Arm’s 98.5% margin business just started shipping chips

Sarah Guo and Elad Gil introduce Rene Haas as CEO of Arm and of SoftBank Group International. Arm’s primary business is still licensing the CPU cores that end up in smartphones, data-center servers, and cars. Some customers, such as Samsung, fab the chips themselves. Most take a design to TSMC. Because those customers span so many markets, Haas says Arm sees the supply chain from every angle: who is short of capacity, who is expanding, and where the pinch is this quarter.

That panoramic view just got more personal. Last March Arm introduced its first physical product, the Arm AGI CPU, later shown at Hot Chips. The company now has to buy substrates, wafers, and memory. Haas’s line is that they are up to their waist in the same soup as everyone else. Guo asks why, after a few decades of selling IP, they started making CPUs. He describes an evolution. Product cycles were not slowing down. Chip manufacturing times were stretching. Getting a solution out faster started to matter more than handing over isolated CPU, GPU, and system blocks.

The intermediate step was compute subsystems: not just the Lego bricks, but the blueprint for how to stitch them. On the IPO roadshow Haas used that Lego analogy. Skeptics said stitching was the chip designer’s job. Demand, he says, was insane, because the package saved time to market and a lot of cost. The physical chip was the next leap. Plenty of licensees can already build Arm-based silicon. Plenty of other companies want an Arm product and do not serve those markets. Meta wanted a general-purpose agentic CPU, nobody on the market could simply sell them one, and they asked Arm to do it together. That is how the company got into hardware.

Haas went around the customer base expecting a fight. CPU IP is only as good as the chip and software ecosystems around it. The pushback was lighter than he feared. More software in the wild, proprietary or open source, helps everyone who builds on Arm, including Nvidia, Amazon, Microsoft, and Google, all of whom make Arm-based server chips. When the product was announced, Jensen Huang, Microsoft’s Rani Borkar, Google’s Amin Vahdat, and AWS’s James Hamilton all showed up to congratulate them. The cost of that peace is operational. Arm remains a fabless company with no plan to build a fab, but it now needs supply-chain operators, memory allocations from Micron and SK Hynix, backend and layout engineers, and bring-up labs. Guo had discovered the old 98.5% gross margin and called it beautiful. Haas, who left Nvidia for Arm in 2013, remembers thinking: no inventory, no RMA, no scrap, what is not to like. Leadership hired from Broadcom, Qualcomm, and Nvidia is how they are putting that muscle back on.

Guo notes that OpenAI had just talked up a fast chip, Jalapeño, built with AI tooling, and that Amazon and others are bringing methods such as formal verification into training-chip work. Haas’s own view is bigger than EDA. He treats AI as a utility that will lift productivity in every industry, a great leveler for companies that are just getting started, whether the field is healthcare, infrastructure, or robotics. Inside Arm, non-engineers already use it everywhere. On the engineering side the gains have been huge.

A chip takes 24 to 36 months, depending on complexity. People assume the calendar is consumed by architecture and RTL. Haas says that is not the largest block. Verification, validation, debug, and documentation are. AI is good at those. He estimates that 80 to 90% of Arm engineers use the tools every day. Turning them off, in his analogy, would be like living in the 1990s with the internet and being told you may go online only between two and four, then walk down the hall to the library. It would be anarchy. “The genie’s out of the bottle, and there’s no stopping that.”

The tools are still weak where it hurts most: RTL generation and best-in-class physical design. Public models train on public data, and the important chip knowledge is proprietary. Arm is working with model makers to fine-tune on that gap. Haas is proud of a quieter advantage. The company’s IP portfolio is rich, and, more important, so is the documentation and the test benches that show how to build the blocks. He has watched other chip firms try to license valuable IP, then discover nobody wrote down how to use it. If it is unusable and untestable, it is untrainable. If it is untrainable, it has no value for AI.

Gil asks whether two or three years of better tooling shrinks the 24-to-36-month cycle to a year or six months. Haas will not pick that horizon. Five-plus years out, for some straightforward designs, he thinks you might go from idea to a GDS2 file, the layout you send the fab, which would take both design and a chunk of verification off the table. You will not press a button and get something 10% faster than Nvidia’s Vera Rubin, 20% cheaper, and 30% more efficient on a given model. In five to ten years, though, he expects the way chips are designed to look startlingly different.

A wider customer base of Veras and Gravitons is, Haas says, only good for Arm. What will matter more is supply-chain acumen: how a design actually becomes an end product. Young AI-chip companies have innovative architectures and plenty of funding. They still need capital at semiconductor scale, memory relationships, substrate vendors, a three-nanometer or six-nanometer line, and advanced packaging. He does not think that constraint lifts in twelve or twenty-four months. As long as the transformer is the unit of energy for training and inference, the workload is compute-heavy and memory-heavy. He expects at least three to five years of a constrained environment.

Gil recites the cascade. Two years ago the bottleneck was packaging. Now people talk about memory. What is next? Haas’s answer is building the data centers. Very few of the projects underway are ahead of schedule or using less labor than planned. Layer on political pressure in parts of the United States to slow or restrict data-center development, and infrastructure itself becomes a headwind. If it were not, wafer and memory capacity probably would be. Several “governors,” not the state kind, will take turns throttling growth. Asked about an AI bubble, he splits the question. Valuation bubbles are a stock-market index problem. Oversupply relative to demand is “not even close,” because the demand, given how the models work, is insatiable.

Guo asks what advice he has for founders in CapEx-heavy industries, given SoftBank’s leverage and Arm’s cash-flow machine. Arm is public, with one very large shareholder. Haas has informal investor meetings with Masayoshi Son constantly, and the two companies can do symbiotic things other chip firms cannot. For young companies his advice is to form strategic partnerships incredibly early: supply chain, private equity, banks. Semiconductors are back as a startup category after a long drought, and access to capital will still be the gate. SoftBank, he says, looks for portfolio companies it can backstop. It has also announced SoftBank Neo, an intent to become a neocloud that can be a first home for young chip firms that would otherwise have to win a design slot at Microsoft or Google.

His SoftBank Group International brief sits under Son’s public four-part strategy: robotics, OpenAI, infrastructure, and Arm. Haas is leading direction at Ampere, Graphcore, and Stack AV, an autonomous-truck company, but he says a better picture is that he is in the room for a lot of Masa’s discussions, helping formulate the strategy and then execute it. That seat gives Arm a bird’s-eye view of capital, energy, and infrastructure. It can also be a home for Arm products. Connecting SoftBank’s verticals to the AGI CPU does not mean Arm is entering the broad merchant-chip business. It might just mean building products back for SoftBank.

Gil’s snapshot of generalized robotics is that it looks like LLMs about a decade in: striking demos of task and environment generalization, even in-context learning, but not wide deployment. Haas agrees one hundred percent. Robotics 1.0 was purpose-built mechanics and software for one job. A new car line meant ripping up the cell. The barrier was high. Robotics 2.0, in his telling, is a machine that can learn from training or from what it sees, a mechanical design general enough to be reprogrammed, and falling cost. Construction, infrastructure, service, and security start to look doable, almost Jetsons. Instagram Olympic races between robots amuse him less. He does not think anyone will watch a robot sports league. The utility is replacing human labor.

On form factor he refuses the usual split. Some jobs are optimized around a person about six feet tall with arms of a certain length, so humanoids have a natural slot. Specialized machines will exist too. Either way, he says Arm will be everywhere: real-time sensing microprocessors out at the fingers, perception and sensing that are Arm-based. Whether the brain in a humanoid today is Nvidia’s or involves Qualcomm, most of those brains already run on Arm. The robotic industry, going forward, “will be powered by Arm.” Early pockets are still foggy because robots are expensive and the business models are not vetted. Distribution centers, factory automation, and last-mile delivery are the obvious first automations. Loosely speaking, he adds, an autonomous truck is a robot of sorts. Stack AV sits in that picture.

Guo later asks where the CPU opportunity actually is: the CPU in the rack, Vera and Graviton, versus agent sandboxes that must call existing software and APIs. Haas rewinds to the ChatGPT explosion, when every question seemed to have the same answer: the accelerator. His counter is blunt. No computing problem ever invented does not use a microprocessor. “It is the heart of everything. All roads lead through it, around it, past it.” As workloads move from training toward reinforcement learning and inference, something still has to orchestrate where the tokens go. The token factory generates them. CPUs are the trucks that take them to users. Until someone invents a system without that mechanism, demand for CPUs sits beside demand for accelerators, inside the same von Neumann picture of CPU, accelerator, and memory.

That picture, he says, is the same in a data center, a car, a robot, a phone, or a wearable. The edge is Arm’s sweet spot. More AI will run on small devices, and you cannot put a 50-watt GPU on your head. Local processing has to happen on a power-efficient CPU, which is the thing Arm has always been good at. The instruction set that was table stakes for running the device is now also the place where on-device AI can live.

Guo piles export controls, robotics rules, chip rules, and data-center rules onto the table and asks how protectionist the United States or the West should be. Haas puts on what he calls his American-citizen hat. Arm’s headquarters are in the UK; roughly 40% of employees are there, 30% in the United States, 30% in Asia. He grew up in semiconductors and remembers the 1980s, when Japan’s memory pricing took share and the U.S. stood up SEMATECH to refortify the industry. Then the internet and SaaS became the rage, and people forgot that semiconductors were a strategic asset. He wants as much of that technology on U.S. soil as possible, and he would say the same, at smaller scale, to the UK. Intel, Micron, more U.S. fabs: national security and supply-chain diversification, not nostalgia.

Export controls framed as stopping China from “winning the race” get a colder answer. His personal view is that it is an infinite game. There is not going to be a winner; the race is not going to be over. You could still end up in a world where a lot of the critical technology is not U.S.-based, which he does not think is a good thing. Cheap goods are not the whole story. Being at the forefront drives innovation and ecosystems. Postwar Detroit pulled Wisconsin, Ohio, and Illinois along with it. Data centers, he argues, work the same way. The cartoon of a Costco box with two cars in the lot and no jobs is one he calls out. Energy, liquid cooling, and the rest of the local stack are skilled work. The electricians’ union has asked not to ban data centers. Those are trained, certified jobs, and you need a lot of them.

Gil sees the anti-data-center campaign as more coordinated than grassroots. Haas’s hypothesis is fear: AI as job loss, even though he does not think the fear is well grounded. The boom has enriched some people and left others with a mortgage, a flat paycheck, and a new thing called AI. Data centers become the bullseye, down to people holding up supposedly tainted water. Gil adds organized media influence and says the industry has to tell the story of jobs and local wealth itself. Haas’s first principle is simpler. Smartphones, the internet, personal computers, fill in your favorite technology: “There is no downside from being the leader.” Gil sharpens it. To be the laggard is to have the script dictated to you. Places that are not in the lead carry the tax burden of being left behind. If AI is close to the final frontier of intelligence, of course you want to be on the wave.

What excites him at Arm in the next year or two is simply being in the center of that. All AI needs compute, and that compute needs to be power efficient, which is the road that still runs through Arm. He has spent a career asking whether the world needs another tablet at 8.9 inches or 9.2. The abundance of opportunity around AI, he says, is unlike anything in that career. He feels fortunate to be leading a company that sits in the middle of it.