2026-05-18 — views $ARM · ARM Holdings · Cortex-X / Neoverse / Mali / CSS
ARM Holdings — the IP layer under every smartphone and increasingly every AI server
ARM licenses CPU IP under 99%+ smartphone SoCs and is the structural read on AI-server CPU shift (NVIDIA Grace, AWS Graviton, Apple). Arm v9 royalty runs 30-50% above v8 per chip — the v9 mandate is the pricing event.
ARM Holdings (NASDAQ: ARM) is the IP layer underneath 99%+ of smartphone SoCs and an increasing share of AI-server CPUs. The business model is pure license + royalty — ARM doesn’t fab chips, it licenses the instruction set + reference cores, then collects per-unit royalty when partner silicon ships.
What ARM actually sells
| Product line | What it is | Where it ships |
|---|---|---|
| Cortex-A / Cortex-X | High-performance application CPU cores | Apple A-series (architectural license), Qualcomm Snapdragon, MediaTek Dimensity — every flagship Android |
| Cortex-N / efficiency cores | Low-power CPUs | IoT, wearables, low-tier mobile |
| Mali / Immortalis GPU | Mobile GPU IP | Mediatek, Samsung Exynos, low-end Android |
| Neoverse V/N/E | Server CPU IP for cloud + AI | NVIDIA Grace, AWS Graviton, Ampere, Microsoft Cobalt |
| Compute Subsystems (CSS) | Pre-validated chiplet/IP packages | New 2024+ business line — sells whole subsystems, not just cores |
The structural pricing event: v9 mandate
The non-obvious story is Arm v9 royalty pricing. Where Arm v8 royalty rates ran ~1-2% of chip ASP, Arm v9 commands ~30-50% higher per-unit royalties because the new ISA mandate (SVE2 vector, MTE memory tagging, confidential compute) is non-optional for new flagship designs.
In practical terms: every iPhone, Galaxy, and Pixel SoC built in 2026+ pays ARM more per unit than the same vendor paid in 2024. The royalty escalator is automatic; ARM doesn’t have to add new customers, just wait for the existing flagship pipeline to refresh.
The AI-server angle: Neoverse + CSS
| Customer | Product | Status |
|---|---|---|
| NVIDIA | Grace CPU (paired with Hopper/Blackwell) | Shipping volume to hyperscalers |
| AWS | Graviton 4 → Graviton 5 | ~50% of new EC2 capacity is Graviton |
| Microsoft | Cobalt 100 (Azure) | Production deployment 2025 |
| Axion (GCP) | Production 2025 | |
| Apple | M4/M5 (architectural license) | Mac + on-device AI workloads |
| Ampere | AmpereOne / OneFamily | Smaller share, present in Oracle Cloud |
The Compute Subsystems (CSS) business sells pre-integrated chiplet packages — not just core licenses — to customers that don’t want to build a server CPU from scratch. CSS shortens time-to-silicon from ~3 years to ~12-18 months, which is what’s letting AWS, Microsoft, and Google catch up on custom silicon vs. NVIDIA Grace’s head start.
Why this matters for AI infrastructure
Two threads:
- Server CPU shift is real. Every cloud hyperscaler is building or buying Arm-based server CPUs for AI-adjacent workloads (control plane, data preprocessing, inference orchestration). x86 is still the volume base, but Arm share of new datacenter purchases is rising ~5-10 points/year. Each new design pays ARM royalty.
- The royalty tail is wide. ARM’s revenue grew 19% YoY through fiscal 2025 with ~95% gross margin — there’s no fab cost, no inventory. As long as the v9 mandate holds and Neoverse adoption keeps expanding, the model is structurally levered to total AI silicon volume without taking on fab risk.
What’s risky
- NVIDIA’s $50B+ commitment to ARM-based Grace is the single biggest concentration. A pivot to RISC-V or in-house ISA on NVIDIA’s part would hurt — but the timeline to replace Arm v9 with anything else is 5+ years.
- Royalty disclosure is opaque. ARM doesn’t break out royalty rate by customer, so estimating the actual v9 lift requires modeling. Bears argue the v9 premium is smaller than bulls claim.
- China exposure (~20% of revenue) is exposed to US export controls. Neoverse V3 has restricted licensing into PRC; if rules tighten, royalty growth from China could turn flat.
Practitioner note
For AI-product builders:
- If you’re deploying Arm-based instances (Graviton, Cobalt, Axion) for inference orchestration, you’re already benefiting from ARM’s volume economics. Run a price comparison on Graviton 4 vs equivalent x86 Intel Sapphire Rapids — typical workload sees 20-40% lower cost-per-vCPU on Arm.
- For on-device AI (mobile, edge, Apple Silicon), every device you ship pays ARM royalty embedded in the chip BOM. Not actionable per-app, but worth understanding the layer that taxes your hardware ecosystem.
- If you build server hardware, Compute Subsystems (CSS) is the leverage point — it’s how custom server CPUs ship in 12-18 months instead of 3 years. Cloud vendors quietly building “ARM but ours” SoCs is the trend that will define 2027 cloud capex.
The under-considered angle: ARM is the only infrastructure IP company where every iPhone, Android, server, and AI cluster pays royalty. That cross-modal exposure — mobile + server + auto + IoT — is what makes the business asymmetric vs. pure-server names like AMD or Intel. The bear case requires multiple unrelated demand pillars to break at once.