ARM — Arm Holdings

Arm Holdings is the closest thing to an intellectual toll road in the semiconductor industry — its architecture runs inside more than 30 billion chips shipped annually, and every one of those shipments generates a royalty. The moat is real, the growth into data centers is genuine, and the migration to higher-royalty ARMv9 architecture creates a multi-year tailwind that does not depend on unit volume growth. At $162.40 per share, however, the stock trades at approximately 93 times non-GAAP fiscal 2026 earnings for a business growing revenues at 20 to 25 percent annually, and the premium asks investors to pay for perfection across a decade of execution that includes a risky new strategy, a dominant shareholder with a margin loan, and a competitive open-source alternative gaining ground in the segments ARM needs most for future growth. Good business, meaningfully overpriced.

---

The artificial intelligence infrastructure buildout has compressed a decade of semiconductor capital spending into roughly three years. Every data center operator, cloud hyperscaler, and enterprise technology company is simultaneously expanding compute capacity, and the chip architectures that power that expansion have become the most valuable intellectual property in the global economy. In this environment, any company with a legitimate claim to foundational position in AI compute commands extraordinary valuations — and few companies have a more foundational claim than Arm Holdings, whose architecture underlies the processors in virtually every smartphone, the custom inference chips being built by Amazon, Google, and Microsoft, and the GPU-adjacent compute infrastructure being deployed by NVIDIA. The investment question for Arm is not whether the business is good. It plainly is. The question is whether the current price compensates for the risks embedded in the multiple the market has assigned it.

Semiconductor intellectual property licensing is among the most attractive business models in technology. A company designs a processor architecture once, then licenses the right to use that architecture to chip makers who build the actual silicon. The licensor collects an upfront fee when the license is signed and a royalty on every chip that ships — typically one to ten percent of the chip's average selling price, depending on the architecture generation and licensing terms. No factory is required. No inventory is held. No raw materials are consumed. The IP licensor's primary costs are the engineers who design the architecture and maintain the standards that make it useful to the ecosystem. Gross margins at a mature IP licensor routinely exceed 90 percent — and Arm's fiscal 2026 gross margin runs at 97 percent, which is essentially the theoretical ceiling for a software-like business. The semiconductor IP licensing market reached approximately $10.2 billion in 2025 and is growing at double-digit rates driven by the proliferation of custom silicon in AI, automotive, and IoT applications.

Arm Holdings was founded in 1990 and spent 33 years as a private company before its September 2023 IPO at $51 per share — the largest semiconductor listing in the history of US markets. SoftBank, which acquired Arm for $32 billion in 2016, retains approximately 87 percent of the company today. The architecture Arm designs — a reduced instruction set computing approach that trades raw clock speed for power efficiency — has become the dominant computing paradigm for battery-powered devices. Over 90 percent of smartphones globally run on Arm processors. More than 310 billion Arm-based chips have been shipped cumulatively, and the installed base of Arm-powered devices generates an ongoing royalty stream that compounds as the world adds connected devices at a rate of 25 to 30 billion new chips per year.

The company earns revenue through two mechanisms. Licensing fees — paid upfront when chip designers agree to use Arm architecture — contributed $1.84 billion in fiscal 2025, growing at 28.5 percent year over year. Royalty revenue — the per-chip stream collected on every device that ships — contributed $2.17 billion, growing at 20 percent. Both streams are growing, but it is the royalty line that reveals the structural story: as the chip ecosystem migrates from the older ARMv8 architecture to ARMv9, and from standalone CPU cores to Arm's newer Compute Subsystems (CSS) products, the royalty rate per chip increases dramatically. ARMv8 carries a royalty of roughly 2.5 to 3.5 percent of chip average selling price. ARMv9 carries approximately 5 percent — nearly double. CSS carries more than 10 percent — double ARMv9 again. In fiscal 2025, ARMv9 chips represented approximately 31 percent of royalty revenue; by the end of fiscal 2026, that figure has crossed 50 percent. Management's long-range target is 60 to 70 percent ARMv9 royalty mix, with CSS growing in parallel. This is a royalty rate expansion story that does not require Arm to ship a single additional chip — it simply requires that the chips already being designed and shipped move up the architecture ladder.

The moat that protects this royalty stream is genuine and specific. It operates on two levels. The first is the software ecosystem. iOS and Android — the two operating systems that together define the entire mobile computing market — are each built on decades of optimization for Arm architecture. An application compiled for Arm iOS runs on every iPhone ever made. Recompiling it for RISC-V, the open-source alternative architecture that is Arm's most credible long-term challenger, would require not just recompiling the app but rebuilding every layer of the OS, the compiler toolchain, the hardware abstraction layer, and the driver stack. Apple has invested twenty-plus years in this integration. Qualcomm has built a licensing business around it. The switching cost, measured in time and money, is simply prohibitive for any application where performance and compatibility matter. The second level of moat is the chip design investment itself. Tape-out costs — what it costs a chip designer to finalize a new processor design and commit it to silicon — run to tens of millions of dollars and twelve to eighteen months of engineering time. Once a chip designer has taped out an Arm-based processor, they are not redesigning it for a different architecture the following year.

The comparative royalty rate table makes the moat visible in economic terms:

Architecture	Royalty Rate (% of ASP)	Injections / Year (analogy)	Status
ARMv8 (legacy)	2.5–3.5%	Baseline	Declining share of new designs
ARMv9 (current)	~5%	~2x v8	>50% of royalty revenue as of Q3 FY2026
CSS (Compute Subsystems)	>10%	~4x v8	21 licenses, 5 customers shipping
RISC-V (competitor)	0%	0x (free to license)	~25% global market penetration, IoT/embedded

RISC-V is Arm's most significant competitive threat, and it deserves direct treatment. The open-source instruction set has achieved approximately 25 percent global market penetration as of early 2026, and RISC-V startups achieved genuine performance parity with Arm's server-grade Neoverse cores in benchmarks published in late 2025. Qualcomm acquired Ventana Micro Systems for $2.4 billion to gain RISC-V data center expertise. Meta acquired Rivos. Alibaba and SpacemiT have shipped RISC-V server processors. The threat is not imaginary. It is, however, structurally limited to domains where the ecosystem moat is thinnest. In IoT and embedded microcontrollers — devices with minimal software stacks and few compatibility requirements — RISC-V is a credible alternative. In data centers running custom AI inference workloads, RISC-V cores are increasingly viable. In smartphones and PCs — where the software ecosystem lock-in is deepest — RISC-V has zero presence in any production device today, and no plausible path to disruption within the next five years. The mobile moat is intact. The question is whether it is sufficient to justify the current valuation when the high-growth segments (AI data centers) are the ones most exposed to RISC-V competition.

The financial profile is dominated by one fact: 97 percent gross margins. Almost none of Arm's revenue goes to cost of goods sold — the architecture has already been designed, and shipping another license costs essentially nothing. Fiscal 2025 revenue was $4.007 billion, up 20.6 percent from $3.233 billion in fiscal 2024. The most recent quarterly report — Q3 fiscal 2026 (ended December 31, 2025) — showed $1.24 billion in revenue, up 26 percent year over year, and management has guided fiscal Q4 2026 (ending March 31, 2026, not yet reported as of this writing) at $1.47 billion — which would bring full fiscal year 2026 revenue to approximately $4.9 to $5.0 billion. Non-GAAP operating margins in Q3 fiscal 2026 were 41 percent, reflecting the operating leverage available in a model where revenue compounds but headcount does not need to match it.

The GAAP picture diverges substantially from the non-GAAP narrative. Stock-based compensation runs approximately $1.2 billion annually — nearly 30 percent of revenue — which is extraordinarily high for a company that is not in an aggressive expansion phase. The GAAP net income in fiscal 2025 was $792 million on $4.007 billion in revenue; the non-GAAP equivalent (adding back SBC and other items) is approximately $1.64 to $1.70 billion. At $162 per share, the stock trades at 216 times trailing GAAP earnings and approximately 93 times estimated non-GAAP fiscal 2026 earnings. Free cash flow, the most honest measure of what the business actually generates after all costs including the employee equity dilution, was only $178 million in fiscal 2025 — a 4.4 percent FCF margin on $4 billion in revenue that is difficult to explain from the income statement alone. Fiscal 2024 FCF was $998 million; management attributes the collapse to working capital timing, and that explanation may be correct. But a business with 97 percent gross margins generating 4 percent free cash flow margins in a given year deserves scrutiny of what is actually happening with the cash, not simply the reassurance that it is timing.

Q3 fiscal 2026's licensing line included a $200 million license agreement with SoftBank — Arm's 87 percent parent company. Related-party license deals of this size, disclosed without detailed breakdown of what rights were transferred, are not disqualifying, but they are a material reminder that minority shareholders in Arm do not control the governance of the company. Masayoshi Son chairs the board. SoftBank qualifies Arm for "controlled company" status under Nasdaq rules, which reduces the required independence of the board and audit committee. More concretely, SoftBank has drawn $8.5 billion against a $20 billion margin loan facility backed by its Arm stake. The margin call threshold at current drawn levels would be triggered at a stock price of approximately $37 — far below today's price — but the facility is expandable to $20 billion, at which point the threshold rises to approximately $87. In the scenario where SoftBank faces financial pressure and draws the facility toward its limit, minority shareholders in Arm become creditors' collateral, not the primary constituency the board is managing for.

CEO Rene Haas has led Arm since February 2022, bringing a background from NVIDIA's GPU computing division and nearly a decade at Arm before assuming the top role. He has managed the company through its IPO and through a significant pivot in strategy: in March 2026, Arm launched its first in-house chip in its 35-year history — a 136-core data center processor branded the AGI CPU, with Meta as its first customer. This is either the most important strategic decision Arm has made in a generation or a significant error, and it is too early to know which. The opportunity is clear: capturing full chip revenue instead of a 5 to 10 percent royalty is vastly more lucrative per unit. The risk is equally clear: Apple, Qualcomm, and NVIDIA — Arm's three most important licensees — now face Arm as a direct competitor in the data center CPU market. NVIDIA liquidated its remaining equity stake in Arm in February 2026, roughly one month before the chip launch. The management team is executing on a strategy that simultaneously targets the largest new market opportunity and threatens the existing revenue streams that fund the current valuation.

The growth runway over the past four fiscal years shows the structural drivers of the royalty rate expansion:

Fiscal Year	Total Revenue	Royalty Revenue	Chips Shipped (B)	Avg Royalty / Chip	ARMv9 % of Royalties
FY2023	$2.68B	$1.68B	~27B	~$0.062	<5%
FY2024	$3.23B	$1.81B	~29B	~$0.062	~5–10%
FY2025	$4.01B	$2.17B	30.6B	$0.071	31% → >50%
FY2026E	~$4.95B	~$2.70B	~32B	~$0.084	>50% → 60%

The table shows two things simultaneously. First, chip shipment volume is growing modestly — from approximately 27 billion to 32 billion units over four years, a 4 to 5 percent annual unit growth rate that reflects global semiconductor demand but is not the primary driver of revenue expansion. Second, and more important, the average royalty per chip collected has risen from approximately $0.062 in fiscal 2023 to an estimated $0.084 in fiscal 2026 — a 35 percent increase over three years, driven entirely by the migration to higher-royalty ARMv9 architecture. If this migration continues toward the 60 to 70 percent ARMv9 target and CSS adoption adds further rate expansion, the average royalty per chip could plausibly reach $0.10 to $0.12 over the next three to four years without requiring any acceleration in unit shipments. That is the structural growth engine: the ecosystem Arm already has, paying more per chip as those chips improve.

The penetration argument for data centers is the most important current expansion story. Arm processors represented approximately 5 percent of data center CPU shipments in 2020. By 2025, that figure had reached an estimated 15 to 23 percent, driven by AWS Graviton5 (now powering more than 50 percent of new AWS CPU capacity), Google Axion, Microsoft Cobalt 200, and NVIDIA's Grace CPU. Cloud AI data center royalties grew more than 100 percent year over year in the most recent quarters. Arm's stated aspiration is 50 percent of data center CPU shipments — a target that, if achieved, would roughly double the royalty revenue available from the segment that currently pays the highest rates. Against an estimated 1.5 to 2 billion data center CPUs deployed globally and growing, the runway from 15-23 percent penetration to 50 percent is substantial. The complicating factor is that this is precisely the domain where RISC-V is advancing fastest and where the hyperscalers (Meta, Google, Amazon) are investing most aggressively in proprietary open-source designs that bypass Arm's royalty structure entirely.

At $162.40 per share, Arm carries a market capitalization of $158 billion and an enterprise value of approximately $163.7 billion, after subtracting $2.1 billion in cash and adding the negligible debt. Against estimated fiscal 2026 revenue of $4.9 to $5.0 billion, this represents an enterprise value to revenue multiple of approximately 33 times. Against estimated non-GAAP fiscal 2026 earnings of approximately $1.75 per share, the forward non-GAAP P/E is approximately 93 times. For context, the semiconductor sector median forward P/E is approximately 28 times. Arm trades at 3.3 times that sector median for a business growing revenues at 20 to 25 percent annually. That growth rate is excellent for a company of this size and margin structure, but it is not the 50 to 100 percent growth rate that would justify 3.3 times the sector median. The premium implies that investors believe Arm will sustain this growth rate for many years, successfully navigate the fabless chip pivot without material loss of licensing revenue, and achieve a much larger royalty base as data center penetration expands. That is a coherent investment thesis. It is not a certain one.

The most credible bear case is not about the mobile moat — that is real and durable. It is about what happens when the company that designs the architecture your entire product line depends on decides to compete with you in your fastest-growing market. NVIDIA, which paid $40 billion in proposed merger consideration for Arm in a deal blocked by regulators in 2022, sold its remaining equity stake in February 2026, one month before Arm's first chip launch. NVIDIA and Arm now occupy the same competitive space for data center AI processors. Qualcomm, which just won a major legal victory establishing that it can continue using its acquired Nuvia CPU cores without ARM's consent, is simultaneously pursuing a damages countersuit against Arm. Apple's in-house CPU design team, which built the M4 processors that benchmark above every x86 alternative, is deep enough into custom silicon that they license Arm architecture at the ISA level rather than using Arm's standard CPU designs — meaning they could, with enormous cost and years of engineering time, eventually develop an alternative if the relationship with Arm deteriorated. None of this collapses the thesis. But it suggests that the "toll road" metaphor — which implies passive, frictionless cash collection — understates how much is being negotiated and contested at the relationship layer below the royalty stream.

The honest answer to the bear on this stock is that the mobile moat is so structurally entrenched that it would take fifteen years of deliberate effort by every major handset OEM simultaneously to displace Arm from smartphones, and that does not happen. The data center royalty stream is growing fast enough that even if RISC-V captures a third of new AI inference workloads, Arm's data center revenue at 93x EPS is still compounding at meaningful rates. These are correct observations. But correct observations about business quality do not determine whether a price is appropriate, and 93 times non-GAAP earnings is not a price that leaves room for the fabless pivot to stumble, for SoftBank to create a governance incident, or for the RISC-V competitive dynamic to develop faster than the current data suggests. At $162.40, the investor is paying for everything going right across every dimension of the strategy simultaneously.

What would change this verdict: a stock price that compresses to a level where the growth rate adequately compensates for the risks. At approximately 50 times non-GAAP FY2026 earnings — roughly $87 per share — the investor begins to pay a premium for growth without paying for perfection. At that level, the royalty rate expansion story provides meaningful upside without demanding that the fabless business, the data center expansion, and the RISC-V competition all resolve favorably simultaneously. The valuation gap from $162 to $87 is not a near-term prediction; it is simply the distance between the current price and one that reflects the actual uncertainty in a business that is genuinely excellent but not permanently immune to competitive, governance, and strategic execution risk.

The architecture is irreplaceable. The price is not.