly five times its current level. Among this, the new chip business is expected to contribute approximately $15 billion in annual revenue opportunities by 2031. In comparison, its traditional IP licensing business is projected to grow to around $10 billion.

Behind this high-stakes gamble lies substantial rewards. Arm's CFO once explained with a simple example: In a $1,000 chip, if Arm only licenses the instruction set, it might earn about $50; if it licenses the core design, it could earn around $100; but if it designs and sells the chip itself, the gross profit could reach as high as $500.

Arm is attempting to trade lower gross margins for a much larger revenue pie.

Of course, Arm will directly break into (Note: This Chinese term means "barge into" or "enter") the data center CPU market long dominated by Intel and AMD, facing them head-on. At the same time, its relationships with cloud giant clients (such as AWS and Google) that also develop their own Arm-based chips will become delicate and complex. Arm must carefully balance its dual identity as both a "supplier" and a "competitor."

What do you think of Arm's "rebellion" and its new CPU? Discuss in the comments section.