2026年3月25日，Arm 宣布将推出第一颗自有矽晶片 Arm AGI CPU，结束了几十年主要靠授权 IP 的模式，过去 Apple、Tesla、Nvidia、Microsoft、Amazon、Samsung、Qualcomm 等公司都基于 Arm 设计在做晶片。这家公司曾有估计每位地球人可对应约三颗 Arm 晶片，从1970年代起源于 Acorn，已走到 SoftBank 于2016年收购并私有化后再次转向；2020年 Nvidia 收购案被叫停后，智慧型手机增长放缓，Haas 于2022年接任 CEO，在保留 SoftBank 90%持股的情况下重新上市。Haas 在访谈中将这一步说成文化与产品转型：重视冒险、快速迭代，以及以 compute platform 思维运作。
Haas 说 Arm 从「仅授权 IP」转成「运算平台公司」，认为 CPU 与作业系统与软体生态（Windows、macOS、iOS、Android、Linux）高度互依；推出自制伺服器 CPU 被他比作 Microsoft Surface 或 Google Pixel，作为可带动整体生态的旗舰参考件。Arm AGI CPU 主打资料中心，核心主张是极高电源效率，并针对 agentic AI 工作负载最擅长，因为即使 GPU 重要，CPU 仍是执行代理式 AI 流程不可替代的核心。Meta 是首发客户之一，早期名单另含 SK Hynix、Cisco、SAP、Cloudflare；Haas 预期这将放大 Arm 软体生态，带来更多 Arm 最佳化与工具链。
资金与供应链压力仍是焦点，但 Haas 认为这次推出不一定直接得罪伙伴，认为 Intel 和 AMD 受压可能比 Nvidia 更大，因为 Arm 直冲 x86 生态份额。制造端完全由 TSMC 承担，参考伺服器会和 Super Micro、Foxconn 等伙伴一起交付。公司新增约2,000名工程师负责后端设计、子系统与商用化，Haas 表示大约85% 的子系统已可直接提供客户，AGI CPU 也沿用已由合作伙伴发布的既有矽平台。传统上新晶片常需2到3代才形成预期影响，但他仍判断首发就有高度机率成为量产候选，而且明确称这不是“bet-the-farm”赌注。
On March 25, 2026, Arm confirmed it would launch its first own silicon, the Arm AGI CPU, ending decades of mainly licensing IP, though companies such as Apple, Tesla, Nvidia, Microsoft, Amazon, Samsung, and Qualcomm had long relied on Arm-based designs. Arm had been estimated at about three chips per person on Earth; after SoftBank bought Arm in 2016 and took it private, growth in smartphones later slowed, Nvidia’s 2020 acquisition attempt failed, and Haas became CEO in 2022 while SoftBank kept 90% ownership in a public Arm. In the interview, Haas framed the move as a cultural and product shift toward risk-taking, fast iteration, and a compute-platform operating mindset.
Haas said Arm has shifted from “IP only” to a compute platform model, arguing that CPUs and software ecosystems (Windows, macOS, iOS, Android, Linux) are tightly coupled. He compared a proprietary server CPU to Microsoft Surface or Google Pixel: a flagship reference product that can lift the entire ecosystem. The Arm AGI CPU is aimed at data centers, with a central claim of high power efficiency and strong suitability for agentic AI workloads, because CPUs remain essential even while GPUs stay important. Meta is the first customer, with early mentions also including SK Hynix, Cisco, SAP, and Cloudflare; Haas expects more Arm software optimizations and toolchains to spread.
Execution and economics remain central risks, but Haas argued the launch is not mainly a direct strike against partners, suggesting Intel and AMD may feel pressure sooner than Nvidia because Arm challenges x86-share dynamics. Manufacturing is done fully by TSMC, with server reference systems built with partners like Super Micro and Foxconn. Arm added about 2,000 engineers for backend design, subsystems, and commercialization, and Haas said roughly 85% of subsystem work can already be handed to customers while the AGI CPU leverages existing silicon already shipped by partners. Although chip ramp usually takes two to three generations to achieve full impact, he said first-volume viability is highly likely and he did not treat this as a “bet-the-farm” bet.