Arm launches its own AI chips, with Meta, OpenAI and Cloudflare as first customers

Arm has done something few expected from it in years: the company that spent three decades selling processor architecture licenses has entered the market for finished AI chips. Among the first buyers of the new hardware are Meta, OpenAI, Cerebras and Cloudflare.

Arm was founded in 1990 in Cambridge and built one of the most successful tech businesses on an unusual model: it does not make hardware itself, but sells the right to use its architecture to anyone who wants it. Today Arm-based processors power iPhone and Android smartphones, Apple's M-series chips, a significant share of cloud providers' server infrastructure and billions of embedded devices worldwide. According to the company, more than 30 billion chips with its architecture shipped in 2023 alone.

Moving into manufacturing its own chips is a fundamental change of business model. On one hand, Arm gains access to higher margins: the maker of a finished chip earns considerably more than the seller of a license. On the other, the company starts competing directly with those who use its architecture themselves: Qualcomm, MediaTek, Samsung and Apple. Arm sells IP to all of them at the same time. It is a delicate balance that will demand skilled management.

The choice of first customers is telling. Meta is actively building its own AI infrastructure and reducing its dependence on NVIDIA. OpenAI is looking for alternatives to the expensive H100 for training and inference of its models. Cerebras specializes in AI chips and is one of the fastest-growing startups in the sector. Cloudflare is rolling out AI capabilities on edge servers around the world. All four need high-performance, energy-efficient chips for AI workloads today.

Details of the new chip have not been fully disclosed yet. What is known is that it is an AI CPU: a central processor with blocks optimized for neural network computation. That sets the product apart from NVIDIA's GPU-based solutions and from Google's specialized TPUs. The CPU approach may find demand where inference at the network edge is needed, where AI and conventional workloads are mixed, or where a predictable power profile matters. Those requirements are especially important for edge infrastructure operators such as Cloudflare.

The move fits a broader trend: the largest tech companies want to control the entire value chain, from architecture to finished hardware. Apple has done this for many years with its M and A series chips. Google builds TPUs for its own needs. Amazon develops Trainium and Inferentia. Microsoft is experimenting with its own AI accelerators. Now Arm joins the list.

For the market, this means growing fragmentation of the AI chip ecosystem. NVIDIA still dominates with a share above 80% of the GPU market for AI workloads, but the number of alternative players keeps growing, and they come from unexpected places. Arm, with its architectural authority, industry ties and experience designing energy-efficient processors, is a potentially serious competitor.

The key question is whether the company can keep the trust of its longtime licensees while becoming their rival in the market for finished chips.

*Meta is designated an extremist organization and banned in Russia.