Arm Holdings CEO Rene Haas at Semafor World Economy

Yeah. No, that's right. And here it is. By the way, this is the ARM AGI CPU. It is the first chip offered by ARM. And one of the trickiest things about this so far has been ARM has been delivering the IP, the blueprint for people who make chips for 33 years. So, when people ask me, you know, what does ARM do? I could just kind of rehearse that statement. And now that we've got this product, I have to think about what that opening looks like. Why did we do this? One simple reason is the market asked us to. The demand for compute is off the charts as everyone knows with AI. One of the big areas that people are demanding more and more performance around is these CPUs that drive all these agents. You know, if you have agents bombarding the data center looking to increase workloads, the only chip that can do that is the CPU. So, Meta, the partner we worked with on this, literally came to us some time back and said, 'Hey, rather than licensing the IP from you, can you build the chip for us?' And so, we did.

No, I mean we dealt very extensively with Santos who runs the data center business and Mark was very very supportive of the effort as well, but Meta's got a lot of things on their plate, right? They do a tremendous amount of things on their own dime. They do their own chips of course, but when they looked at the ROI of doing the chip themselves using the ARM IP or having someone like us deliver the solution to them, we both collectively looked at it and the answer ended up being, you know what, the better answer here for both companies is ARM is the supplier.

That's right.

Including me now.

Yeah. No, that's the right question to ask. Yes, TSMC is the builder of the chips for us. So, we don't build chips ourselves. The folks who aren't familiar in the audience, something like this has a lot of people involved in the supply chain who actually make the device. The chip itself is fabricated at a company called TSMC. They're the largest in the world. Then it's packaged by one of our partners in Korea. We have a long relationship with TSMC because many of the chips that have ARM today. In fact, a lot of TSMC's output uses ARM. So, we knew the people at TSMC, we knew the engineers, we knew the management. So, when we came to them and said this was something we were looking to go off and do, we understood and knew how to negotiate for supply chain for allocation etc etc. Demand has been off the charts for this. We are looking for more capacity. If there's anyone from TSMC in the room, we are looking for more capacity. Because if we had more capacity, we could sell more.

Yeah, they are very very disciplined in ensuring that when they have a relationship with a partner like ourselves that there's a relationship downstream, there's a customer, there's demand. So yes, there was a lot of checks and balances in the system and Meta was great in terms of partnering around that. And just to explain a little bit more deeply about this capacity problem because people hear gosh we need more chips, we need more memory, the amount of time it takes to build one of these factories and then order the equipment to go into those factories is a minimum of three years, probably five years. The capital is tens of billions of dollars. So if you're in TSMC and you're having to make a judgment call about how much demand will I have, you need to think three to five years out. So here we are in 2026. In 2022 there was no ChatGPT moment. There was no Anthropic, there was no huge demand for these AI data centers. So what you're seeing here is the result of, and they weren't the wrong decisions at the time, but they were decisions made based upon the best lens that they had relative to the overall demand picture. And here we are. And in the memory case, it's even more acute. If you go back to 2022, 2023, you looked at Micron and Samsung and SK Hynix, they're the largest suppliers of memory. They were all losing money. They were all negative. So if you're in that world in that time and you're having to now think about expanded DRAM capacity in 2026, you're going to be conservative.

Yeah. One thing, and I was really struck by, and I lived in China about 10 years ago. I was there 2014, 2015, 2016, right after I joined ARM. I came over and joined them in China for a number of different roles and at that time China put together a Made in China 2025 program. If you look back, Xi Jinping put some policy in place and at that time the imports of semiconductors were greater than the imports of energy and China as we know has a lot of capability on the energy side but imports a lot. So they were very very focused on semiconductor supply chain whether it came to fabs, equipment, people who could design chips etc etc. They were very circumspect about that policy and the Made in China 2025 policy around semiconductors was really about being self-sufficient. That was one of the goals. Now they're not there yet, but they're making lots of headway in that space. And if you look at China's history over 5,000 years, self-sufficiency is a mantra in terms of how China's operated. So fast forward, it's starting now to some level, but not so much on the global side that people are realizing semiconductors are incredibly important. They power everything that we do. If anyone lost their key fob during COVID and it took a year to get it back, that's because the tiny chip that was inside there was way under supplied and underbuilt and you couldn't get any of them. So to your big idea, I think the world needs global resiliency. I think there should be fabs on every continent. I think there should be equipment makers on every continent. I think we should be looking at getting private capital involved in more and more startups that can do the type of heavy equipment that's required and then ultimately have these factories because the other thing, it's good hygiene to have supply chain resiliency just the way countries would love to have their own energy sources, countries should have their own sources for chips.

This is why you're seeing data centers being built across the world. This is why you're seeing demands for energy. It's really this compute. And what does that mean in terms of compute? When you type in a query into ChatGPT or Claude and it spits back the answers incredibly fast for you. Those are tokens. Those are words being generated by some data center. And those words cost money to generate. Someone's thinking to create the word. That's compute. Those words have to be distributed, etc., etc. And for those who use those tools, you'll now notice that ChatGPT always asks you a follow-up question. You know, would you like me to do this? Would you like me to do that? And the engine continues to feed on itself. So, it's really the demand for compute. So, it's Meta, it's OpenAI, it's Microsoft, it's Amazon, it's Google, it's everybody.

I would hope so. We're certainly trying. They are absolutely all candidates to buy our chip for sure.