ARM CEO Rene Haas on launching the company's first CPU product, AI's compute demand, and the SoftBank relationship

Arm's First Product, AI Compute, and the Control Tower View

Arm Holdings has spent decades licensing CPU architecture to chipmakers. In late March, it crossed into a new category: shipping its own product.

ARM AGI

The product is ARM AGI, a CPU targeting agentic AI workloads. Haas says it delivers twice the performance at equivalent power versus the competition, and demand has been "extremely strong" since launch. The timing was partly fortunate — agentic workloads spawn large numbers of parallel processes that lean heavily on CPUs, creating a demand surge that Arm's own product landed directly into. Arm is now, by Haas's description, feeling supply pressure from both sides: end customers pushing for more IP licensing, and its own direct customers pushing for more ARM AGI supply.

Moving from IP licensor to product company required new capabilities Arm didn't have internally. Haas brought in people with back-end design, customer validation, and supply chain operations experience — disciplines that don't translate from building IP. He runs the new division the way he learned from spending 25 years working alongside founders, including Jensen Huang at Nvidia and Masayoshi Son at SoftBank: deep in the operational detail, personally involved in supply decisions, and focused on what comes next in the product roadmap.

“We have huge, huge demand for the product, which is ARM AGI CPU driven by somewhat selfishly, it's a great product. It's two x performance at the same power as the competition... ARM's core business is compute. We do CPUs. And CPUs are table stakes. Every digital product needs a CPU no matter what it is... 350 billion devices have shipped based on ARM.”

Scaling, not moments

Haas is skeptical of the "AGI moment" framing. ChatGPT and Claude Code are mile markers on a longer scaling curve, not binary inflection points. The more significant frontier, in his view, is synthetic data — using compute and models to generate training sets for problems that don't have Internet-scale datasets behind them, particularly in engineering and science. He doesn't think scaling laws have broken down. More compute continues to produce more capable systems, and the hardware roadmap at Arm is already planning for products shipping in 2030, 2031, and 2032.

AI inside Arm

About 85–90% of Arm's workforce is using AI in some form. The most operationally interesting use case involves royalty forecasting. Arm's financial model requires accruing royalties one quarter ahead of collection, then reconciling the estimate against actuals in what Haas calls a "true up." The forecasting model functions like a macroeconomic hedge fund, pulling in market indicators to predict chip shipment volumes. AI has made those forecasts materially more accurate. Legal teams use it to review contracts against standard terms. Finance uses it across multiple functions. In meetings, teams use ChatGPT or Claude in real time to size markets and check facts — not always 100% accurate, Haas says, but "80 to 90% close enough to start making decisions."

Chip design and AI

Chip design's biggest time sink is verification and bug triage, not circuit design itself. AI is already helping Arm run automated bug triage over weekends, prioritize high-severity issues, and in some cases fix them without engineer intervention. The upside isn't dramatic yet — development cycles haven't been cut in half — but Haas argues they would be longer without the tools. Full AI-assisted chip design, where an idea goes from concept to tape-out at TSMC with minimal human intervention, remains years away. The datasets aren't mature enough and the closed-loop systems aren't in place, but he expects it to arrive.

Robotics

Haas puts the robotics tipping point around 2030, and thinks the industry will both take longer and grow bigger than current expectations. The bill of materials on humanoids won't drop to $4,000, in his view, making outright purchase economics difficult. Robots-as-a-service is the more viable model — subsidized units deployed as labor replacements in contexts where human workers are already scarce. The hardest unsolved problem is dexterous manipulation. Human fingers apply variable pressure, sense texture, and switch between threading a needle and lifting 20 pounds. Replicating that in hardware and software remains genuinely hard. Once it's solved, Haas thinks the impact on infrastructure, construction, and industrial labor will be transformational.

The SoftBank relationship

Haas first met Masayoshi Son in 2016, shortly after SoftBank acquired Arm. Son's original thesis centered on a trillion connected devices. His current views have evolved, but the operating style is consistent: big bets, long-horizon thinking, moving quickly on large trends. Haas describes his own role as translating that ambition into executable plans — he shares Son's appetite for scale but runs a ground-level operating business that has to ship products and hit quarterly numbers.

350 billion devices have shipped on Arm architecture, giving the company what Haas calls a "control tower view" of the entire semiconductor industry. That visibility — across every device category, every major chipmaker, every emerging workload — shapes how Arm builds its product roadmap and where it places long-term bets.