[News] Arm Unveils First AGI CPU on TSMC 3nm; Meta Among Customers, Promising Up to 2× Performance vs. x86

Chip IP provider Arm has ventured into chip design. According to Commercial Times, Arm announced it will launch its first in-house processor for AI data centers, the “Arm AGI CPU,” targeting emerging agentic AI applications. The move marks a strategic shift from an IP supplier to a full platform provider. As the report notes, TSMC will manufacture the Arm AGI CPU using a 3nm process.

Arm also highlighted the chip’s performance and cost advantages. As noted in its press release, the Arm AGI CPU delivers more than 2× performance per rack compared with x86 CPUs, while enabling up to $10 billion in CAPEX savings per GW of AI data center capacity.

On the customer front, Meta is among the first wave of partners and is participating in co-development, integrating the Arm AGI CPU into its data center architecture alongside its in-house MTIA AI accelerators. Arm also confirmed additional commercial traction with partners including Cerebras, Cloudflare, F5, OpenAI, Positron, Rebellions, SAP, and SK Telecom, its press release notes.

Within the hardware ecosystem, Arm is partnering with OEMs and ODMs such as ASRock Rack, Lenovo, Quanta Computer, and Supermicro. Early systems are already available, with broader rollout expected in the second half of the year. As Bloomberg notes, Arm expects the chip to generate about $15 billion in annual revenue within five years.

Performance and Architecture Highlights

Looking ahead, Arm estimates that as AI applications scale, CPU compute demand per GW of data center capacity will rise more than fourfold, driving strong demand for next-generation AI-focused CPUs. Compared with traditional x86 architectures, the new chip balances high throughput, energy efficiency, and architectural simplicity, as its press release indicates.

Regarding performance, Arm says the chip delivers up to 136 Arm Neoverse V3 cores per CPU, offering strong performance across core, SoC, blade, and rack levels. It also provides up to 6GB/s of memory bandwidth per core with latency below 100ns. In terms of system design, the chip has a 300W TDP and assigns a dedicated core per program thread. It supports high-density 1U server chassis for air-cooled deployments of up to 8,160 cores per rack, as well as liquid-cooled systems delivering more than 45,000 cores per rack.

From a broader market perspective, as CNBC notes, Arm’s move underscores a renewed surge in CPU demand. NVIDIA, which has emerged as the leader in AI GPUs, recently told CNBC that CPUs are “becoming the bottleneck” as agentic AI reshapes compute requirements. Meanwhile, Bloomberg notes that Arm’s push into chip design could complicate its relationships with customers. Many of the largest buyers of data center silicon, including Meta, run in-house chip programs, and nearly all of them license technology and designs from Arm.

Read more

• [News] NVIDIA Reportedly Plans Arm-Based SoCs, with Lenovo and Dell Among Early Adopters
• [News] Arm Reportedly Weighs Chiplet and Solution Development, Raising Customer Tensions

(Photo credit: Arm on X)