Arm Holdings CFO Jason Child said the company expects more than $2 billion in customer demand for its AI chips across fiscal years 2027 and 2028—a figure far larger than Arm anticipated when it first entered chip manufacturing. The gap between that demand and Arm's ability to fill it is the defining constraint on the transition from chip-design licensor to chip maker.
"The opportunity has gotten much larger than what we'd even planned when we started it," Child said. The bottlenecks are now visible: memory, fabrication capacity and power availability.
Arm announced its AI chip entry in March with the AGI CPU, a processor designed for agentic AI workloads—applications where AI systems operate autonomously over extended tasks. The AGI CPU sits alongside Nvidia GPUs in the same server racks, giving Arm a footprint in the infrastructure Nvidia already dominates.
Child was direct about the supply wall. "When you're trying to build chips, you have to kind of get in line," he said. Because Arm is starting from zero in manufacturing rather than expanding existing production, it will take a couple of years to ramp up its share of fabrication and memory capacity. Taiwan Semiconductor Manufacturing Co. has disclosed its own wafer and throughput constraints publicly. Child named TSMC capacity and memory as Arm's two biggest bottlenecks. "The demand on compute is for the most part insatiable," he said.
To close the supply gap faster, Arm is acquiring smaller companies. Child cited the purchase of networking startup DreamBig as the model. The deal gives Arm an integrated compute-and-networking solution—Arm's Neoverse compute cores combined with Arm-designed networking IP in a single system—reducing system cost and improving efficiency. The logic mirrors Nvidia's 2019 acquisition of Mellanox, which gave Nvidia end-to-end control of the interconnect layer in its server platforms.
Arm has completed roughly 20 acquisitions over the past 20 years, mostly of smaller private companies. Child said the company evaluates deals on two criteria: whether a target expands Arm's total addressable market or increases its ability to develop capabilities internally. He did not rule out larger transactions. "Large M&A is certainly a capability," he said.
"Delivering silicon is definitely more complicated" than licensing chip designs, Child acknowledged—a direct statement that Arm's new business model carries operational weight its licensing model did not. As a licensor, Arm collected fees and royalties without touching manufacturing. As a chip maker, it must secure fabrication slots, lock in memory supply and manage power constraints at data center scale.
The financial market has priced in the transition's success. Arm ADRs are up 152 percent year-to-date and 101 percent over the past 12 months, outpacing the iShares Semiconductor ETF's 127 percent gain over the same period. Arm's market value has surpassed $300 billion, with SoftBank holding an 87 percent stake.
Arm reported 22.4 percent revenue growth in its most recent earnings, driven by higher licensing and royalty fees as AI chip designers built more Arm-architecture cores into their products. That growth came entirely from the licensing model; the AI chip business has not yet contributed meaningful revenue. The $2 billion demand figure for fiscal 2027 and 2028 represents the revenue line Arm is racing to open.
The supply math is harder than the demand math. High-bandwidth memory—the type used in AI accelerators—is sold out at SK Hynix through the end of 2025. TSMC's advanced node capacity is similarly constrained, with major customers including Apple, Nvidia and AMD competing for the same wafer slots. Arm enters that queue as a newcomer with no existing volume relationship, which is why Child's two-year ramp timeline is the operative constraint on when the $2 billion demand converts to actual revenue.
The DreamBig deal also reflects a deliberate effort to differentiate on system integration rather than raw compute. For AI inference workloads, the speed at which data moves from memory to processor—memory bandwidth—determines latency more than clock speed. A chip that integrates compute and networking in one package reduces that latency and cuts the number of components a data center operator must buy and manage separately. That system-level value proposition is how Arm intends to compete against established AI chip vendors without matching their manufacturing scale on day one.
