ARM — The Moment the IP Company Holding AI Compute's Instruction Standard Becomes a Silicon Company
The ARM architecture has captured half of AI data center CPUs. On top of the proven engine of rising v9/CSS royalty-per-chip mix, a second growth engine has been added 35 years on — its own in-house silicon AGI CPU. We dissect the compute architecture layer.
ARM is the company that holds the CPU instruction standard in the AI data center compute stack in a vendor-neutral way. It has already captured half of hyperscaler CPU compute, and on top of the proven engine of rising v9/CSS royalty-per-chip mix, a second engine has been added 35 years on — its own in-house silicon AGI CPU. The business and the moat are solid, but at roughly 215 times FY2026 non-GAAP earnings per share, flawless execution is already priced in, so the return depends on the timing of entry.
Reader's Brief — 30-second TL;DR
Advanced
Why Now
The key upside triggers are quarter-after-quarter upward revisions to the AGI CPU order backlog, continued doubling of data center royalties, and v9 attach rate entering its target band. The first variables to break are a major customer's switch to RISC-V (a royalty-free open instruction set) and the supply overhang created by SoftBank's roughly 90% stake. Qualcomm's June 24 Investor Day and the FY27 Q1 results on August 5 are the twin axes of summer volatility.
Winners ?? Losers
Beneficiaries: ARM (high-ASP v9/CSS mix + in-house AGI CPU silicon), the ecosystem of ARM-based hyperscaler in-house chips. Pressured/competing: the x86 camp (Intel, AMD) defending data center share, the RISC-V camp (Qualcomm/Ventana, backed by Google and Meta, plus China's ecosystem) defining the ASP ceiling. A formalized RISC-V switch by major licensees or a SoftBank block sale would be the first to trigger a de-rating.
1. Industrial Coordinate — Why a CPU Must Sit Beside the GPU
The infrastructure story always starts with the GPU. But no data center runs on GPUs alone. Open up a single AI server and a host always sits beside the multiple GPUs, and what that CPU does is half the system.
First, data supply. The GPU computes fast but cannot fetch its own data. Reading training data from storage, preprocessing it, and pushing it into GPU memory is all the CPU's job. If this supply is slow, the tens-of-thousands-of-dollars GPU sits idle waiting for data — the GPU idling the industry fears most. Second, command. The orchestration of splitting work across thousands of GPUs, coordinating communication, and removing failed nodes to restart all runs on the CPU.
So GPU investment and CPU investment are not substitutes but proportional. The larger the GPU cluster grows, the more CPUs to feed and command it must grow too. NVIDIA bundling its own ARM-based CPU Grace with the GPU in one package, and AWS, Google, and Microsoft each building their own ARM server CPUs — Graviton, Axion, and Cobalt — are the proof.
Power as the New Currency
The bottleneck in data center construction is no longer money but power. Because sites, substations, and transmission carry multi-year lead times, hyperscalers solve a constrained-optimization problem of extracting maximum compute within a secured power budget. Under this constraint, every watt is allocated to GPUs first, and the power budget allowed to CPUs grows ever tighter.
Here the architecture diverges. ARM, which started in mobile and made performance per watt the number-one principle of its design philosophy, holds a structural edge over x86 in running more cores on the same power. The harsher the power constraint — that is, the larger the AI buildout — the more mechanically ARM's adoption incentive strengthens.
Agentic AI — The Variable That Grows the Proportionality Coefficient Itself
If CPUs growing in proportion to GPUs is the first-order logic, agentic AI is the second-order variable that grows that coefficient itself. Single-shot question-and-answer inference has the GPU working and the CPU assisting, but agents are different. A single request becomes a workflow of dozens of steps — planning, tool calls, search, code execution, retries — and the branching, state management, and data movement between those steps is all CPU work. ARM estimates agentic AI lifts per-data-center CPU core demand to roughly four times prior levels, and cites a 2030 data center CPU market of over $100 billion.
Where ARM sits is the very top of this vast flow, the compute architecture layer. The CPU is not the GPU's shadow but a necessary layer that grows in proportion, and half of those CPUs already run on ARM.
Full access requires a free account
Sign in with Google to unlock the full body of every free report instantly.
This site runs on ads — the tier system rewards community contributions.
Comments
This report is provided for informational purposes only and does not constitute a recommendation to buy or sell any financial instrument. Investment decisions should be made based on your own judgment and responsibility. The analysis and opinions contained herein are based on information available at the time of writing and are subject to change.