Arm just broke its 35-year history: from intellectual property licensing to factory-free chip making

There's a joke in a recent conversation between Ben Thompson (Stratechyry) and Rene Haas, CEO of Arm, that I think pretty much sums up what's going on: Thompson basically said that Haas was the only CEO he interviewed whose company did nothing more than license intellectual property (IP), and now Haas is "just another factory-less chip guy, like NVIDIA's Jensen Huang or Qualcomm's Cristiano Amon." Haas laughed, then admitted that it was true. That's what makes this story worth sitting down and analyzing carefully. Arm, a company that for 35 years lived only by licensing its microprocessor architecture (ISA - Instruction Set Architecture) to others for production, now designs and sells its own complete chip. For a company whose business model is built on the principle of "not competing with customers", this is not a small step.

From the venture to make CPUs for PDAs to the heart of every mobile device

Arm was born from a joint venture between Acorn Computer, Apple and VLSI Technology, with the initial goal of designing a low-power CPU to run on early personal assistant devices (PDAs). Because it is a general-purpose CPU architecture, not optimized for any single task, Arm is forced to grope to find its strengths over time, like many other successful startups still follow that path: just do it, find what you are good at, then stick to customers who need that exact thing.

Arm started out by making CPUs for PDAs. Image source: XDA

From IP licensing to CSS, to making your own chips for Meta

The intermediate step before Arm decides to produce its own chips is the CSS: Compute Subsystem model, which means that instead of just selling discrete design blocks, Arm packages the entire computing subsystem into a more complete package for customers to integrate faster into their system-on-chip (SoC).

The partnership with Meta has opened a new direction for ARM The story with Meta (Facebook) starts with CSS. According to Haas, around mid-2025, Meta approached Arm to discuss implementing CSS for their AI data center, but the conversation gradually turned to another question: instead of Meta licensing and designing it itself, could Arm directly make the chip itself? The reason, according to Haas, is simply the return on investment (ROI) problem: Meta already has many teams working on other internal chip projects such as MTIA: Meta's self-developed AI acceleration chip, so the question is whether it is worth making another chip yourself or letting Arm do it instead. The result is the first chip under the new model, called the Arm AGI CPU, and it works from the start, ready for production. Arm announced this milestone at an event called "Arm Everywhere", the first independent product launch in the company's history, instead of traveling to events like Computex as usual. Haas described it as "a landmark day" for Arm.

Why CPUs are still important in an age where everyone only talks about GPUs

Part of Arm's reason for taking this gamble lies in the belief that CPUs still play a key role in AI infrastructure, even though most of the industry's attention is on AI accelerators (GPUs) like NVIDIA's. Haas uses a rather expensive image: the volume of tokens being generated is now so large that it has to be calculated "by the pound, by the truckload", and he believes that in the near future, each processing core in the CPU may run its own AI agent, launch a virtualization process, finish processing and then stop, meaning that the number of cores and performance per core will become more and more important.

Risk: Arm doesn't have a factory, and conflicts of interest are inevitable

The price to pay for this move is real risk. Arm does not own a semiconductor factory, so it must work through a specialized chip design partner (ASIC) to reserve processing output. When asked if he was concerned if this chip business grew beyond production capacity, Haas admitted that it was a real risk, especially in the context of high bandwidth memory (HBM - High Bandwidth Memory) being a common bottleneck for the entire industry when increased CPU demand leads to increased memory demand.

But the bigger risk, in my opinion, lies in Arm's new position in the value chain. When asked directly if the only licensing customer capable of truly competing with Arm in terms of chip design, Qualcomm, was upset, Haas half-jokingly asked to "off the record", then affirmed that Arm only entered this segment because it was Meta who proactively proposed, not because Arm arbitrarily sought customers to compete with its own licensing ecosystem. That's a reasonable explanation, but it doesn't erase the fact that the line between "architecture-neutral licensor" and "directly competing chip seller" has now blurred significantly, and each of Arm's remaining licensing customers, from Qualcomm to Samsung to MediaTek, is surely having to rethink what its relationship with Arm in the next 5-10 years will look like.