March 26, 2026
By Bob O'Donnell
It’s rare to see a decades-old company take a completely different path in the process of creating a new product. It’s even more surprising when, at a quick glance, that new product appears to be a competitor to the company’s primary customers.
With the launch of its new AGI CPU chip, however, Arm has done all of that. Rather than simply providing the architectural design IP for a new processor—as it has traditionally done—the company is taking its own design into production and will be selling the chip to companies such as Meta, OpenAI, SK Telecom, and more. While that does represent a bit of a competitive challenge to traditional Arm licensee customers such as Broadcom, Nvidia and others, Arm was quick to point out—and rightfully so—that the market opportunity for AI-focused silicon is so large and so diverse that there’s plenty of room for all.
This first Arm chip, in fact, was developed along with Meta—yet another Arm chip design licensee—and is intended to work alongside Meta’s own custom AI accelerators (the recently announced MTIA chips). Meta recognized that it needed a large number of CPU cores to help efficiently run agentic AI workflows and found the new Arm AGI CPU to be a perfect fit for its requirements. Similarly, other customers are also finding that the unique capabilities and low-power draw per core of these Arm CPUs fit the unique demands of their environments—particularly in comparison to x86-based options.
In fact, one of the reasons Arm chose to take this strategic right turn and follow the path of product creation is because the company recognized that many of these kinds of situations exist. But if you look at the development from a bigger picture perspective, it’s also the next logical step in the company’s multi-year process of driving higher value creation and helping build more complete solutions. Several years back, Arm extended its individual CPU and GPU core designs into more elaborate SOC (System on Chip) architectures that allowed its licensee customers to build these more sophisticated chips. Then the company created what it called CSS (Compute Subsystems), which took things even further by starting to specify other critical sub-components that made the process of designing full products much easier. Some of the elements of CSS also ensured that companies who were using Arm’s designs could get the full performance capabilities out of Arm’s chip architectures.
With the debut of the AGI CPU, Arm is taking that solutions-focused process even further. Specifically, it has created a chip to be built on TSMC’s 3nm process node that leverages the complete capabilities of its latest Neoverse V3 CPU architecture. One of the key differentiators of the AGI design is the amount of single-threaded CPU cores that it includes. AGI supports up to 272 cores in a single air-cooled 1U two CPU blade server (136 per CPU). In a standard air-cooled rack design based on 36kW of power with 30 blades that translates to 8,160 CPU cores. A liquid-cooled option is expected to offer an impressive 45,000 cores in a single rack.
In addition to large numbers of CPU cores, Arm’s AGI dual chiplet design supports fast access and high throughput to large amounts of memory. The system features memory and IO on the same chiplet, which enables 110ns memory latency and high throughput (up to 6 GB/sec), uses DDR5-8800 DRAM, and supports access to up to 6 TB of memory per chip. Those capabilities make it very well-suited for the types of massively parallel agentic AI workloads that have suddenly made CPUs nearly equally important to GPUs for large AI factories. That, in turn, makes the timing of the AGI CPU’s release—it’s expected to be available in the second half of this year—nearly perfect. The rise of tools like OpenClaw and other autonomous AI agents has put a huge focus on the need for CPUs to handle certain critical parts of the workloads, such as orchestrating accelerators, managing memory, scheduling, and other tasks that these tools are generating.
It’s important to note that Arm will continue to develop and license chip designs—including the new AGI CPU—for its existing customers. The company’s productization efforts are an add-on to its existing business model, not a replacement of it. The company also hinted that this is the first of what will likely be several Arm-branded chips, so there are more to come—and potentially across other markets as well.
From a strategic perspective, the announcement is incredibly important for Arm on many levels. First, of course, is the ability to leverage its multi-decade semiconductor design expertise and turn out world-class finished products that can meet a variety of different applications. Directly related is the company’s ability to literally cash in on that expertise in a more meaningful way. The amount of impact that Arm chip designs have had on the entire computing world is enormous and, quite honestly, it hasn’t felt like the company has gotten the monetary recognition for all the work it’s done in a manner that it’s deserved. With the start of chip sales, Arm should be able to enjoy the significantly higher revenues and profit margins that finished devices often command. In fact, the company is predicting that it can increase its revenues by an impressive 600% by 2031 because of these moves.
Even beyond these ramifications, however, the Arm AGI CPU release also cements the reality of an extremely diverse semiconductor ecosystem as a direct result of the explosion in AI computing. Even just a few years back, the potential blowback that Arm might have received for releasing its own chip would have been huge. Now, with the amount of industry-wide support that Arm announced at its Arm Everywhere launch event, they made it clear that the TAM for AI accelerated computing is so large that there’s plenty of opportunity for all. Plus, the amount of customization and specialization in chip design necessary to meet the specific needs of different organizations and workloads moving forward will undoubtedly drive even more new ideas and product launches in the future.
The bottom line is there’s little doubt in my mind that Arm will be able to expand both the range of companies that license their designs and buy their chips. At the same time, there’s also plenty of opportunities for other types of chip architectures as well. It’s not just about GPUs or CPUs anymore when it comes to AI. Instead, developments like Nvidia’s Groq-based LPUs, Qualcomm’s ongoing AI 200 series inference accelerator efforts, the custom chip designs from major cloud and AI providers, and many others will all be highly viable options in an increasingly heterogeneous semiconductor world.
Here’s a link to the original column: https://www.linkedin.com/pulse/arm-targets-new-path-agi-cpu-silicon-bob-o-donnell-q09fc/
Bob O’Donnell is the president and chief analyst of TECHnalysis Research, LLC a market research firm that provides strategic consulting and market research services to the technology industry and professional financial community. You can follow him on LinkedIn at Bob O’Donnell or on Twitter @bobodtech.
| 
