Arm's Quiet Push Into AI Silicon With Samsung Tests New Boundaries

Arm Holdings just took another step beyond its traditional role as a designer of blueprints. The British company is now supplying the architecture for a dedicated AI accelerator that Samsung will turn into a complete system-on-chip. Built on Samsung's 2nm process, the part targets on-device inference. It promises lower latency and less reliance on distant cloud servers. Details emerged in recent days. Arm provides the AI accelerator architecture and register-transfer level design. Samsung's System LSI business integrates the full SoC. Its foundry division manufactures the chip on the SF2 2nm node and handles advanced packaging. The project kicked off after Arm approved non-recurring engineering fees in late August 2026. Shares rose about 3% on the news. Yet analysts caution against reading too much into it. This isn't the data-center jackpot many investors hoped for. That prize sits elsewhere in Arm's portfolio. Two Distinct AI Plays Emerge The Samsung effort focuses on consumer devices. Phones, tablets, and other gadgets gain the ability to run complex models locally. Power efficiency matters here. So does keeping sensitive data on the device. Cloud calls introduce delay and privacy risks. This new accelerator aims to shrink both. Contrast that with Arm's bigger bet. The company now sells its own finished silicon for the first time in 35 years. The Arm AGI CPU, built with Meta as lead partner, packs up to 136 Neoverse V3 cores across two chiplets on TSMC's 3nm process. It targets agentic AI workloads. These involve orchestration, reasoning, and coordination across accelerators inside massive data centers. Performance claims stand out. Arm says the AGI CPU delivers more than 2x the performance per rack compared with current x86 platforms in validated configurations. One air-cooled rack holds 8,160 cores at 36kW. A liquid-cooled version scales to over 45,000 cores at 200kW. Potential capex savings reach $10 billion per gigawatt of capacity. Those numbers come from internal modeling and early partner feedback. Rene Haas, Arm's CEO, pointed to surging demand. 'Demand now exceeds $2 billion across fiscal 2027 and fiscal 2028,' he said on a recent earnings call. High-performance computing made up 66% of revenue in the latest quarter. Royalties from data-center designs more than doubled year over year. But the Samsung project operates in a different world. High-volume mobile chips bring thinner margins than custom server processors. Still, success here hands Samsung a critical reference design for its struggling foundry business. The SF2 node needs wins to compete against TSMC's N2. A real AI accelerator SoC provides exactly that validation. The partnership doesn't stand alone. In 2024 the two companies joined ADTechnology and Rebellions on an AI CPU chiplet platform. That effort uses Arm Neoverse CSS V3 subsystems, Rebellions' NPU with Arm CMN interconnect, and Samsung's 2nm GAA process. Early tests show 2-3x better efficiency on large language models such as Llama 3.1 405B. 'Datacenters are under pressure to keep up with the fast-paced evolution of AI workloads and require new power efficient solutions that can meet the increasing need for greater performance,' Eddie Ramirez, Arm's vice president of go-to-market for infrastructure, said in a Samsung Semiconductor announcement. Jinwook Oh, CTO of Rebellions, added context around scalability. His company's REBEL accelerator brings high-bandwidth memory. Combined with Arm subsystems and Samsung's process, the platform looks toward future AI infrastructure needs. Recent coverage reinforces the split focus. A September 7 analysis from 24/7 Wall St. notes that investors paying 298 times earnings for Arm may overstate this deal's immediate impact on data-center growth. The real story remains the AGI CPU and its pipeline. Yet on-device AI gains importance. Models keep growing. Running even distilled versions in the cloud costs money and adds latency. Local inference solves both problems for many consumer applications. Samsung's Exynos chips already incorporate Arm's latest Scalable Matrix Extension 2 technology. That extension boosts matrix math inside the CPU itself, reducing trips to discrete accelerators. Tests on object detection show up to 70% better performance versus prior generations without SME2. Such gains matter when thermal limits and battery life constrain mobile devices. They also open the door to more flexible software that assigns workloads across CPU, GPU, and NPU depending on the task. Arm's broader platform strategy supports these moves. The company now offers intellectual property, pre-integrated Compute Subsystems, and actual silicon. That range gives customers choices they lacked before. Hyperscalers build their own designs on Neoverse. Others buy the AGI CPU directly. Still others license the new accelerator IP for custom parts. Partnerships multiply. Meta co-developed the AGI CPU and plans to deploy it alongside its own MTIA accelerators. OpenAI appears on launch partner lists and may serve as an end customer for the Samsung SoC project. Cloudflare, Cerebras, SK Telecom, and SAP also committed early. Manufacturing questions linger. The AGI CPU uses TSMC's 3nm. The new accelerator uses Samsung's 2nm. That split lets Arm test both leading foundries while giving Samsung a high-profile design win. Success on SF2 could attract more customers to the Korean foundry, which trails TSMC in advanced nodes. Challenges remain. Arm's legal dispute with Qualcomm over licensing terms heads to trial in late 2026. Any shift in economics could affect server chip projects. Competition from NVIDIA, AMD, and custom ASICs stays fierce. Yet Arm's energy-efficiency heritage gives it an edge in power-constrained environments, whether phones or dense data-center racks. So the Samsung collaboration marks progress on two fronts. It extends Arm's reach into dedicated AI accelerators for edge devices. And it strengthens ties with a vital manufacturing partner at a moment when advanced process technology decides winners. The data-center goldmine may still lie with the AGI CPU. But this smaller, mobile-first effort could prove just as strategic over time. Industry watchers will track tape-out schedules and early silicon results. If the 2nm accelerator delivers on latency and efficiency promises, expect more custom AI projects to follow. Arm no longer simply draws the maps. It's helping build the vehicles that travel those roads. And Samsung gains a chance to show its foundry can compete at the bleeding edge.