Arm Holdings (ARM) — the CPU IP underneath every AI chip, and how the v9 royalty stepup works

The standard $ARM story is that Arm Holdings licenses CPU designs to chip companies and collects a small royalty per chip shipped. Smartphones use Arm. Servers are starting to use Arm. Therefore as AI compute grows, royalty revenue grows. The story is half-right.

The half it gets wrong is the rate. Arm v9 — the current ISA generation — carries a royalty rate roughly 2× the v8 rate on a per-chip basis. Every NVIDIA Grace CPU, every AWS Graviton4, every Apple M4/M5, every Google Axion, every Microsoft Cobalt 100 is an Arm v9 design — not v8. The mix shift toward v9 is a structural ASP expansion that runs independent of unit growth. Then there's a second layer most retail explainers skip entirely: Compute Subsystems (CSS) — pre-validated chiplet IP that moves Arm one rung up the value chain from "we license the instruction set" to "we license the system reference design," with royalty per AI accelerator rising again. This piece walks through what Arm actually does, how the royalty model works, where it sits in the AI compute stack, and why the bull case rests on a step function the bear case doesn't price.

What does Arm Holdings actually do?

Arm doesn't make chips. They license the blueprint for chips.

Two things get licensed. The first is the Arm instruction set architecture (ISA) — the language a CPU speaks. ARMv8 and ARMv9 are the current generations. Once a chip designer licenses the ISA, they can build their own CPU core that runs Arm code (Apple's Firestorm / M-series cores are the canonical example — Apple's own CPU design, Arm's ISA, Apple pays a royalty per chip shipped). The second is the Cortex / Neoverse CPU IP — fully designed CPU cores Arm ships ready to drop into a chip. Cortex is the consumer line (Cortex-A78, A720, X4). Neoverse is the server line (Neoverse N2, V2, V3) — that's what's inside Graviton, Axion, Cobalt, and the Grace CPU NVIDIA pairs with every Blackwell GPU.

Customers pay two things. Upfront license fees when they sign — the right to use the ISA or a specific Cortex/Neoverse design family. Then per-chip royalties on every chip they actually ship — a small percentage of the chip's average selling price, paid to Arm forever. The license fees are lumpy and front-loaded; the royalties are the recurring annuity that grows with the unit base.

The reason the model scales is that Arm doesn't carry the fab cost, doesn't carry the inventory risk, doesn't ship anything physical. They sell IP. Gross margins north of 95% on royalty revenue — closer to a software company's economics than a chip company's.

How they make money: the v9 stepup

Two revenue lines:

License revenue — upfront fees when a customer signs a new IP agreement (a new core family, a new ISA generation, a new architecture-level license). Lumpy, multi-quarter recognition, growing as more hyperscalers + fabless customers sign up for custom-silicon programs.

Royalty revenue — per-chip fees on every Arm-based chip that ships. This is the structural annuity. The royalty rate isn't disclosed publicly per-customer, but Arm has confirmed in earnings calls and 20-F filings that v9 carries roughly 2× the per-chip royalty rate of v8. That means as v9 mix climbs (and it's been climbing every quarter), royalty revenue grows even if unit volumes are flat. The v9 → v10 transition somewhere out a few years is the next analogous stepup.

Compute Subsystems (CSS) is the next leg. Instead of licensing just the CPU core, CSS is a pre-validated chiplet — CPU complex + interconnect + memory subsystem + I/O — that a customer can drop into their own silicon program with substantially less integration work. The royalty per CSS-licensed chip is higher than a bare Cortex/Neoverse license because Arm is doing more of the design. The 2026 ramp of CSS-licensed hyperscaler programs is the visible piece of this rung-up the value chain.

Customer concentration is real and worth naming. Arm's anchor customers are essentially every hyperscaler + every major fabless: NVIDIA, Apple, AWS, Google, Microsoft, Meta, Qualcomm, MediaTek, Samsung. Apple alone is reported to be ~20%+ of Arm's annual royalty revenue depending on the year. The top five customers are well over half. That's the standard concentration risk for any IP licensor — and it's roughly the same shape as TSMC's customer concentration, with the same response: when the customers ARE the hyperscaler set, the concentration is the AI capex cycle, not a single counterparty risk.

Where it sits in AI compute

Arm sits inside the AI Compute Accelerators bubble — but at a different layer than every other name in it.

Every AI accelerator needs a host CPU. The GPU/TPU/ASIC doesn't run by itself — there's a CPU coordinating workload, managing memory, handling I/O. NVIDIA's solution is the Grace CPU — Arm v9 Neoverse V2, paired one-to-one with each Blackwell GPU on the GB200 / GB300 / Rubin platforms. AWS's solution for general compute is Graviton (Arm v9 Neoverse V2 in Graviton4); for AI training, the Trainium accelerator pairs with an Arm-based host. Google's TPU pods use Arm-based hosts via Axion. Microsoft's Maia AI accelerator pairs with Cobalt 100 (Arm Neoverse-based). Meta's MTIA accelerator runs on Arm hosts.

The structural read: every single named AI accelerator program from a hyperscaler ships paired with an Arm CPU. Arm gets a royalty per accelerator unit sold — without doing the AI silicon itself.

Top correlated names with $ARM (252-day pairwise, QA data):

| Ticker | Correlation | What it tells you | | --- | --- | --- | | $BESIY | 0.66 | BE Semiconductor — hybrid bonding tools; same AI-capex demand | | $MRVL | 0.55 | Marvell — custom AI ASIC silicon; pairs with Arm hosts | | $AMD | 0.52 | AMD — competitive but data-center-driven | | $TSM | 0.52 | TSMC — the fab that makes Arm-licensed chips | | $QCOM | 0.50 | Qualcomm — Arm licensee, current v9 dispute |

The correlations cluster around the AI-capex thesis. The cleanest pure-play comp inside the bubble is the fab (TSMC) — Arm is the IP licensor what TSMC is to the fab side: every name in the bubble passes through both companies' books one way or another.

The numbers

| Metric | Value | As of | | --- | --- | --- | | Market cap | $367B | 2026-05-29 | | Shares outstanding | ~1.04B | 2026-05-29 | | Stock price | $353.31 | 2026-05-29 (close) | | YTD return | +213% | 2026-05-31 leaderboard | | Sector / Industry GICS | IT / Semiconductors | 2026-05-31 | | Analyst PT mean | $232 | 2026-05-31 | | Forward P/E | ~98× | per QA semis-forward-pe watchlist | | ETF exposure (QA tracked) | SOXX at 0.53% weight | 2026-05-14 | | Float overhang | SoftBank owns ~88% | post-IPO 2023 |

Two numbers in that table are worth unpacking. First — the stock is ~52% above the analyst consensus PT mean ($353 spot vs $232 PT). That's not a recommendation observation, it's a tape observation: consensus is trailing the rally, and the rally is being driven by retail + index-passive flows on a thesis (every AI chip has an Arm core) that institutional models haven't fully re-priced. Second — the forward PE of ~98× is the richest multiple on QA's semis-forward-pe watchlist. The bull case has to keep delivering each quarter or the multiple compresses fast.

The bull case

• v9 mix shift = structural ASP growth. Every new hyperscaler custom CPU + every NVIDIA AI platform shipping today is v9, not v8. The ~2× royalty rate runs independent of unit volume. Each quarter of v9 mix lift translates into royalty revenue growth without needing the underlying chip market to grow.
• CSS chiplet IP is a second royalty layer. Compute Subsystems licensing moves Arm up the value chain — royalty per CSS-licensed chip is higher than a bare Cortex/Neoverse license. 2026 is the first material year of CSS-licensed hyperscaler programs.
• Universal AI-host pairing. Every named AI accelerator from a hyperscaler (Grace, Trainium, Axion+TPU, Cobalt+Maia, MTIA hosts) ships with an Arm CPU paired one-to-one. Arm gets paid per accelerator unit without doing the AI silicon itself.
• Apple silicon = 100% Arm v9. M-series, A-series (iPhone), S-series (watch), R1, every chip Apple ships is Arm v9 v9.x. The Apple royalty stream is the closest thing to a recession-resistant annuity in the model.
• Hyperscaler custom-silicon proliferation is accelerating, not slowing. AWS Graviton4, Google Axion GA, MSFT Cobalt 100, Meta MTIA — the move from x86 to in-house Arm CPUs at every hyperscaler is structural. Each new program is a new royalty stream that didn't exist five years ago.

The bear case

• Customer concentration is real. Top five customers are well over half of royalty revenue. Apple alone is ~20%+ in most years. The concentration is the AI capex cycle (same as TSMC), which is fine until capex slows.
• Qualcomm v9 license dispute pending. The Oryon CPU in Snapdragon X / Snapdragon 8 Elite is licensed under Arm v9 terms Qualcomm disputes (inherited from Nuvia acquisition). Outcome materially affects PC + handset royalty per-chip economics for one of the top five customers.
• SoftBank overhang. SoftBank still owns ~88% of the float post-IPO. Any secondary offering pressures the multiple. The 2024 partial sell-down was orderly; future ones won't necessarily be.
• Multiple is extreme. Forward PE ~98× is the richest on QA's semis-forward-pe watchlist. Every percentage point of v9 mix has to print or the multiple compresses fast.
• Stock is ~52% above analyst PT mean. Consensus is trailing the rally. Either consensus revises up (bull) or the gap closes from the other direction. The rally has run on retail + index-passive flow, which is fragile to risk-off rotations.
• RISC-V is a real long-tail threat. Open-source ISA, no royalty. Western Digital's storage controllers, NVIDIA's GPU control processors, Google's data-path silicon have all migrated portions of design from Arm to RISC-V. The displacement is slow, but it's the secular bear case if v10 doesn't deliver another step function.

How to access

ARM trades on NASDAQ as $ARM. Clean US listing — no ADR conversion, no foreign listing complexity. US-retail access through any standard broker; for the operational stack QA uses (multi-currency margin, fractional shares, alerts), see /stack/ibkr.

ETF exposure tracked in QA's data: iShares Semiconductor ETF (SOXX) holds ARM at 0.53% weight as of 2026-05-14. That's a smaller weight than what's structurally true on a market-cap basis — the QA ETF coverage is partial; ARM is also held in NASDAQ-100 trackers (QQQ, QQQM) at meaningfully higher weights, in semiconductor sleeves (SOXQ, PSI, SOXX), and in AI thematic ETFs. If you buy $QQQ or $SOXX, you buy a slice of ARM.

For the live bubble correlation and cross-watchlist context: /stocks/arm.

How this compares to the rest of the AI compute stack

NVIDIA's CUDA moat is software lock-in around a specific chip family. HBM is the tightest bottleneck and concentrates rents in three memory companies. Arm is the third layer: it's a protocol-level monopoly on the CPU IP that every AI chip is paired with. Different mechanism (license + royalty), same structural feature (every dollar of AI infra spend touches one of these names somewhere). The three together — NVIDIA (compute), HBM (memory), Arm (host CPU IP) — define the rent-capturing layer of AI infrastructure today.

If you're trying to build a thesis about "who gets paid when the AI capex cycle runs," all three deserve a slot. They aren't substitutes; they're stacked on top of each other in every datacenter rack.

What to watch

• Next earnings: Arm reports on a UK fiscal calendar — Q1 FY26 (calendar Q1 2026) reported May 2026; Q2 FY26 expected early August 2026. v9 mix percentage and royalty revenue growth rate are the two numbers the tape reads first.
• Qualcomm dispute resolution. Settlement, court ruling, or license restructure — any of the three resolves a hanging tail risk on ~10-15% of royalty revenue.
• SoftBank secondary activity. Any 13F-style disclosure of secondary offerings or planned sell-downs would re-price the float overhang.
• CSS revenue disclosure cadence. Arm hasn't broken out CSS-specific revenue yet. The quarter they start disclosing is the quarter the chiplet thesis becomes investable as a discrete line, not a paragraph in management commentary.
• NVIDIA Rubin platform shipping. Rubin continues the Grace pairing. Each Rubin GPU shipped = one Arm v9 Neoverse royalty. The Rubin ramp timing (currently CY2026 H2) is the next visible AI-accelerator royalty step.
• Bubble shift: if the AI Compute Accelerators bubble breaks correlation with $NVDA, the structural read changes — Arm rides the bubble correlation, and the bubble rides NVDA.

Bubble shifts and rule-based alerts on names like $ARM as they cross watchlist thresholds are part of /pro.

---

Live data on this ticker: /stocks/arm — price, ETF holdings, bubble correlation, bot positions.

Bubble context: /bubbles/semiconductors — the AI Compute Accelerators cluster ARM sits in and how it's moving.

Adjacent reading: NVDA's CUDA moat and HBM is the tightest bottleneck in AI — the other two protocol-level rent layers of the AI compute stack.

QuantAbundancia is educational research. Nothing here is investment advice. See /disclosures.