Pulse News

**Why Arm’s CEO Sees Room for Nvidia Competitors**

By Anissa Gardizy

Feb 26, 2025, 6:00am PST

Cloud and chip executives have predicted for two years that Nvidia's dominance in artificial intelligence server chips would crater, but that hasn't come to pass. Yet Rene Haas, CEO of Arm Holdings, whose chip designs power part of Nvidia's systems as well as those of major competitors, feels comfortable saying it's only a matter of time before sales of other AI chips grow meaningfully as well.

**The Takeaway**

- Arm CEO says specialized 'inference' chips to run AI can take off

- He says 'we're nowhere close' to artificial general intelligence

- He sees potential for consumer AI devices that attach to clothing

"You will start to see some specialized chips for running AI because so many of the biggest AI developers and cloud providers are developing alternatives to Nvidia chips," Haas said. Haas isn't sure which Nvidia challengers will have the most success. Chip startups, cloud providers such as Amazon, new entrants such as OpenAI, and traditional chip firms all have a shot, he said. That's because companies like OpenAI will eventually spend more money running existing AI, known as inference, than training new models, an arena where Nvidia's chips have been especially useful. The demand for chips for comparatively less-intensive tasks will provide an opening to the Nvidia challengers, he said.

Few people are closer to AI chips and data centers than Haas. Arm licenses designs for central processing units that are pervasive in smartphones and data centers, including CPUs that sit next to Nvidia's Blackwell graphics processing units. Arm's shares have more than doubled since the company was publicly listed in late 2023, though licensing designs to Nvidia makes up a small part of Arm's business. (Nvidia reports quarterly earnings later Tuesday.) Although Arm benefits from Nvidia's sales, Haas may not be neutral about Nvidia's prospects; Arm is reportedly creating its own AI chips, something he declined to comment about.

Haas also sits on the board of SoftBank, Arm's majority owner, and is involved in Stargate, a joint venture with OpenAI to develop AI data centers in the U.S. that promises to spend a jaw-dropping $500 billion over the next four years. In an interview, Haas also discussed the future of AI devices and how Elon Musk's XAI has impacted the industry. This interview has been edited for length and order.

Last week The Information reported that OpenAI projected its spending on inference compute (running AI models) will surpass its spending on training models in 2030. Is that in line with your predictions?

"If you ascribe to a notion that at some point in time... [gains from] training [new models using large server clusters] asymptotes to some extent, where you're doing just primarily reinforcement training [improving the model after the initial training run] and not developing these giant new frontier models. Whether that's five years away...I guess 2030 is only five years away. Good grief. Five years away, 10 years away. That's an interesting judgment call. Let's just say for a moment that they're right. There will be a point where that will happen, and inference will be the dominant workload by quite a large, large proportion."

What you're seeing now is that we're nowhere close to [artificial general intelligence, AI that's on par with humans in handling complex tasks]. I was reading this morning about these software engineer agents and how good they are, and they can pass certain math tests, but are they really able to write spectacular code to the level of a midlevel engineer? I think when you start getting to that level, and everyone has different definitions for AGI, you'll start to see that we're on to something. We're clearly not there yet, which is why people are spending so much money on training. So whether that number is 2030, let's just say for a moment that it is. I could see the shape of the curve looking that way.

**In the long term, do you think companies will run AI inference on Nvidia GPUs?**

"It's still a little early to tell where that all goes right now, because so much of the installed base is Nvidia. So it's almost, you know, a default that the products are running on Nvidia hardware that are inference in the large hyperscalers [major cloud providers such as Amazon Web Services, Microsoft, Google]. Smaller [chip] companies like Cerebras and Groq that are having success with running inference in smaller data centers, some of them are just folks who can't get their hands on [Nvidia] GPUs. Right now, it is a little early because people might be looking at 'I've sunk all this money into GPUs for training, I've got a sunk cost in my data center on hardware, and I'm not sure I can flip it out for inference [chips made by other companies].' But I think over the next number of years, as the models get larger, more and more is going to go into specialized GPUs or solutions for training, and I think you'll see a divergence. You will start to see some specialized products for inference. All the hyperscalers are doing their own [chips]. I'm not going to talk out of school of confidentiality, but we have a good understanding of what all those things are, and in many cases, the CPU can help a lot."

**Who could take market share from Nvidia in the inference market? Cloud providers? Startups? AMD?**

"I think it's a little early to handicap it. Nvidia is a phenomenal company. They've got a great position in terms of the installed base. On the other hand, there's a lot of things going on with innovation [by chip startups], whether it's around fiber optic substrates [which improve performance compared to traditional electrical parts that connect chip components], co-packaged optics, different memory configurations, in-memory compute [speeding up data movement between the memory and processing units]—it's a very interesting time in the semiconductor world, where for many years there was not a lot of innovation."

**What are you sensing from the large cloud providers in terms of their custom AI chips?**

"We're kind of still waiting for them to take off in a big way. Do you see there being a specific year when that happens? It's really, really hard to do. You know, we've had a lot of experience just on general-purpose compute with [Amazon's] Graviton [chips for non-AI tasks] and also with Microsoft and Google, [which design AI chips]. Graviton already has got a lot of momentum....Microsoft and Google aren't there yet [with AI chips that appeal to their cloud customers], but it's taken a lot of time. And I think the challenge is that the models are moving so fast relative to the hardware that [for the cloud providers] to do their own things, it's really, really challenging."

**In five years, could you see [companies like AWS, Microsoft, Google or Meta Platforms] no longer investing in [their own AI] chips?**

"It's possible. I mean, it's a big investment for companies whose end business is either selling ads or selling subscriptions or running Instagram to have a core team developing chips for themselves. It's very, very tough to do in terms of just company focus and in orientation. Nvidia has done an amazing job of just staying way ahead of the competition."

**Do you have a perspective on the chip OpenAI is developing or think developing a chip might be something XAI considers doing?**

"You know, I can't say, because as you can imagine, if these folks are doing custom chips, they're probably talking to Arm."

**There are reports that Arm is also launching an AI chip this year. Are you?**

"I will promise you, if and when we're ready to do it, you'll be the first to know."

**What do [XAI's] Grok model advances mean in terms of the need for AI infrastructure and sites like [XAI's] Colossus supercomputer?**

"The one thing probably more than anything else was the example that Elon [Musk] put that thing up just so fast. I think that puts huge, huge pressure across the entire ecosystem."

**Do you think there will be more Stargate-like projects in the U.S.?**

"I don't know if SoftBank can do many more Stargates, but in terms of the idea, it's a good question. [Masayoshi Son, CEO of SoftBank] connected to something very early on that made a lot of sense: You need access to capital; you need access to other assets. One thing that's not well documented is that SoftBank has a very large energy business, SoftBank Energy, which does a lot of work for hyperscalers [like] Google and Microsoft. SoftBank [also] brings something very unique to the party, including Arm [and] a lot of technology partners. And I think that's a bit of what it's going to take. I think just the pure capital financing piece of it, without bringing technology, is going to be tricky. I think other people will try it, I don't have any doubt of that, but to pull it all together, it's going to be challenging."

**It's still unclear to me if AI inference will primarily run from data centers or if it will primarily run from edge devices, like phones. What is your perspective?**

"It's not illogical to think that you could be running some significant AI workloads on an edge device. You'll always, I think, have some level of hybrid—some running in the cloud, some running locally. There may be some things that...from a privacy standpoint, a security standpoint...you're going to want to run...locally. Or even a smaller device, like earbuds. There's only so much power that goes into those devices. So what has to happen is the AI algorithms and the compute associated with them to run the inference has to get a lot more efficient."

**What kind of device makes the most sense here?**

"The challenge with the phone is that it's a pretty efficient product. It does a lot of things. And what the phone has taught us over the last 15 years or so is people don't like carrying multiple things that don't really add to the experience. So I think it'll have to be something that kind of adds on to what you're doing. Could it be clothing and apparel [to which AI devices can be attached]? It actually could. Obviously, there's a lot going on with eyewear and things of that nature. It needs to be somewhat natural. But I could see accessories adding on the things that you have today that do some level of enhancements. I haven't yet seen the 'Oh, I got to have this thing and carry it as well.'"

Anissa Gardizy is a reporter at The Information covering cloud computing. She was previously a tech reporter at The Boston Globe. Anissa is based in San Francisco and can be reached at anissa@theinformation.com or on Twitter at @anissagardizy8.