COMPUTEX 2026 Arm CEO Keynote: The Foundation for the Agentic AI Era

Author: Renee Haas, CEO, Arm Collection: Computex 2026

Summary

Arm CEO Renee Haas delivered a 49-minute keynote at Computex 2026 in Taipei, framing the rise of agentic AI as the defining inflection point for compute architecture. The central thesis — first articulated at Arm's "Arm Everywhere" event on March 24, 2026 — is that agentic workloads require 4x more CPU cores in the same power envelope compared to pre-agentic AI deployments, because agents never sleep, spawn sub-agents, and demand continuous orchestration that GPUs alone cannot handle. Arm claims this multiplier is now being validated industry-wide, with CPU total addressable market (TAM) projections that once looked aggressive at $120 billion in five years now being superseded by analyst estimates of up to twice that figure or more.

The keynote's primary product announcement was the Arm AGI (Agentic General Intelligence) CPU — branded "The Agentic CPU" — which is now in production, manufactured by TSMC in Taiwan. The chip ships in two rack configurations: an air-cooled rack with 8,000 cores at 36 kilowatts, and a liquid-cooled rack with over 45,000 cores at 200 kilowatts. Both deliver 2x performance per rack versus a comparable x86 system in the same power envelope, yielding an estimated $10 billion in capital expenditure (CAPEX) savings for hyperscale data center operators at 10-gigawatt scale. Early AGI CPU customers announced include Meta, Rebellions, SAP, Cerebras, OpenAI, and SK Telecom (from March 24), with Oracle (via Oracle Cloud Infrastructure) and ByteDance added at Computex as the newest adopters. A multi-generational roadmap — AGI CPU 2 and AGI CPU 3 — was disclosed, paired with Arm Compute Subsystems (CSS) V3, V4, and V5 lineups.

The keynote's second major arc covered the PC market, where Arm asserted it now uniquely addresses two bifurcated use cases: efficient on-the-go computing (long battery life, always-connected) and extreme agentic performance platforms. The highlight was the Nvidia RTX Spark, co-developed with Arm and MediaTek using Arm CSS, featuring a 20-core Arm-based Grace CPU, a Blackwell RTX GPU with 1 petaflop FP4 AI performance, 128 GB unified memory at 600 GB/s via NVLink Chip-to-Chip (C2C), running full Windows natively on Arm. Jensen Huang joined Haas on stage to discuss how agentic systems will make PCs autonomous workers running 24/7, with the RTX Spark capable of running a 100-billion-parameter model fully compressed into MVFP4 format. Haas closed by summarizing Arm's position as a unified "one agentic platform, cloud to edge," supported by 22 million developers — the largest compute developer community globally.