ARM Sells Its Own Silicon

Aug 19, 2026 · 5 min read

In AI hardware, there’s not a very wide field of competitors. Some describe it as almost a zero-sum game. For instance, xAI and the folks behind Colossus have said that they now only want to use Nvidia’s GPUs. Based on a dominant role in selling such AI hardware, Nvidia’s stock has blossomed to eclipse even those of Apple and Microsoft by market cap. That’s just one of many indicators showing that, in the new world of AI microprocessors, Nvidia is king. And there aren’t a lot of runners-up.

That’s part of what makes the news of ARM Holdings so interesting: in March, news broke that the firm is going to be selling its own chips.

“The company, a British unit of Japan’s SoftBank, on Tuesday announced plans for the first silicon product that Arm will design and sell since its founding in 1990,” wrote Don Clark for the New York Times. “It is a microprocessor aimed at data centers running artificial intelligence tasks.”

Other news includes the announcement that Meta will be an early customer.

What does it mean, that ARM is going to be competing in this space? Well, if you look behind the curtain, the sort of incestuous org chart here gets pretty weird.

The History of ARM

The reality is that ARM had early designs, for Nokia phones and other devices like PDAs, that were based on different design philosophy, something called RISC or Reduced Instruction Set Computing.

The idea is that if the chip doesn’t have to do everything, if it operates only on the tasks that the device needs done, your chip can be smaller and more efficient, energy-wise.

That’s not what Nvidia is doing: the front-runner is making AI-native tech, specifically geared toward LLM production. But question marks remain about how ARM’s products will work in the current market.

MORE FOR YOU

Nvidia Uses ARM

So digging around, I found that not only does Nvidia use ARM CPUs in architectures like Grace and Rubin, but it also tried to acquire the firm in the last decade. Regulators, it’s reported, squashed the deal in 2022, because it violated an anti-trust principle: that the biggest seller of AI chips shouldn’t also have ownership of a relevant designer like that.

That didn’t stop Nvidia from acquiring Run:ai, and Mellanox. But that’s a different thread altogether.

How it Works

I’ll just give you this explanation that I got from GPT, and in the interest of having a human in the loop, I’ll try to translate:

“Remember why NVIDIA couldn't buy Arm?

Because regulators worried that NVIDIA would control a technology that its competitors needed.

Now reverse the situation.

Arm itself is becoming a competitor to companies that license Arm.

That's potentially uncomfortable because Arm has privileged knowledge of the ecosystem and, importantly, controls the underlying architecture/IP.”

Even the model knows that this whole thing is strange and unusual.

It’s just too close for comfort. As for GPT’s claim that ARM and Nvidia can cooperate, offering, for example, a CPU that sits next to a custom GPU, I’m having a hard time getting my head around that one, too. But the above suggests that the two companies are capable of cooperating.

Thoughts from the Top

Here’s more on the rationale, from ARM’s perspective, specifically from Rene Hass, CEO, in a recent interview around that news of the ARM chip going on the market.

Contending that “we entered this (market) because Meta asked us to,” Hass had this to say:

“I had a question at the investor conference about competing with Nvidia,” he said, broaching the topic. “And I said, you know, a month ago, no one would have asked about any Arm person competing with anybody. So it’s wonderful to have these kinds of conversations; the market is underserved and there aren’t choices. There isn’t a product from Qualcomm, there isn’t a product from MediaTek, there isn’t a product from Infineon, there just isn’t.”

To say that there’s a lot more in this interview is an understatement. If you want to know more, about the specifics of the infrastructure, about the market context, and about all manner of design wonkery, read it in detail.

Up to Date

It’s probably predictable that you can’t find a whole lot of updates on the ARM news past March. I did find analysis July 29 from Anju Kushwaha at Vucense, who characterized the situation this way:

“Infrastructure leads should evaluate the “Sovereignty Score” of their cloud providers based on their dependency on proprietary silicon versus open-standard hardware. … this shift signals the end of the ‘licensing-only’ era for ARM, as it competes directly with Nvidia and custom ASIC providers for dominance in the 2026 AI compute market.”

And here’s some of that common language again around new market entry:

“This move marks a critical juncture in chip-stack sovereignty: by controlling both the design and the physical silicon, ARM is attempting to challenge Nvidia’s dominance while offering cloud providers a more integrated alternative.” (italics mine)

Now, I liked this part of Kushwaha’s analysis that focuses on the changing needs of hardware users:

“Nvidia’s dominance is built on massive Tensor Cores that are incredibly efficient at dense matrix multiplication—the brute-force math required to train massive Foundation Models. However, in 2026, the bottleneck has shifted from training to inference. Autonomous AI agents do not need to constantly crunch dense matrices; they need to perform highly branched, logical ‘if-this-then-that’ reasoning. ARM’s SVE architecture allows the AGI CPU to dynamically adjust vector lengths on the fly, optimizing for the sparse, unpredictable activation patterns of next-generation transformer-based agents.”

So, matrix multiplication recedes in importance?

I guess we’re going to have to wait and see how all of this shakes out. It certainly might be something that engineers or top brass ask each other about at conferences. Stay tuned.