Arm Releases First Ever AI Chip, With Meta As Initial Customer

[0:00] Chips with designs by Arm are inside nearly every smartphone, Apple computer, [0:06] Nvidia server rack, Amazon, Google, and Microsoft data center. Now, Arm's going to [0:11] compete with most of these mega customers too, joining their ranks as a [0:15] fabulous chip company because it's now in a whole new business, making physical [0:19] silicon. It was really designed for the AGI era and that it's been incredibly [0:24] low power, it's incredibly high performance, and it's sort of ruthlessly [0:28] optimized. This is Arm's new in-house central processing unit, Arm AGI, built [0:34] specifically for power efficient inference with Meta as its initial [0:38] customer. It's just not an option in the world for us, right? It's all about [0:43] supply chain diversity. For the first out chip to be embraced by Meta, I'll admit [0:47] I'm a little bit surprised. It's very tough to get the first chip right. We're [0:52] here in Austin, Texas for an exclusive first look at this brand new chip lab [0:56] where Arm's making it happen. We literally have a bunch of servers [0:59] in there and people all around the world are logged into it and working on [1:03] optimizing the silicon and getting it ready for our customers like Meta, [1:07] customers like OpenAI, F5 Networks, SAP. For more than 35 years, the UK-based [1:12] chip design firm has been selling its architecture, or instruction sets, to [1:16] almost all the largest chip makers and collecting royalties on every Arm-based [1:21] chip too. It offers a lower power alternative to the traditional x86 PC [1:26] and server architecture used by CPU giants Intel and AMD. It's been a [1:29] long time since we've been able to build a chip like this. We really don't [1:31] even know how we're going to power data centers that we're building in 2028 and [1:35] 2029. And Arm is really good at low power at a specific performance level. [1:42] But what do Arm's top customers think of Arm becoming a competitor in the race [1:46] to build the best AI chips? We talked to top leaders at Meta and Arm to find out, [1:51] and asked why CPUs are having a renaissance as agentic AI takes off, and how the new [1:56] Arm can succeed in such a crowded field of AI chip makers today. [1:59] Before this new chip, Arm didn't make anything tangible. Rather, selling [2:13] licenses to its core designs as an alternative to traditional x86 architecture, [2:18] the instruction set that's dominated since Intel introduced it nearly 50 years ago. [2:22] x86 has been tried and true. It can run pretty much anything. What Arm has [2:29] brought to the table with the royalties and licenses is you can just create the chip that [2:35] you want with nothing else. [2:39] Basically, from 12 founders working out of a turkey barn in Cambridge in 1990 to becoming [2:43] the lead chip architecture in almost every mobile phone by the early 2000s, thanks to [2:48] early deals with Apple and Texas Instruments. [2:51] Then, Arm broke into architecture for data center chips with the launch of its Neoverse [2:55] platform in 2018. Amazon took it mainstream that same year when it launched its first [3:00] custom processor, Graviton, based on Arm Neoverse. [3:04] Muhammad Awad joined Arm in 2018, too. [3:06] It's been over 1.25 billion chips. [3:09] Chips shipped into the data center based on Neoverse since we launched. [3:13] Neoverse gained traction as more and more hyperscalers got into custom silicon. [3:18] By the time Arm went public for the second time in 2023, it was the largest public offering [3:22] in nearly two years. [3:24] Its instruction sets, which are more customizable than x86, allow fabless chip makers like Amazon, [3:30] Google, and Microsoft to tailor their designs, integrating other Arm technologies like interconnect, [3:35] memory support, and compute subsystems. [3:38] If Arm didn't exist, [3:39] then all of those companies that have their own processors wouldn't be able to create [3:45] their own. [3:46] Arm's long held the building blocks to make a CPU. [3:49] So why choose now for its next big move? [3:52] In the agentic world, what we're seeing is that, you know, you've got all these GPUs [3:56] which are now firing off agents and all those agents need to go off and execute. [4:02] And it's the CPU that does the execution. [4:04] Before it was people. [4:06] People were the bottleneck in that process. [4:07] But now with AI spawning all these agents. [4:10] That bottleneck has been removed. [4:12] And so the demand for CPUs to keep up with all that is really going through the roof. [4:16] Arm has been hinting at this for perhaps the last four earnings calls. [4:20] Over a year, this lab has been in the works for 18 plus months. [4:24] Why is now the right time to launch? [4:26] One of the things that we always said was we were going to launch when we had the silicon [4:29] back. [4:30] We knew it was working well. [4:31] We knew that we had customers, strong customer demand. [4:33] You know, you think about all those tick boxes and we're there. [4:37] Congratulations on launching Arm's first data center chip. [4:38] We're so proud of you. [4:39] Thank you. [4:40] Thank you. [4:41] Congratulations to Arm. [4:42] Congratulations. [4:43] Congratulations. [4:44] Indeed, dozens of customers who license Arm's architecture to build their chips have publicly [4:48] endorsed Arm's break-in to compete with them in physical silicon. [4:51] It's a trillion dollar market. [4:53] And what we're seeing over and over again is actually our partners coming out and understanding [4:57] and realizing this is actually great for the industry. [5:01] Customers like AWS and Google and Microsoft and Nvidia have all come out and said, hey, [5:06] this is a good thing for the ecosystem. [5:08] At least eight customers. [5:10] Have already committed to purchasing Arm's new CPU and about 50 of Arm's existing customers [5:15] signaled support for the chip at the event where it was announced. [5:18] Big names like Nvidia, AWS, Google, Samsung, Broadcom, Microsoft, Micron and TSMC. [5:26] A lot of your customers have been supportive. [5:29] Not everybody. [5:30] Where was Qualcomm today? [5:31] I think that's a question for Qualcomm. [5:33] You know, you know, they've been a strong partner of ours in the past. [5:37] I look forward to having a strong relationship with them in the future. [5:40] Qualcomm bases its flagship Snapdragon PC and smartphone processors on Arm. [5:45] But it's working on an Arm-based server CPU that will compete with this one. [5:49] And the two have traded lawsuits since 2022 over the right to make certain chips on Arm [5:54] technology. [5:55] How about AMD? [5:56] How do you anticipate that reaction? [5:57] AMD is another great partner of ours. [5:59] If you look at kind of some of the stuff that they're doing on the FPGA side, Xilinx and [6:03] some of the embedded platforms, you know, those are all areas where we are very deeply [6:07] engaged with AMD. [6:08] And so I look to continue those strong partnerships. [6:10] I neither want to minimize nor maximize their potential entry into the merchant silicon [6:17] market. [6:18] I mean, we are going to run as fast as we possibly can to meet our customer requirements. [6:24] And competition overall is a good thing. [6:26] It keeps us on our toes and keeps us running very fast. [6:28] AMD is second only to Intel in server CPU market share. [6:32] The inventor of x86, Intel also runs a foundry for manufacturing chips. [6:37] We obviously compete with Intel in some areas. [6:39] But we partner with them in a bunch of areas as well. [6:41] If you look at their IPU products, these are their networking products that go into managing [6:45] the data center. [6:46] They're all based on Arm. [6:47] They're a licensee of ours. [6:49] You know, we've worked closely in the past and continue to work closely with Intel foundry [6:53] and looking for ways to partner with them in the future. [6:56] Like nearly all fabless AI chip makers, Arm currently manufactures its CPU at Taiwan Semiconductor [7:02] Manufacturing Company. [7:04] Made on TSMC's 3 nanometer node, Arm's CPU is entirely fabbed in Taiwan for now. [7:09] Although TSMC does have a chip manufacturer in Taiwan. [7:10] We have a 3 nanometer fab coming to Arizona soon. [7:13] Do you have plans to bring it to Arizona when the 3 nanometer fab opens? [7:18] You know, we would love to manufacture here. [7:20] It really comes down to, you know, what our customers are looking for. [7:23] Once the chips come back from TSMC, Arm needs to test early iterations. [7:28] So Arm built this in under two years. [7:30] An entirely new $71 million lab with three separate rooms and a ballooning team in Austin, [7:36] Texas. [7:37] You can see we're in an active construction site. [7:39] When we started in Austin, it was well over $20 million. [7:41] And we were probably, you know, 10 to 15 people. [7:44] Today we're over a thousand, growing rapidly. [7:47] The initial bring up for the chips happens at this first room in the Austin lab. [7:51] This is a validation board. [7:52] This is the board they get plugged into to kind of allows us to basically pull out all [7:55] the different signals from the chip and just ensure that everything is operational as we [8:00] would expect. [8:01] Next, the chips are put into server racks next door for more testing. [8:05] Underneath these heat sinks, you have two different Arm AGI CPUs, but they are connected [8:11] and acting as a single CPU. [8:13] Hundreds of cores that are acting together, hundreds of cores that can access all the [8:15] memory on the system. [8:17] This is great for things like database workloads. [8:20] Then they'll come here to a third and final stage of testing, eventually. [8:23] This is going to be a full set of automated tester equipment so that we can do things [8:27] like failure analysis, so that we can do pre-screening of parts before they go out. [8:31] We're literally building the runway as we're taking off in some cases. [8:34] Finally, up to 64 of the new CPUs, that's a total of 8,700 cores, are placed in server trays, and they're all connected to the CPU. [8:36] And that's what we're looking for. [8:37] And that's what we're looking for. [8:38] That's what we're looking for. [8:39] And that's what we're looking for. [8:40] That's what we're looking for. [8:41] These are cores that slide into a full CPU-only air-cooled rack, something Arm tests in a [8:45] small mock data center on site. [8:48] It's a dense configuration Arm is betting will appeal to data center customers around the world. [8:52] You can get two times the performance per watt than you can from an x86 rack. [8:57] So that means twice as much performance in the same footprint, in the same power. [9:01] How much of a game-changer is that when there's a lot of power constraints facing data centers? [9:05] It's huge. [9:06] So if you think about a one gigawatt data center, there's something like 10 million CPU cores [9:10] today that are going to be on that. [9:11] Yeah. [9:11] So if we're going to be in that with agentic, we think that's going to grow to 100 million [9:14] or more. [9:15] That's going to represent an incredible amount of power. [9:17] Awad says the power efficiency largely comes from Arm's decision to ruthlessly optimize [9:22] the AGI chip specifically for artificial general intelligence, hence the name. [9:27] It's not worried about supporting a lot of those legacy applications and that legacy [9:30] software that in some ways becomes a burden for some of the existing designs because it [9:35] creates overhead in terms of silicon, it creates overhead in terms of power, it creates overhead [9:39] in terms of performance. [9:41] In AI's world, wattage is like liquid gold. [9:45] So you can imagine a world where if you have a best-in-class CPU that's giving you the [9:51] best performance per watt that you possibly can, that opens up more wattage for other [9:56] parts of your infrastructure, right? [9:58] Memory is another necessity in short supply for AI chip makers, now including Arm. [10:02] We had a pretty good understanding of the timeframe at which it was going to go to production. [10:07] So we've been working to secure those supply chains for quite some time now. [10:10] Waifer capacity is another necessity. [10:12] It's another industry-wide crunch. [10:14] Concerns about getting enough 3 nanometer capacity, TSMC's most advanced nodes are quite [10:18] in demand. [10:19] Yeah, absolutely. [10:20] Demand is a challenge. [10:22] But because of our relationships, because of that forward-forward planning that we've [10:24] worked through with them, we're feeling pretty good about where we're at on that. [10:31] Unlike other tech giants like OpenAI, who've quietly pulled back on data center plans, [10:36] Meta remains all in on its AI infrastructure spending spree, making it a logical choice [10:41] as the first customer for Arm's CPU. [10:42] It's really a way for us to have a better future. [10:43] It's really a way for us to have a better future. [10:44] It's really a way for us to have a wide, diverse set of options to purchase from. [10:50] It's like in today's world, you really only have a couple of players. [10:54] And so this adds yet another player to the ecosystem for us. [10:57] Paul Saab has been with Meta for 18 years, and he's one of five engineers making sure [11:02] every piece of Meta software can run on Arm. [11:05] It will power any workload that we throw at it. [11:09] It was meant to basically be a full replacement, drop-in replacement for compute CPUs. [11:11] And it's a great way to get the best out of your system. [11:12] It's a great way to get the best out of your system. [11:13] It's a great way to get the best out of your system. [11:14] It's a great way to get the best out of your system. [11:15] And be transparent to our developers. [11:18] Saab was also around when Facebook launched the Open Compute Project in 2011, now with [11:23] hundreds of member companies like Arm and Nvidia committed to open hardware designs [11:28] that help reduce data center energy consumption and costs. [11:31] The first conversations we had with Arm were, hey, if we build this, we don't want to keep [11:37] this only within the company. [11:39] We're not like a chip company that's trying to build sales channels to sell chips. [11:43] We wanted it to be available to the whole world. [11:45] Meta has a total of 30 planned or operational data centers, from its Hyperion site with up [11:50] to five gigawatts of capacity in Louisiana to one gigawatt sites under construction in [11:55] Ohio and Indiana. [11:57] Meta is also reportedly looking to lease space at the giant Stargate site in Texas, where [12:01] OpenAI and Oracle scrapped plans to expand up to 10 gigawatts of capacity. [12:06] Meta got into a little bit of a bind with their Lama models. [12:09] Their last Lama models that came out in operation, they didn't work as well. [12:14] So they got behind. [12:16] And they also recognized we don't have enough compute power to do what we need to do. [12:23] And they increased capex, they built new data centers, they went to third parties and got [12:30] capacity from them. [12:33] And to fill all these data centers, you need a lot of chips. [12:35] Meta's filling those data centers with GPUs from both Nvidia and AMD, with huge new deals [12:40] announced in February. [12:42] And with its own line of Meta training and inference accelerators in the future. [12:43] Meta is also looking to build a new data center, which will be built in the future. [12:44] Meta is filling those data centers with GPUs from both Nvidia and AMD, with huge new deals [12:45] announced in February. [12:46] For that, it's been making since 2023, unveiling four new MTIA chips in March. [12:50] CPUs to power all those accelerators will come from Nvidia and now Arm. [12:54] In a way they helped co-develop this processor. [12:59] Although financials of the deal weren't disclosed, Meta announced in January that it plans to [13:04] spend between $115 and $135 billion on AI in 2026. [13:09] For Arm, there's big opportunity in partnering with the social media giant that makes about [13:13] 50 times more in annual revenue than it did last year. [13:15] it does. I also think that the company is looking for higher revenue [13:20] numbers. Let's say they get 5% of Meta's $115 to $135 billion CapEx going into the future. [13:29] That is a game changer on the on the top line for them. [13:33] Awad says ARM's CPU isn't currently export controlled, so it can also be sold in China, which [13:38] made up about 19% of ARM's revenue in 2025, a percentage that's been steadily declining since [13:44] 2023. As for which U.S. customers might buy the ARM CPU next, NVIDIA is especially [13:49] interesting. NVIDIA tried to buy ARM for $40 billion, but was shut down by regulators in [13:54] 2022. NVIDIA has since sold off the rest of its stake in ARM. [13:59] How is that relationship now? [14:01] Our relationship is great. They're one of our strongest partners. [14:03] We partner with them on all aspects of their designs. [14:05] On our side, we look at our ARM AGI CPU as something that may exist alongside some of their [14:11] products in a data center. [14:13] We don't see any reason why it can't. [14:14] I think Jensen knows he needs more CPUs to sell more GPUs, so he's very pragmatic. [14:20] ARM wouldn't disclose pricing for the CPU, but called it competitive, saying it aims to help [14:24] companies get access to compute even if they can't afford to make their own in-house [14:28] processors. Moorhead predicts it'll be in the thousands of dollars. [14:32] You know, there's a lot of ways to get ARM data center chips today, but you have to have 1,000 [14:37] engineers, $500 million budget to to go. [14:43] And. [14:44] Create it. So there's definitely a market need. [14:48] Is there a proving point that that we should wait to see to say, all right, ARM did this, it was [14:52] successful, the yields were good. [14:54] We've got AI workloads running at scale on ARM made CPUs. [14:59] Well, we've got silicon back now. [15:01] It's running real workloads and we've got silicon in customers hands in the labs. [15:06] They are actively testing it, verifying it, qualifying it to go to production. [15:11] And we expect to be in production later this year.