Last price
$166.73
CPU ISA licensor; structural AI demand tailwind.
Live quote sourced from Yahoo Finance. Prices cited in narrative below reflect the original memo date and may be stale.
Arm licenses CPU instruction-set architecture (ISA) and designs for mobile, edge, and emerging server markets. As AI workloads shift to edge devices (phones, IoT, autonomous vehicles) and custom silicon becomes standard, Arm's footprint expands. Licensees (Qualcomm, Apple, Amazon) design custom cores around Arm ISA; more AI functionality per device drives more Arm design wins and higher royalties. Thesis: enabler—demand for Arm-based AI chips is structural.
ARM is structurally positioned as the inverse NVDA trade within the thesis. NVDA loses inference share to custom silicon; ARM gains royalty share because every custom ASIC that Meta, Google, or Amazon designs is an ARM ISA license. The thesis inversion is clean: NVDAs headwind is ARMs tailwind.
Meta MTIA, Google TPU, Amazon Trainium, Apple Neural Engine, Qualcomm custom cores—all are ARM-licensed ISAs. As hyperscalers and OEMs abandon merchant GPU and move to custom silicon, they still need an ISA. ARMs footprint in custom-silicon is nearly universal. Every custom ASIC design (and there will be 10x more by 2030) pays a design license (typically $1–5M per chip) plus per-unit royalties (typically $0.50–2 per chip). This is a high-margin, recurring revenue stream that scales with custom-silicon TAM.
| Licensee | ISA Use | Royalty Growth | Risk |
|---|---|---|---|
| Meta (MTIA) | Custom inference ASIC | ++ | Medium (captive, low switching cost) |
| Google (TPU) | Custom training/inference | ++ | Medium (captive, low switching cost) |
| Amazon (Trainium) | Custom inference ASIC | ++ | Medium (captive, low switching cost) |
| Apple (Neural Engine) | Mobile neural coprocessor | + | Low (Apple is unlikely to switch) |
| Qualcomm (custom cores) | Mobile/edge AI accelerators | ++ | Medium (competitive pressure from in-house) |
| Mobile SoC ecosystem | All Snapdragon, Exynos, MediaTek | + | Stable, no switching risk |
Hyperscalers (Meta, Google, Amazon, Microsoft) are designing new ASICs every 18–24 months. Each design pays ARM. At 5+ hyperscalers × 2+ designs per year, ARM gets 10–15 new design licenses annually, growing royalty base by 20%+ CAGR.
Unlike one-time license deals, per-unit royalties scale with custom-silicon volume. If Meta ships 100M MTIA chips over 5 years, ARM collects per-unit royalties on every chip. This is an annuity business growing with AI infrastructure.
Every Snapdragon, Exynos, MediaTek chip is ARM. Mobile AI adoption (on-device inference) is a structural tailwind ARM captures automatically.
Hyperscalers have invested billions in ARM design expertise and ecosystem (LLVM, compilers, tools). Migrating to RISC-V or x86 would require retraining thousands of engineers. ARM benefits from installed base lock-in.
RISC-V is free (no royalties). If a hyperscaler commits to RISC-V for edge-AI inference (where RISC-V is credible), ARM loses that royalty stream. One large defection (e.g., Amazon moving Trainium to RISC-V) would be a material shock.
Apple, Qualcomm, and Amazon have teams designing ARM alternatives. If any of these move a significant design volume off ARM, it directly reduces royalty growth.
EU regulators questioned whether ARM is charging excessive design-licensing fees. If forced to reduce per-chip royalties or design-license costs, ARM margin pressure is real.
ARMs Cambridge headquarters are in the UK, but the companys software ecosystem and customer base are globally distributed. A Taiwan strait escalation could trigger licensee diversification away from ARM toward RISC-V or in-house ISAs.
ARM is the purest reverse-NVDA trade in the Sequoia thesis context. NVDA loses inference share to custom silicon and hyperscaler vertical integration. ARM gains royalty share because every hyperscaler ASIC that ships is an ARM-licensed design. The question is whether ARMs royalty moat is defensible against RISC-V and vertical-integration risk. Answer: for the next 3–5 years, probably yes—switching costs are high, ARM ecosystem is mature, and the first 10 hyperscaler custom-silicon designs are ARM. But beyond 2028–2030, if RISC-V adoption accelerates or a major licensee defects to open ISAs, ARMs growth story deflates. Verdict: Highly Positive for the next 24 months, with a structural headwind emerging by 2028–2030.
Structural demand from edge AI and custom silicon; royalty model provides high-margin, low-COGS growth.