Arm Holdings plc (NASDAQ: ARM) — A Flawless Toll-Road at a Fictional, Float-Driven Price

An independent equity research note Report date: 2026-06-10 · Price reference: $324.86 (2026-06-09 close) · Fiscal year: ends March 31

⚡ Author’s Take

This block is the author’s own subjective opinion and general information only — not investment advice. The analysis that follows takes no position and contains no price target; it discusses valuation only as embedded expectations and scenarios.

Verdict: AVOID at $325 — a genuinely great business at a perfection-plus price; HOLD/accumulate-only-on-deep-weakness. Not a short (the ~13% float is a squeeze trap). Tag: “Flawless company, fictional price.”

Arm is one of the highest-quality franchises in technology — a 92.5%-gross-margin royalty toll-road sitting at the root of a 350-billion-chip installed base, >99% mobile share held for a decade (Greenwald’s share-stability moat test passed outright), a v9/CSS pricing escalator that lifts royalty-per-chip every cycle, and a real, accelerating data-center beachhead. None of that is in dispute. What I am unwilling to underwrite is the price: ~$347B market cap = ~70x sales, ~384x trailing GAAP earnings, ~152x a non-GAAP forward number that adds back >$1B of stock comp. The reverse-DCF is brutal — to justify today’s price at a mature ~35x P/E, Arm must earn ~$10B of net income (~11x FY26’s $904M), implying roughly 4x the revenue, ~3x the net margin, and a decade of sustained ~20%+ growth, all at once. Management’s own ~$25B-by-FY31 ambition, if fully delivered, merely justifies the current price — it is the base case the market has already banked, not upside. Meanwhile operating margin fell YoY (R&D +38% outran revenue +23%), SBC-adjusted free cash flow has been ~nil for three straight years, ~30% of revenue is now related-party (the SoftBank “consulting” line quintupled to $704M in one year and is mostly uncollected), insiders have bought exactly zero shares in the open market, and the sell-side’s own mean target ($234) sits ~28% below spot.

This is, in my read, less a fundamental story than a market-structure story: SoftBank owns ~86%, the float is ~13%, and the marginal price-setter is passive/momentum flow into a thin, crowded ($13% short) security — a scarcity premium that can persist far longer than valuation logic allows and unwind violently (SoftBank supply, lock-up/index dynamics, a growth wobble, beta ~3.4). Framing: quality-compounder-at-the-wrong-price / late-cycle momentum — not contrarian value. My directional zone: fundamental base-case equity value lands around ~$160–180B (~$150–170/share) on excellent execution, with a defensible accumulation zone only below ~$150 (~25x a plausibly-normalized earnings power). Conviction: medium. The single fact that would flip me bullish: durable evidence of operating leverage — R&D growth decelerating below revenue while data-center royalty keeps doubling, i.e., margin actually inflecting toward 30%+. The single fact that would flip me bearish-with-conviction (if it were shortable): a credible SoftBank monetization/float-normalization event landing into any royalty-growth deceleration.

1. Executive Summary

Arm Holdings plc is the world’s dominant supplier of CPU architecture intellectual property — it designs and licenses the instruction-set architecture (ISA), CPU cores, GPUs, system IP, and pre-integrated Compute Subsystems (CSS) on which the vast majority of the world’s computing devices run. It does not manufacture chips; it collects an upfront license fee and a perpetual per-chip royalty. In FY2026 (year ended March 31, 2026) Arm earned $4,920M of revenue (+22.8% YoY) at a 92.5% gross margin, split royalty $2,613M (53%) / license & other $2,307M (47%). The royalty line is a long-tail annuity: ~$2.6B of FY26 revenue was recognized on chips designed in prior periods, and more than 350 billion cumulative Arm-based chips have shipped.

The business quality is not the question — the price is. Arm’s moat is genuine and, in Greenwald’s taxonomy, the strongest combination available: demand-side captivity (ecosystem/software lock-in across 22M+ developers) reinforced by economies of scale (industry-wide R&D amortization no single rival can match). Mobile applications-processor share has held >99% for over a decade — a textbook pass of the market-share-stability test — and the v9 architecture (~2x the v8 royalty rate) plus annual CSS refreshes give Arm a demonstrated, recurring pricing escalator. Data-center is the growth frontier: management claims ~50% of new compute at top hyperscalers and ~100% of DPUs/SmartNICs, with data-center royalty “more than doubling year-on-year.”

But the financial conversion and the governance are weaker than the franchise. Operating margin compressed to 18.5% in FY26 (from 20.6%) because R&D — at $2,776M, 56% of revenue — grew +38%, faster than revenue. ROE is only ~11% and ROIC ~14%, modest for an asset-light IP business, dragged by a post-IPO equity base and the R&D-consuming model. Most importantly, stock-based compensation of $1,052M (21% of revenue) roughly equals reported free cash flow ($979M) — so on an SBC-burdened basis, Arm has produced essentially no owner free cash flow in any of the last three years; the headline FCF is dilution-financed. Layer on a controlled-company structure (SoftBank ~86.4%, float ~13%), ~30% related-party revenue (the SoftBank consulting line quintupled to $704M and is largely uncollected; Arm China royalties are self-reported by an entity Arm does not control), a CEO incentive plan tied to a $2T market-cap milestone, and zero insider open-market buying, and the capital-allocation/governance picture is “not yet proven, elevated risk.”

Valuation embeds perfection-plus. At ~70x sales / ~300x EV/EBITDA / ~384x GAAP earnings, Arm is roughly 3x the EV/Revenue of the next-richest large semiconductor peer and trades richer than every comparable despite a middle-of-the-pack growth rate and a below-pack GAAP operating margin. The reverse-DCF requires ~4x revenue, a near-tripling of margin, and a decade of multiple persistence — simultaneously. The sell-side mean target (~$234) sits below spot; the price is best understood as part fundamental, part low-float scarcity/flow premium. The business will very likely still be a dominant, growing franchise in a decade. Whether an investor buying at $325 is compensated for the embedded expectations is a different — and far less favorable — question. No recommendation and no price target appear below; the analysis is framed as embedded expectations and scenarios.

2. Business Overview

What Arm is. Arm Holdings plc architects, develops, and licenses the world’s most widely deployed CPU instruction-set architecture (ISA) and an expanding portfolio of complementary compute IP: CPU core designs, Mali GPU and NPU accelerators, system IP (interconnects, memory controllers), pre-integrated Compute Subsystems (“CSS”), and the development tools/software that surround them [FACT — Arm FY26 20-F, filed 2026-05-26]. Founded in 1990 as an Acorn/Apple/VLSI joint venture, Arm sells designs, not chips — it sits at the root of the semiconductor value chain and lets licensees (chip makers, OEMs, cloud service providers) build their own silicon on Arm’s architecture. As of March 2026, more than 350 billion Arm-based chips had been reported shipped cumulatively, and the platform supports a community of more than 22 million software developers [FACT — 20-F]. A material new development: in March 2026 Arm announced it will, for the first time, sell its own production silicon — the Arm AGI CPU for data centers, with production expected by end-CY2026 and Meta as lead partner/co-developer [FACT — 20-F; Q4-FY26 call, 2026-05-06]. This is a strategic departure (discussed below) that puts Arm in partial competition with its own licensees.

How Arm makes money — two revenue lines. The model has two components, and the split is the single most important structural fact about the business:

[FACT — 20-F, FY26 vs FY25; total reconciles to the $4,920M anchor.]

1. License & other revenue — upfront/access fees for the right to use Arm IP: Technology Licensing Agreements (TLAs, single design), Architecture Licenses (ALAs, the right to build a custom CPU compliant with the Arm ISA — what Apple and Qualcomm use), and the subscription bundles Arm Total Access (ATA) and Arm Flexible Access (AFA). ATA gives larger customers access to a portfolio (including the latest cores) for an annual fee with manufacturing/tape-out rights baked in; AFA serves smaller players with older cores and a tape-out fee on use [FACT — 20-F]. Licensing is lumpy — a single deal can range from “hundreds of thousands to hundreds of millions of dollars” [FACT — 20-F] — and is concentrated in a relatively small number of customers, so quarter-to-quarter timing swings the line (FY26’s +25% was partly “fluctuation in timing and size of multiple high-value license agreements” and backlog conversion) [FACT — 20-F].

2. Royalty revenue — a per-chip fee on substantially every Arm-based chip shipped, set either as a percentage of the chip’s average selling price (ASP) or a fixed fee per unit, and rising as more Arm IP is integrated into a chip [FACT — 20-F]. This is the annuity. Because royalties are owed on every chip ever shipped across a ~30-year installed base of Armv7/v8/v9 designs — and chips ship for many years after a one-time license — the royalty base is a long-tail, compounding stream. FY26 royalty grew +21% on “an improved mix of products with higher royalty rates per chip, such as Armv9 technology” [FACT — 20-F]. The accounting underlines the recurring nature: of FY26 revenue, $2,617M was recognized from performance obligations satisfied in prior periods — i.e., royalties on previously-shipped designs [FACT — 20-F, Note 4].

Why royalty-per-device is rising structurally (the v9/CSS step-up). Arm’s growth thesis hinges less on chip unit growth (mobile units are flat-to-down) than on extracting more dollars per chip:

• Architecture generation: v8 → v9. Each generation adds capability (and, for v9, AI/security instructions). Management’s repeated framing is that v9 carries roughly double the royalty rate of v8 [INTERPRETATION — consistent management commentary across FY24–FY26 calls; treat the “~2x” as management’s figure, not independently audited]. v9 is now the standard in premium/flagship Android — “all the major Android OEMs are now ramping smartphones with chips based on both Armv9 and CSS” [FACT — Q3-FY26 call, 2026-02-04].
• CSS (Compute Subsystems). Pre-integrated, pre-verified configurations of Arm CPU+GPU+interconnect that customers buy as a hardened block, cutting their design cost and time-to-market — and commanding a higher royalty than discrete IP [FACT — 20-F]. Critically, in smartphones CSS is now an annual refresh: “every smartphone cycle, we deliver a brand-new CSS… the royalty rates are generally increased year-on-year… It’s all moving to CSS now. And as a result of that, we get price every year” [FACT — Q3-FY26 call]. That converts a one-time architecture step-up into a recurring price-escalator.
• More Arm content per chip. Mobile royalty alone is ~43% of total royalty revenue [FACT — 20-F]; as customers add Arm GPUs/NPUs/system IP and stack 100+ Arm cores in data-center chips, royalty-per-device climbs independent of unit volume [FACT — 20-F, “More Value Per Chip”].

End markets. Arm reorganized FY26 around three “AI domains”: Edge AI (smartphones, IoT/embedded, consumer electronics — mobile applications-processor share >99% for many years), Cloud AI (data-center server chips, DPUs/SmartNICs, networking), and Physical AI (automotive ADAS/IVI, robotics) [FACT — 20-F]. The standout is Cloud AI: data-center royalty “continues to more than double year-on-year,” and Arm cites ~50% market share with top hyperscalers and “close to 100% market share” in data-center DPUs/SmartNICs [FACT — Q4-FY26 call]. Geographically, ~64% of revenue is ex-US and the PRC is ~18% of total (down from 22% in FY24) [FACT — 20-F].

Recurring vs. non-recurring. Royalties (53% of revenue) are a high-visibility annuity on an installed base that only grows; ~$2.6B of FY26 revenue came from previously-satisfied obligations [FACT — 20-F]. Licensing (47%) is genuinely lumpy and customer-concentrated. The mix is shifting toward the lumpier line in dollar terms only because licensing grew faster off large FY26 deals — but the structural franchise value lives in the royalty stream and its v9/CSS escalator.

Verdict. This is an exceptionally high-quality revenue architecture: a 92.5% gross-margin IP licensor sitting at the root of a 350-billion-chip installed base, collecting a per-unit annuity that rises with each architecture generation and CSS refresh even when unit volumes stall. The business is not capital-intensive and its royalty line is among the most visible, durable revenue streams in technology. The two caveats are (a) the lumpy, concentrated licensing half, which makes any single quarter noisy and flatters/depresses growth on timing alone, and (b) the AGI-silicon pivot, which trades pristine IP economics for a lower-margin, capital-heavier, customer-conflicting hardware business — a deliberate bet on data-center TAM that the market must underwrite. On the IP franchise itself: best-in-class. The model’s quality is not in question; its valuation and the silicon detour are.

3. Industry Dynamics

The structure: IP licensing vs. merchant silicon. Arm occupies a singular position in the semiconductor stack — it is the dominant supplier of CPU architecture IP, a layer upstream of every fabless designer, foundry, and OEM. The economic logic is powerful: rather than each chip company funding its own ground-up ISA and core development (a multi-billion-dollar, decades-long undertaking), the entire industry collectively funds Arm’s R&D ($2,776M in FY26, ~56% of revenue) and rents the output [FACT — 20-F]. Arm captures a thin slice of an enormous downstream value pool — Arm-based chips power virtually all of the world’s ~1.2B+ annual smartphones plus tablets, TVs, IoT, autos, and a rising share of servers. The royalty take is a small fraction of chip ASP (management does not disclose a blended rate; the widely-cited industry estimate is ~1–2% of chip value, and Arm’s own framing is that it captures “pennies” relative to the silicon and system value it enables) [INTERPRETATION — Arm does not publish a blended royalty rate; the ~1.7% figure circulated by analysts is an estimate, OPEN QUESTION whether it holds as CSS/v9 lift the take-rate]. The 92.5% gross margin confirms the structure: this is a royalty toll-road, not a manufacturer.

Profit pools and where capital sits. The merchant-silicon profit pool is large but fragmented and cyclical (Intel, AMD, Qualcomm, MediaTek, the hyperscalers’ in-house teams, Nvidia). Arm sits above the cycle to a degree — it earns on shipments, not on capacity or pricing wars — but it is not immune: royalties accrue when customers ship, so a semiconductor downturn that cuts unit volumes flows directly into Arm’s royalty line [FACT — 20-F, “cyclical by nature”]. The asymmetry that matters: Arm’s costs are R&D (people), incurred years before the royalties they generate, while revenue is a back-end annuity. That timing mismatch is the source of both the operating leverage (when designs ship at scale) and the risk (if a generation under-adopts after the R&D is sunk).

The competitive landscape.

• x86 (Intel/AMD) — the incumbent in PCs and the majority of the installed server base. Arm is taking share here, not defending: management claims ~50% of new compute at top hyperscalers is now Arm, and predicts Arm will be “the largest market share by CPU type” by end-of-decade [FACT — Q4-FY26 call]. x86’s moat (the Windows/server software ecosystem) is real but eroding as cloud workloads (Linux, containerized, recompilable) blunt the legacy-software lock-in that historically protected x86. This is the rare case where the incumbent’s moat is the structural opportunity for the challenger.
• RISC-V — the genuine structural threat. An open, royalty-free ISA. The pull is obvious: zero license cost and no dependence on a single commercial licensor (a real concern after the failed Nvidia deal and the Qualcomm dispute). Realistic assessment: RISC-V is gaining where the software-ecosystem moat is weakest — deeply embedded microcontrollers, IoT, fixed-function accelerators/state-machines, and China (where avoiding a Western-licensed ISA is a geopolitical objective). It is not displacing Arm where the ecosystem matters most — application processors running rich OSes (Android/iOS/Linux), where decades of compiled, validated software create switching costs RISC-V’s immature toolchain cannot yet match [INTERPRETATION; Arm names RISC-V explicitly as a competitor and notes “some of our customers are also major supporters of the RISC-V architecture” — 20-F]. The threat is real and grows from the bottom up; the question is whether it climbs into Arm’s high-value mobile/data-center core (OPEN QUESTION — the key long-term moat test.
• Internal/custom silicon — a tailwind, not a threat. The most important industry trend is hyperscalers and OEMs designing their own chips (Amazon Graviton, Google Axion, Microsoft Cobalt, Nvidia Grace/Vera). Crucially, most of these are built on Arm — so the custom-silicon wave increases Arm’s royalty base rather than disintermediating it. “NVIDIA, Amazon and Google are already using Arm-based CPUs as head nodes,” and DPU/SmartNIC share is “close to 100%” [FACT — Q4-FY26 call]. The risk embedded here is that the same customers have the scale to fund a RISC-V migration if Arm over-prices — a latent check on Arm’s pricing power.

Regulation and geopolitics — three distinct exposures.

1. Arm China — an unusual, uncontrolled structure. Arm accesses the PRC market almost entirely through a single related party, Arm Technology (China) Co. Ltd., via the IPLA, under which Arm China is the exclusive distributor/sublicensor of Arm IP to PRC customers [FACT — 20-F]. Critically: “Neither we nor SoftBank Group control the operations of Arm China, which operates independently of us,” and Arm holds only a ~4.8% indirect economic interest (the controlling stake was transferred to a SoftBank subsidiary in March 2022) [FACT — 20-F]. Arm depends on Arm China self-reporting royalties accurately and paying on time, and has historically “had issues obtaining timely and accurate information from Arm China… received late payments” [FACT — 20-F]. This is a genuine governance/credit anomaly: ~16–18% of revenue flows through an entity Arm does not control and cannot fully audit; the IPLA runs to 2048.
2. US/UK export controls. BIS has repeatedly tightened controls on advanced-compute IP and services to China; January 2026 rule changes and new 25% semiconductor tariffs add policy risk; a future expansion could cut Arm (directly or via Arm China) off from PRC customers [FACT — 20-F].
3. The failed Nvidia acquisition (2020–2022). Nvidia’s ~$40B bid to buy Arm was abandoned in early 2022 after FTC/EU/UK competition opposition — a clear regulatory signal that Arm’s neutral, “Switzerland of semis” position is viewed as structurally important, and that no single chipmaker will be allowed to own it [FACT — public record, deal terminated Feb 2022]. SoftBank’s ~86% ownership and the thin ~13% float are the legacy of that failed exit.

Marathon capital-cycle read. Capital is flooding into compute, but the supply-side dynamics cut in Arm’s favor. The boom is in AI silicon and data-center capex — and Arm earns a royalty on a rising share of it without funding the fabs, so it captures the upside of the capex super-cycle while bearing little of the capacity risk. The one place capital is being deployed against Arm is RISC-V: hyperscalers, China’s state apparatus, and a wave of RISC-V startups are funding an alternative supply of ISA. In Marathon terms, Arm’s high returns (92.5% gross margin) are precisely the signal that attracts capital seeking to erode them — and RISC-V is that capital. The mitigant is the ecosystem barrier (the supply side of software cannot be rebuilt with money alone), but the capital-cycle warning is genuine and is the bear’s strongest structural card.

Verdict — structurally attractive, with a geopolitical asterisk and a long-fuse threat. This is one of the most attractive structures in technology: a neutral, asset-light IP toll-road at the root of a secular, AI-driven expansion in compute, capturing rising value-per-chip while shifting capacity/cyclical risk onto licensees. The industry’s barriers (the global Arm software ecosystem) are formidable and self-reinforcing. The two structural deductions are (1) the Arm China governance/concentration anomaly — ~16–18% of revenue through an uncontrolled, self-reporting related party — and (2) the RISC-V capital cycle. Net: a structurally good — at the franchise level, excellent — industry, but not a fortress; the moat must be actively defended at the high end, and the China structure is a permanent overhang.

4. Competitive Position

Naming the moat in Greenwald’s taxonomy. Arm’s advantage is the strongest combination in the framework: demand-side customer captivity (ecosystem lock-in / switching costs) reinforced by economies of scale (industry-wide R&D amortization). Both mechanisms are present, and they compound.

1. Demand-side captivity / switching costs. The Arm ISA is the substrate on which the world’s software runs — “the vast majority of the world’s software, including the operating systems and applications for smartphones, tablets and PCs, data centers and networking equipment” — and 22 million developers write to it [FACT — 20-F]. To abandon Arm a customer must re-port, recompile, re-validate, and re-tape-out across an entire software and tooling stack, and accept the risk that third-party apps, drivers, and OS layers may not follow. The 20-F states the captivity directly: “>99% share in mobile applications processors for many years, by virtue of all key mobile operating systems depending on Arm processors” [FACT]. This is Greenwald’s switching-cost moat in its strongest form — not habit (frequent, low-stakes), but software/system lock-in (infrequent, high-stakes, network-reinforced).

2. Economies of scale via shared R&D. This is the deeper, more durable advantage. Arm spends ~$2.78B/year (56% of revenue) on R&D [FACT — 20-F], funded by royalties from the entire industry. No single rival — not Intel, not a hyperscaler, not the RISC-V community — can match an R&D budget amortized across 350B+ chips and every major chipmaker simultaneously. A would-be competitor would have to fund comparable CPU/architecture R&D against a fraction of the volume base, i.e., at a structurally worse cost-per-unit. Greenwald’s “share of the relevant market” test is decisive here: Arm’s scale is in CPU architecture IP specifically, where its share is overwhelming, and the fixed cost (R&D) is spread over the largest possible base. This is the self-reinforcing loop — dominant share → lowest cost-per-design → more R&D → wins the next generation → retains dominant share.

Greenwald’s market-share-stability test — the decisive evidence. The framework’s sharpest moat test is share stability over 5–8 years (<2pt change = formidable barriers). Arm’s mobile applications-processor share has held >99% for many years [FACT — 20-F] — essentially perfect stability over more than a decade, through the entire smartphone era, Apple’s vertical integration, MediaTek/Qualcomm’s rise, and repeated RISC-V hype cycles. By Greenwald’s own standard, the moat test passes unambiguously in mobile. The financial confirmation is the 92.5% gross margin and the v9/CSS pricing escalator: sustained extraordinary economics with stable share is the textbook signature of genuine competitive advantage.

Where the moat is being WON, not defended (less stable, more contestable). Data center is the opposite of mobile — Arm is the insurgent, going from negligible to ~50% of new compute at top hyperscalers and ~100% of DPUs/SmartNICs, with data-center royalty “more than doubling year-on-year” [FACT — Q4-FY26 call]. This is high-quality growth but it is not yet a stable, defended position: the server software ecosystem is younger, more recompilable, and the customers (hyperscalers) are exactly the parties with the scale and motive to fund RISC-V if Arm over-reaches. Contestable share won quickly can be lost; the mobile moat is decades-set, the data-center “moat” is a beachhead.

Pressure-testing pricing power. v9/CSS step-ups are direct evidence of pricing power: Arm raises the royalty rate each smartphone cycle via a new CSS, and customers pay because the alternative (in-housing the architecture) is more expensive than the increase [FACT — Q3-FY26 call: “we get price every year with the royalty increase year-on-year”]. But the power is bounded, and the bounds are named:

• Apple holds a perpetual architecture license and designs its own cores — Arm earns a royalty but exerts little design or pricing leverage over its largest, most sophisticated licensee.
• Qualcomm is both a major customer and an antagonist: the Arm/Qualcomm–Nuvia license arbitration (a 2024 jury trial in which Arm largely lost — Qualcomm was found entitled to use the Nuvia-derived custom cores under its existing licenses) demonstrated that even Arm’s contractual leverage over a giant licensee has real limits. A licensor that cannot fully control how its largest customers use its IP has a moat with a ceiling.
• Hyperscalers can fund RISC-V at the margin; their tolerance for royalty increases is the practical cap on Arm’s data-center pricing.

So pricing power is real and demonstrated, but it is the pricing power of an indispensable supplier negotiating with a handful of giant, well-resourced customers — not the unconstrained pricing power of a monopolist over fragmented buyers.

Direct comparison vs. peers (cross-read against peer company filings and market data, June 2026):

• vs. Nvidia. Both “sell the picks.” Nvidia’s moat is CUDA software lock-in + leading-edge GPU design + integrated systems; Arm’s is the broader, thinner CPU-ISA toll across all compute. Nvidia captures far more dollars per system and earns merchant-silicon margins on hardware; Arm captures a thin royalty across vastly more units at higher gross margin but lower absolute revenue. Tellingly, Nvidia builds its CPUs (Grace/Vera) on Arm — Arm is upstream of even Nvidia.
• vs. Qualcomm. QCOM is simultaneously a customer (licenses Arm IP and the ISA), a royalty-payer, and a litigation adversary. QCOM’s own moat (modem IP, handset integration) is downstream of Arm’s. The relationship illustrates the bounded nature of Arm’s power over scaled customers.
• vs. x86 (Intel/AMD). The mirror image: x86’s legacy-software moat is the asset Arm is steadily eroding in the cloud. Arm’s architecture-licensing model out-innovates a duopoly that must fund its own (increasingly capital-destroying, in Intel’s case) silicon.

Durability assessment. The mobile/edge moat is among the most durable in technology — software lock-in plus industry-funded R&D scale, validated by >99% share over a decade. The threats are asymmetric in time: RISC-V is a slow-fuse, bottom-up erosion (real in microcontrollers/China, unproven in high-value app processors and servers); the Apple/Qualcomm cases show the moat has a ceiling against the largest customers; and the AGI-silicon pivot risks self-inflicted erosion — by competing with licensees in data-center silicon, Arm could motivate those same customers to diversify toward RISC-V to reduce dependence on a now-competing supplier (OPEN QUESTION — the central tension of the new strategy).

Verdict — a durable, genuine moat (demand captivity + scale economies), not an eroding standard — but it is being actively contested at both ends. By Greenwald’s own market-share-stability and ROIC tests, Arm’s mobile/edge franchise is a textbook formidable barrier: >99% share for a decade, 92.5% gross margin, and a demonstrated annual pricing escalator. The scale-of-shared-R&D advantage is structurally near-impossible to replicate with money alone, because the software ecosystem cannot be bought. This is a real moat, full stop. But it is not a fortress without weak points: pricing power is bounded by a handful of giant, litigious, in-housing-capable customers (Apple, Qualcomm, the hyperscalers); the data-center position is won, not defended, and therefore less stable; RISC-V is a genuine long-fuse threat funded by exactly the parties most able to sustain it; and the silicon pivot risks converting customers into rivals. The moat is durable, but durability assumes Arm keeps out-innovating and does not over-monetize — and the current valuation appears to price flawless execution against every one of these contests. The competitive advantage is real; the margin for error is thin.

5. Growth History and Forward Opportunities

Arm grew total revenue from $2,679M (FY23) to $4,920M (FY26), a three-year CAGR of ~22.5% [FACT; FY26 20-F, MD&A]. Growth accelerated through the period — +20.7% in FY24, +23.9% in FY25, +22.8% in FY26 — and, critically, the composition improved: the durable, annuity-like royalty line is now growing in line with the lumpier license line, while the underlying chip-shipment base compounds. As of March 31, 2026, more than 350 billion cumulative Arm-based chips had shipped [FACT; FY26 20-F], each contributing a per-unit royalty for years after the one-time license is booked.

Royalty vs. license decomposition (FY26)

(FACT; FY26 20-F MD&A. Within these totals, related-party revenue — predominantly Arm China and SoftBank — rose +82% to $1,499M, 30% of total, up from 21%, a concentration and related-party-quality flag carried forward below.)

Royalty (+21%, $2,613M). Two structural drivers, both confirmed in the filing and on the calls:

• Armv9 + CSS mix-up. Management attributes royalty growth to “an improved mix of products with higher royalty rates per chip, such as Armv9 technology” [FACT; FY26 20-F MD&A]. Arm has guided historically that v9 carries roughly double the royalty rate of v8, and CSS layers another step-up. All major Android OEMs are now ramping chips on both v9 and CSS [FACT; Q3-FY26 call, 2026-02-04]. This is the mechanism by which smartphone royalty grows despite a flat-to-declining unit market — content/ASP uplift, not volume.
• Data-center share gains. Data-center royalty “more than doubled year-over-year” for multiple consecutive quarters [FACT; Q3/Q4-FY26 calls]. Management claims ~50% CPU share among top hyperscalers and “close to 100%” share in data-center networking (DPUs/SmartNICs) [INTERPRETATION — management figure, unaudited; treat as hypothesis]. The named vehicles are real and shipping: AWS Graviton, Google Axion, Microsoft Cobalt, NVIDIA Grace/Vera. Smartphone applications processors were ~43% of royalty revenue in FY26 [FACT; FY26 20-F], so mobile concentration is falling as data center scales.

License (+25%, $2,307M). Growth is real but lumpy by design — management explicitly warns it varies “due to timing and size of high-value deals” [FACT; Q4-FY26 call]. The cleaner underlying signal is annualized contract value (ACV), up 22% YoY [FACT; Q4-FY26 call] and remaining performance obligations (RPO) of $2,071.4M at March 31, 2026, of which ~28% converts to revenue within 12 months and ~21% in months 13–24 [FACT; FY26 20-F]. RPO explicitly excludes future royalty receipts, so it understates the true contracted backlog. The license model has migrated toward subscriptions — ATA and Flexible Access — which convert one-off licensing into recurring portfolio fees. One caution: the SoftBank technology-licensing/design-services arrangement contributed $200M of the $819M Q4 license revenue [FACT; Q4-FY26 call] — a related party inflating the license line.

Forward opportunities. The forward story is overwhelmingly data center and AI. Management’s framing (ASSUMPTION/INTERPRETATION — company guidance, not independently verified): a data-center CPU TAM of >$100B by 2030; >$2B of booked Arm AGI CPU demand across FY27–FY28 (vs. ~$1B at launch); a stated $15B data-center revenue ambition by ~FY31 and a goal to double the IP business to ~$10B. Meta is the lead co-developer of the AGI CPU; SAP, Cloudflare, F5, SK Telecom, OpenAI, and Cerebras are cited design wins [FACT — as claims; Q4-FY26 call]. Secondary vectors: automotive/robotics (“Physical AI”), AI edge/PCs, and the optionality from Arm moving into production silicon/chiplets itself — which raises revenue-per-chip dramatically but introduces gross-margin dilution, working-capital, and channel-conflict risk with its own licensees (OPEN QUESTION; flagged in 20-F risk factors and the Changes section below).

Quality of growth. The growth is organic (the Artisan business was divested to Cadence in Aug 2025, not acquired), high-incremental-margin (near-zero COGS), and increasingly annuity-like on the royalty side — a genuine long-tail recurring base. Against that: license revenue is structurally lumpy and can pull forward; related-party (Arm China/SoftBank) revenue is 30% of the total and growing faster than third-party revenue, which lowers the auditable quality of reported growth; and the headline data-center figures rest on management’s own unaudited share claims.

Verdict: High-quality growth. Revenue compounds in the low-20s% organically, the royalty base is a durable annuity that re-rates upward with each architecture generation (v9 ≈ 2× v8, CSS higher still), and the data-center vector is real and accelerating off a small base. The quality knocks are real but second-order: license lumpiness, a rising related-party mix (30%), and reliance on management’s unaudited share claims. Growth is not the problem with this business — price and profit conversion are.

6. Financial Quality

Margin structure: spectacular gross margins, thin operating margins. Arm is a pure-IP licensor with almost no cost of goods — COGS was just $121M on $4,920M of revenue (2.5%) in FY26, yielding a gross margin of ~92.5% (94.9% in FY25, 92.8% in FY24) [FACT; FY26 20-F income statement]. That places Arm among the highest-gross-margin public companies in existence, well above fabless peers (NVDA ~75%, QCOM ~56%, MRVL ~60% on the cross-read comps).

The problem is what happens below the gross line. R&D was $2,776M — 56% of revenue — in FY26, up from 50% (FY25) and +38% YoY in absolute terms. SG&A added another $1,115M (23%). The result: operating income of just $908M, an 18.5% operating margin [FACT; FY26 20-F].

(FACT; reconciled to FY26 20-F.)

The operating-leverage question is the crux of the thesis — and the answer so far is negative. Despite revenue rising 23% in FY26, operating margin fell from 20.6% to 18.5%, because R&D grew faster (+38%) than revenue. (INTERPRETATION) Arm is choosing to consume its world-class gross margin in reinvestment — funding the v9 roadmap, CSS, and the AGI CPU/data-center push — rather than letting it drop to operating income. That can be a defensible compounding choice, but it means the textbook “asset-light IP company → 40%+ operating margins at scale” outcome has not yet appeared, and the trajectory in FY26 went the wrong way.

FY24 distortion (normalize it). The $99M FY24 operating income (3.1% margin) is not a fundamentals collapse — it reflects the IPO-related share-based compensation recognized at the September 2023 listing [FACT]. SBC was $1,037M in FY24 vs. the gross-profit base of $2,999M; normalizing out the IPO-cliff SBC, FY24 operating profitability was broadly continuous with the ~25% FY23 level. Use FY25/FY26 as the post-IPO run-rate, not FY24.

EBITDA reconciliation. FY26 EBITDA of $1,157M ties to operating income ~$908M + D&A $249M plus net non-operating items (interest income $111M, equity-investment income, and gains); D&A of $249M is rising as Arm capitalizes more (capex jumped to $545M from $219M in FY25) [FACT; FY26 20-F].

Free cash flow — the central quality knock. Reported FCF looks healthy: CFO $1,524M − capex $545M = $979M FCF (20% FCF margin) in FY26. But the single largest add-back inside CFO is share-based compensation of $1,052M [FACT; FY26 20-F cash flow statement]. SBC is a real economic cost — it dilutes owners — yet it is added back to arrive at CFO. Adjusting FCF for SBC:

(FACT; FY26 20-F cash flow statement.) The conclusion is blunt: on a SBC-adjusted basis, Arm has generated essentially no positive owner free cash flow in any of the last three years. Reported FCF is real cash, but it is funded by paying employees in stock rather than cash — i.e., by diluting shareholders. At ~21% of revenue, Arm’s SBC intensity is extreme even for a tech IP company. This is the most important number in the financial profile.

FY25 CFO drop explained. CFO fell to $397M in FY25 despite NI of $792M — a >$700M gap. The drivers [FACT; FY26 20-F cash flow statement]: a $218M deferred-tax outflow and a large working-capital drain — accounts receivable −$331M, contract assets −$412M, accrued compensation −$152M, and “other liabilities” −$381M, the last driven by employer payroll taxes payable on the vesting of IPO-era RSUs. FY26’s bridge confirms this in reverse: a $302M favorable swing “due to lower employment taxes payable on vested shares.” So FY25 CFO was depressed by a non-recurring post-IPO RSU-tax bulge plus contract-asset build — not a deterioration in the business.

Returns on capital — low for an asset-light model. ROE ≈ NI $904M / equity $8,286M ≈ 11% (FY26). That is low for a 92%-gross-margin IP business, and the reason is structural: the IPO and accumulated SBC inflated the equity base to $8.3B (including $1.6B goodwill and $4.4B retained earnings) against modest operating profit. ROIC is similarly modest — NOPAT ≈ $908M × (1 − ~22%) ≈ $710M against invested capital of roughly equity $8.3B + debt $0.4B − cash & STI $3.6B ≈ $5.1B, i.e. ROIC ≈ 14% (ASSUMPTION — depends on cash-netting/NOPAT normalization). Both sit well below NVDA’s (ROE >90%) and QCOM’s high-teens-to-20s% ROIC. For a business with no factories and near-zero COGS, sub-15% ROIC reflects (a) the R&D-consuming operating model and (b) a bloated post-IPO equity base — not a low-quality asset base.

Balance sheet — a fortress. Cash and short-term investments total $3,601M ($2,751M cash + $850M STI) against total debt of only $432M — net cash of ~$3.2B — with equity of $8,286M and no meaningful leverage [FACT; FY26 20-F]. Liquidity is a non-issue. Current deferred revenue (contract liabilities) is $294M, up from $209M, consistent with the growing subscription/ATA mix. The balance-sheet quality note is $1,623M of goodwill (stable, largely legacy) and the Arm China consolidation.

Dilution and accounting quality. Shares outstanding rose from 1,044M (FY24) → 1,063M (FY25) → 1,068M (FY26) — only ~2.4% on the face, which understates the dilution pressure because the economic cost is the $1.05B annual SBC, not the visible share-count drift. Revenue recognition mixes upfront license recognition (lumpy) with over-time royalty accrual (estimated from licensee-reported shipments — a genuine estimation-risk area flagged as a critical accounting judgment). The $131M gain on the Artisan divestiture (Aug 2025) is a one-time item inflating FY26 pre-tax income and CFO quality — normalize it out. The 141% surge in related-party license revenue warrants scrutiny as a potential timing/pull-forward item (OPEN QUESTION).

Verdict: Economics do NOT yet improve with scale — the bull case is unproven. Arm has the gross-margin structure of an exceptional business (92–95%, near-zero COGS, fortress net-cash balance sheet). But that is where the good news ends on the current numbers. Operating margin compressed in FY26 (18.5% vs. 20.6%) as R&D outran revenue; ROE/ROIC are mediocre (~11%/~14%) for an asset-light licensor; and — decisively — after deducting the ~$1.05B of stock-based compensation that funds it, free cash flow has been roughly nil for three straight years. The reported $979M FCF is a dilution-financed figure, not owner earnings. The entire investment case rests on future operating leverage that has not appeared yet; FY26 went the other way. Until it does, this is a great business model converting very little to true owner cash.

7. Capital Allocation

Arm runs a capital-allocation policy that is unusually simple to describe and unusually hard to evaluate: it returns nothing to shareholders, retains every dollar of cash, and reinvests almost all of it in R&D — and, as of FY26, in a capital-intensive pivot into its own silicon. Whether that is brilliant or value-destructive depends on a single unproven bet, and the analysis is clouded by a controlling shareholder whose related-party dealings now move the P&L.

Returns of capital: zero, by design. Arm pays no dividend and operates no buyback: “We intend to retain earnings, if any, for use in our business and do not anticipate paying any cash dividends in the foreseeable future” [FACT; FY26 20-F, Dividend Policy]. As an English plc, Arm also faces a distributable-profits constraint. For a business this profitable on paper, paying nothing is defensible only if reinvestment earns above the cost of capital. The contrast with fabless peers is stark: Qualcomm returned ~100% of FCF in FY25, ran buybacks at ~3x SBC, and shrank its share count ~3–3.5%/year [QCOM 2026-06-10 report]. Arm does the opposite — share count rose from 1,044M (FY24) to 1,068M (FY26) [FACT], net dilution. The bull case is that a 20–25% grower with a $25B FY31 ambition should plow capital back; the bear case is that “retain everything” is the default of a company whose true free cash flow, after the real cost of its stock-comp engine, is roughly nil (FCF $979M ≈ SBC $1,052M).

R&D intensity: the core allocation decision. FY26 R&D was $2,776M, ~56% of revenue [FACT; 20-F], deliberately ramped “to focus on long-term returns and to replicate the strong position that we maintain in smartphones” in automotive, networking, and cloud; government grants/credits offset $149M. R&D at 56% of revenue is extraordinary even for a designer (Qualcomm ~20%+). The evidence for productive R&D: v9 royalty-rate uplift, CSS/Neoverse adoption, ~50% Arm-compute share at hyperscalers. The evidence against: R&D now partly funds a chip business (AGI CPU) at ~30%+ gross margins — i.e., diluting the franchise’s margin structure — with a payoff years out and unproven. This is no longer pure IP-royalty R&D; it is empire-extension R&D, and should be judged at a higher bar.

M&A: a sensible divestiture, a small bolt-on. Artisan divestiture (closed Aug 26, 2025): Arm sold its Artisan physical-IP/foundation-IP business to Cadence, recognizing a $131.0M pre-tax gain [FACT; 20-F]. Clean and on-strategy — shedding a commoditized, low-royalty line to concentrate on the high-value compute platform. DreamBig Semiconductor (announced Oct 2025): Arm agreed to buy DreamBig for ~$265.0M cash for advanced networking/chiplet capabilities, expected to close by Q2 FY27 [FACT; 20-F] — a small, logical tuck-in supporting the chiplet/complete-solution strategy, but itself a vote of confidence in the vertical-integration pivot rather than the asset-light model that made Arm valuable. (Context: Nvidia’s ~$40B attempt to buy Arm collapsed under antitrust opposition in Feb 2022, after which SoftBank IPO’d Arm in Sept 2023 — the origin of the current minority-float structure.)

Capex: the pivot is now in the cash-flow statement. FY26 net cash used in investing rose $290M YoY, driven by a $326M increase in purchases of property and equipment “driven by data center and office expansions and computer hardware purchases” [FACT; 20-F]. Operating lease obligations are $549M with $76M of signed-but-not-commenced leases (data-center capacity). For a company prized precisely because it was asset-light, a rising capex line is a structural change to the investment case, not a footnote.

SBC as the real cost of the comp model. SBC was $1,052M in FY26 (~21% of revenue) — almost exactly equal to reported FCF. Aggregate senior-management remuneration was $115.85M for FY26; equity awards to senior management totaled $91.67M [FACT; 20-F]. Unlike Qualcomm (buybacks ~3x SBC, mopping up dilution), Arm has no offsetting buyback, so the dilution is real and cumulative. On a fully-burdened basis, Arm generated close to zero true free cash flow in FY26.

Compensation structure and incentive alignment. CEO Rene Haas: base salary $1.35M (frozen for FY27); annual bonus 125% of salary, weighted 50% Revenue / 50% Non-GAAP Operating Income; annual PSU target up to 14x base salary, 75% on financial/strategic measures and 25% on relative TSR vs. the S&P 500 IT Sector Index [FACT; 20-F]. Most strikingly, the proposed FY27 policy adds a one-time Value Creation Plan (VCP) award of 425,000 PSUs tied purely to market-cap milestones: $1.0T by Mar-2029, $1.5T by Mar-2030, $2.0T by Mar-2031 [FACT; 20-F]. (INTERPRETATION) The annual plan is reasonable; the VCP is aggressive and denominated in share price, not returns on capital or per-share value — it aligns management with momentum and a soaring multiple rather than disciplined reinvestment economics, exactly the wrong incentive if the data-center bet turns out to be a capital trap. SoftBank approval is a formality.

SoftBank as controlling shareholder — the conflict that dominates everything. SoftBank Group beneficially owns ~86.4% (as of May 21, 2026); public float ~13% [FACT; 20-F]. Under the Shareholder Governance Agreement, SoftBank can elect the entire board, holds pre-emptive (anti-dilution) rights minority holders lack, approves related-party transactions, and holds consent rights down to 5% ownership [FACT; 20-F]. Arm is a NASDAQ “controlled company.” CEO Haas is also CEO of SoftBank Group International and sits on SoftBank’s board. Related-party revenue is now enormous and growing: ~$1,499M in FY26 (~30% of revenue, +82% YoY), from two engines — (1) Arm China (IPLA): ~16% of total revenue, self-reported by an entity Arm does not control (Arm holds only ~4.8% indirect economic interest; IPLA runs to 2048); and (2) the SoftBank Consulting/licensing agreement: revenue exploded to $704.4M in FY26 from $145.5M in FY25 (~5x), with $645.8M sitting in contract assets (recognized but not yet collected), plus a separate $300M fixed payment due in FY27 [FACT; 20-F]. (INTERPRETATION — the most important capital-allocation red flag.) Roughly a third of Arm’s revenue, and the fastest-growing third, comes from its own controlling parent and an uncontrolled affiliate; the SoftBank line — ~14% of total revenue, up ~5x in one year, mostly uncollected — is exactly the kind of related-party item a skeptic must heavily discount. A material and rising share of Arm’s reported economics is not independently verifiable and is subject to the parent’s discretion, with the 13% float bearing the risk and essentially no minority protections.

Insider & SEC-sweep read. The Form 3 corpus (all dated 2026-03-18) is the initial Section 16 ownership slate (SoftBank, Masayoshi Son, Haas, CFO Jason Child, CLO Spencer Collins, Chief Architect Richard Grisenthwaite, and the director bench). Across the entire ~28-filing Form 4 corpus (Mar–Jun 2026), there were ZERO open-market purchases (no code P) [FACT; SEC Form 4 filings, EDGAR]. Activity was exclusively RSU/PSU vesting (A/M), shares withheld for tax (F), and routine sales (S), with CEO/CFO sales under Rule 10b5-1 plans adopted June 11, 2025; the annual vest/grant clustered May 15–19, 2026. (INTERPRETATION) The tape is neutral-to-mildly-cautious, not bullish: not a single conviction open-market purchase at any price, while executives systematically sold into a stock that roughly doubled. Management’s upside is captured through equity grants, not personal cash conviction. SoftBank (86.4%) neither bought nor sold in the open market in this window; its stake is the dominant overhang, and any future selldown to improve float would be a structural supply event for the minority.

Verdict — has management allocated capital intelligently? Mixed, trending cautious. The defensible pieces are real: the Artisan divestiture was clean, DreamBig is logical, the annual comp structure is acceptable, and reinvesting in R&D suits a genuine 20%+ grower with a wide IP moat. But three things prevent a positive verdict. First, retaining 100% of cash while generating ~zero true post-SBC FCF and diluting shareholders is the absence of a return policy masked by growth. Second, the AGI-silicon pivot is now consuming capex and will dilute gross margins, and the new VCP rewards a $2T share-price outcome rather than returns on the capital deployed — momentum-aligned, not value-aligned. Third, and most serious, ~30% of revenue is related-party, the largest piece quintupled in a year and is mostly uncollected, and Arm China royalties are self-reported by an entity Arm cannot audit reliably. Until the data-center reinvestment demonstrably earns its cost of capital and the related-party revenue proves durable and arm’s-length, the rating is not yet proven, with elevated governance risk.

8. Changes and Headwinds — Last Two Years

The last two years reframed Arm from a sleepy, asset-light IP-royalty monopoly into something more ambitious, more capital-intensive, and more contested. Five changes matter; most cut against the simplicity that made the franchise so attractive.

1. The AGI CPU / own-silicon pivot — the biggest strategic change in Arm’s history. In March 2026 Arm announced it will sell its own production data-center silicon — the Arm AGI CPU (136 cores), with Meta as lead partner and co-developer; first production-chip revenue expected in Q4 of FY27 [FACT; 20-F; Q4-FY26 call]. Reported demand has grown to >$2B across FY27–28 (double the launch figure), though Arm is holding its revenue outlook at $1B due to supply-chain capacity (memory, wafers, packaging, test). Stated FY31 ambition: $15B AGI silicon + $10B IP = $25B revenue, >$9 EPS; first-gen silicon gross margin ~30%+ [FACT — as claims; Q4-FY26 call]. (INTERPRETATION) This is a genuine pivot from a ~95%-gross-margin, near-zero-capital licensor to a (partial) chip vendor that must buy wafers/packaging/test, carry inventory, and earn ~30% margins. Three risks: (a) margin dilution as silicon scales; (b) channel conflict — Arm will compete with its own licensees (Nvidia Grace, Amazon Graviton, Microsoft Cobalt, Google Axion); management insists products “run in tandem” and don’t cannibalize, but that is a hypothesis, not evidence; © capital intensity (capex +$326M in FY26). The Marathon capital-cycle lens flags a late-cycle signal: high returns + a soaring multiple drawing capital into a newly capital-intensive line, justified by an AI-capex boom that may mean-revert.

2. Qualcomm/Nuvia arbitration — a moat-ceiling event. Arm sued Qualcomm/Nuvia (Aug 2022) after terminating Nuvia’s Architecture License, arguing Qualcomm’s Nuvia-derived Oryon cores were not properly licensed. The December 2024 jury largely went against Arm: it found the disputed technology was licensed to Qualcomm and that Qualcomm had not breached, deadlocking on whether Nuvia breached. On Sept 30, 2025, the court granted Qualcomm judgment as a matter of law that Nuvia did not breach; Arm has appealed to the Third Circuit (pending), and Qualcomm’s countersuit is set for trial Q4 calendar 2026 [FACT; 20-F]. (INTERPRETATION) Materially negative for the perceived ceiling of Arm’s contractual leverage: its attempt to use ALA-termination to force a re-license at higher rates failed. For a business whose moat rests on the bindingness of its license terms, a public courtroom loss to its most aggressive licensee caps how much pricing power Arm can extract from the highest-value (architecture) tier; the pending countersuit adds a live downside tail.

3. Artisan divestiture (Aug 2025) and reorganization. Sold Artisan foundation-IP to Cadence for a $131M gain (noted in the Capital Allocation section); the FY26 organization was recast around Edge AI, Cloud AI, Physical AI domains [FACT]. Both are coherent — prune the low-value physical-IP tail, reorganize around where compute demand is. Modestly thesis-strengthening — these are the good changes, signaling focus.

4. v9 / CSS ramp and data-center share gains. Arm-based compute now ~50% share with top hyperscalers [Q4-FY26 call]; v9/Neoverse CSS drive royalty-rate uplift; named wins include SAP, Cloudflare, F5, SK Telecom; Nvidia/Amazon/Google use Arm CPUs as head nodes [FACT — as claims]. The strongest pillar of the bull case and genuinely thesis-strengthening — the IP/CSS franchise is winning the cloud at rising royalty rates. Caution: AMD targets holding ~50% of the CPU TAM and Intel remains relevant, so the share gains are contested, not a coronation.

5. RISC-V, China/export controls, and the SoftBank orbit. The 20-F repeatedly flags the free, open-source RISC-V ISA as a structural threat, noting many customers “are also major supporters of RISC-V” — a slow-burn moat-erosion risk the AGI-CPU pivot ironically amplifies by giving licensees another reason to diversify. China: PRC revenue fell to 18% (FY26) from 22% (FY24); Arm China royalties (~16%) are self-reported via the uncontrolled IPLA; January 2026 BIS rule changes and new 25% semiconductor tariffs add policy risk [FACT; 20-F]. SoftBank AI commitments: SoftBank’s Stargate Project (Jan 2025, with OpenAI and Oracle) names Arm a “key technology partner,” but Arm has received “no formal commitments” that its technology will be used [FACT; 20-F] — narrative tailwind for the multiple, unconfirmed economics, conflicted relationship.

Verdict — do these changes strengthen or weaken the thesis? Net weaken, on balance. The IP/CSS data-center share gains and the Artisan/reorg focus are genuine positives that keep the core franchise compounding. But the two largest changes are negatives for a skeptic: the AGI-silicon pivot trades the pristine asset-light, ~95%-margin model for a capital-hungry, ~30%-margin business that competes with its own royalty-paying licensees, justified by an AI-capex cycle that may not persist; and the Qualcomm/Nuvia loss publicly capped Arm’s contractual leverage over its highest-value licensees. Layer on RISC-V’s slow erosion, the conflicted SoftBank orbit, and an insider tape showing zero conviction buying, and the two-year trajectory is one of rising ambition and rising risk — Arm has become a more exciting story and a less clean business.

9. Risk Analysis (Risk Matrix)

Catastrophic and total-loss risk. (INTERPRETATION) The asymmetry is the key insight: business-catastrophe risk is LOW; multiple-catastrophe risk is HIGH. Arm the business is among the most durable franchises in technology — a universal architecture, ~350B chips shipped, a 22M-developer ecosystem, near-100% DPU/SmartNIC share, and a royalty annuity on an installed base that does not vanish in a downturn. There is no debt-driven solvency risk (net cash ~$3.2B), no single-product dependency, and the switching costs are real. Arm is not a candidate for a zero. The stock, however, embeds a different risk. A ~$347B capitalization requiring roughly a quadrupling of revenue, a tripling of margin, and a decade of multiple persistence — on a security whose price is partly set by passive/momentum flow into a ~13% float — can lose a large fraction of its value without the business doing anything wrong. A reversion toward even an optimistic base case (see the scenario analysis below) implies an equity value on the order of half today’s; a growth deceleration coinciding with a SoftBank supply event could compress it further and faster (beta ~3.4). The catastrophic risk here is not insolvency — it is paying ~70x sales for a great business and watching the multiple, not the franchise, do the destroying. (OPEN QUESTION: what fraction of the price is fundamental value versus low-float scarcity premium? The gap between spot and the ~$234 consensus suggests it is not trivial.)

10. Valuation (Embedded Expectations)

10.1 The starting point: a multiple of multiples. At the 2026-06-09 close of $324.86, on ~1.068B shares, Arm carries a market capitalization of ~$347B and — net of ~$3.2B net cash — an enterprise value of ~$344B [FACT]. Against FY2026 results, the trailing multiples occupy a category almost no other profitable large-cap company shares:

Per an own-history valuation index, Arm’s composite valuation percentile sits at 90.9 versus its own ~3-year post-IPO history (P/B 98.8, P/S 98.8, P/E 75.1) — near the most expensive it has ever been on every basis [FACT]. The 200-day moving average (~$150) sits less than half of spot; the capitalization reflects a re-rating concentrated in recent months, not a settled level.

10.2 Cross-sectional comp set — extreme on every axis. Cross-reading peer semiconductor companies on a like-for-like basis (peer multiples from company filings and market data; ARM as above):

(INTERPRETATION) Arm trades at roughly 3x the EV/Revenue of the next-most-expensive name (MRVL ~27x, AVGO ~25x) and ~14x Qualcomm’s EV/Revenue. The richest peers earn their multiple two ways Arm does not: faster growth (AVGO +48%, AMD +38%, TSM +35%, MRVL +28% all out-grow Arm’s +22.8%) and higher operating margins (NVDA ~60%, TSM ~51%). Arm is simultaneously the most expensive, a middle-of-the-pack grower, and a below-pack GAAP-margin name. The “pure-IP model has higher incremental margins” retort is real (see the Financial Quality section) but does not survive the arithmetic gap.

10.3 Reverse-DCF / embedded expectations — what must be true. Rather than launder assumptions into a forbidden number via a forward DCF, the honest exercise is to invert the price — hold a defensible mature multiple fixed and solve for the earnings Arm must reach:

• Earnings at a mature P/E. To justify ~$347B at 35x → ~$9.9B GAAP net income (~11x FY26’s $904M); at 30x → ~$11.6B (~13x); at 40x → ~$8.7B (~9.6x).
• EBITDA at a mature EV/EBITDA. At ~25x on ~$344B EV → ~$13.8B EBITDA, vs FY26’s $1,157M — a ~12x increase.
• Translate to revenue. At FY26’s ~18% net margin even a best-case mature ~45–50% margin (only NVDA achieves this; Arm never has) puts ~$9.9B net income at ~$20–22B revenue — ~4.0–4.5x FY26. At a still-doubled ~35% margin, required revenue rises to ~$28B (~5.7x).
• Implied trajectory. Reaching ~$20B+ from $4.92B implies ~26% revenue CAGR for 6 years, ~22% for 7 years, or ~17% for a full decade.

(INTERPRETATION) The price embeds a triple-compounding bet, each leg of which must hit: (1) ~20–26% revenue CAGR for ~a decade with zero deceleration; (2) a near-tripling of net margin (~18% → ~45%+) even while funding a lower-margin silicon push; and (3) the multiple not de-rating over that decade. Tellingly, the bull’s own scaffolding (management’s “~$15B data-center + ~$10B IP” by FY31) sums to the ~$20–25B revenue the base of the price already requires — so the stock is not pricing the ambition as upside; it is pricing the ambition as the base case, and pricing flawless margin leverage on top.

10.4 The SBC problem in valuation. FY26 SBC (~$1,052M) exceeds GAAP net income and reported FCF; owner FCF net of it is ~nil. Management’s headline is non-GAAP EPS of $1.77 (vs GAAP ~$0.85), and the ~152x “forward P/E” is computed off a non-GAAP consensus — itself flattered by adding back >$1B of a real, recurring dilution cost. On a fully-loaded basis the cash-economic multiple is materially worse than even ~384x trailing GAAP suggests.

10.5 Scenario analysis (rough equity-value ranges, not targets).

(INTERPRETATION) The asymmetry is the headline: in the bull case, an investor at $324.86 earns roughly the current price plus modest upside — the bull case is the justification of the price, not a source of return. The base case (excellent execution, doubled margin) implies equity value roughly half of today’s; the bear case a ~three-quarters draw-down. (ASSUMPTION: exit multiples are generous — a mature IP licensor that has stopped compounding has historically commanded 20–25x, not 35–40x.)

10.6 What the market prices correctly vs. incorrectly. Correctly: a genuine, durable architectural moat (the ISA, ~350B chips, 22M+ developers, ~100% DPU/SmartNIC share); a real secular data-center tailwind (royalty “more than doubled YoY”); the v9/CSS pricing escalator. Too optimistically: not the existence of these tailwinds but their pace, margin conversion, and the absence of any setback simultaneously — ~20%+ CAGR for a decade with no air-pocket; margins tripling toward NVDA-like levels despite 56%-of-revenue R&D and a lower-margin silicon push; the royalty escalator meeting no large-customer pushback (Qualcomm is the live precedent it does); RISC-V not eroding the tail; Arm China/export friction not impairing a revenue node; and SoftBank’s control never translating into supply. Each is individually plausible; the price requires the joint outcome. No price target — but the embedded expectation is best described as priced for perfection-plus.

11. Variant Perception

Consensus belief. Arm is the toll-collector on compute’s migration to AI and energy-efficient architectures — an “AI tax” that compounds as Arm CPUs win data-center share from x86, as v9/CSS lifts royalty rates, and as AI proliferates to the edge. In this view Arm is a capital-light royalty compounder with a widening moat where operating leverage will eventually convert ~20%+ growth into rapidly expanding margins. The structural oddity the consensus must be read against: Wall Street’s mean target (~$234) is roughly 28% below spot; ratings are mixed (3.8/5). So the “consensus” is bifurcated — the narrative is bullish, the aggregated target says the stock has overshot fundamentals. The marginal price-setter is not the fundamental analyst.

Strongest bull case. A real, internally-consistent thesis: (1) the v9/CSS royalty-rate escalator is a structural per-chip price increase already in the mix; (2) data-center share gains with royalty “more than doubling YoY” and ~100% DPU/SmartNIC share; (3) AGI-silicon TAM expansion — the Arm AGI CPU (Meta lead partner) with >$2B demand booked FY27–28 opening a new silicon line on top of IP; (4) operating leverage finally arriving if R&D growth decelerates below revenue, stepping the ~18% margin toward 35–45%. Sum the parts to ~$15B data-center + a doubled ~$10B IP business and you approach the ~$20B+ revenue the price requires. (All four legs are management hypotheses requiring validation; the data-center-royalty doubling and v9 mix shift are best-evidenced.)

Strongest bear case. Not a dispute of the moat — a dispute of the price for it: (1) ~70x sales / ~300x EBITDA with margin leverage that has not arrived (FY26 op margin fell); (2) SBC-funded FCF (owner cash net of dilution ~nil); (3) RISC-V erosion of the long tail and a customer negotiating lever; (4) a pricing-power ceiling on big customers (the Qualcomm precedent); (5) ownership structure — SoftBank ~86–90%, float ~13% (~139M sh), institutions ~95% of float — a thin, crowded float where momentum/flow sets the price, with any SoftBank monetization a structural overhang; (6) quantitative tells — beta ~3.4, short interest ~13.3% of float, consensus target ~28% below spot; (7) China/Arm-China governance and export risk on a related-party node.

The float/ownership-structure variant — the most distinctive read. (INTERPRETATION) The central variant perception here is not fundamental but market-structural. Spot above the mean analyst target — a stock holding despite covering analysts seeing ~28% downside — is most coherently explained by who sets the price. With SoftBank ~86–90% and only ~13% floating (institutions ~95% of that), the marginal buyer is disproportionately passive/index and momentum capital, not valuation-sensitive money. A thin float amplifies moves (beta ~3.4), lets modest demand sustain a price fundamentals don’t support, and creates two-sided tail risk — upward squeeze potential (13% short into a thin float) and a structural overhang the day SoftBank sells or float normalizes to index free-float weighting. The view: a meaningful share of Arm’s ~$347B cap is a scarcity/flow premium on a controlled, low-float security — not a fundamental valuation the earnings support. That premium can persist long and unwind fast; it is not anchored to cash flows.

The 3–5 assumptions that matter most, and what would falsify each.

The bull must win on all five; the bear needs only one or two to break. That asymmetry — combined with a price already above the covering analysts’ average target — is the heart of the variant perception: less a debate about whether Arm is a good business (it is) than about whether a great business inside a controlled, ~13%-float security is worth ~70x sales when the operating leverage is still a forecast, not a fact.

12. Fact vs. Interpretation Table

13. Open Questions

1. Where does the blended royalty take-rate actually sit, and is it durably rising? Arm does not disclose a blended % of ASP; the ~1.7% industry estimate is unverified, and the v9/CSS escalator’s durability against large-customer pushback is unproven.
2. Will operating leverage ever arrive? FY26 R&D (+38%) outran revenue (+23%) and margin fell. Is the 56%-of-revenue R&D a temporary land-grab that will decelerate, or the permanent cost of staying ahead of RISC-V and funding silicon?
3. What are the true unit economics of the AGI CPU? ~30% gross margin is a guess; capex, inventory, packaging/test capacity, and channel-conflict costs (alienating Graviton/Cobalt/Axion customers) are not yet visible.
4. How real and collectible is the SoftBank consulting revenue? $704M (5x YoY), $645.8M in contract assets (uncollected), $300M more due FY27 — is this durable arm’s-length revenue or parent-directed smoothing?
5. What is Arm China actually paying, and can Arm verify it? ~16% of revenue self-reported by an entity Arm does not control, with a documented history of late/inaccurate payment.
6. What is SoftBank’s monetization plan? Any selldown to improve float is a supply event for the 13% minority; the timing and size are unknown and dominate the technical setup.
7. What fraction of the price is fundamental vs. low-float scarcity premium? Unanswerable with precision, but the spot-vs-consensus gap suggests it is material.

14. What Must Be True (Bull and Bear, with Falsification Tests)

For the BULL (today’s ~$347B is justified or low):

1. Revenue compounds ~20–26% for ~a decade to ~$20–25B. Falsified by: two+ consecutive quarters of decelerating royalty growth or a soft multi-quarter guide.
2. Net margin roughly triples (~18% → ~40–45%) as R&D growth decelerates below revenue and silicon scales profitably. Falsified by: R&D continuing to grow ≥ revenue, or blended margin falling as AGI silicon mixes in.
3. The v9/CSS royalty escalator holds; no successful large-customer challenge. Falsified by: a second Qualcomm-style ruling, a major licensee renegotiating rates down, or royalty-per-chip flattening.
4. RISC-V stays bottom-of-market; SoftBank does not flood the float. Falsified by: a high-volume design publicly moving to RISC-V, or a SoftBank secondary/lock-up/index-reweight supply event.

For the BEAR (the multiple de-rates materially):

1. Operating leverage never inflects — margins stay ~18–20% while the market eventually demands cash earnings. Falsified by: op margin stepping durably above ~25% on decelerating R&D.
2. The growth rate normalizes toward mid-teens as mobile saturates and data-center share gains slow against AMD/Intel pushback. Falsified by: sustained 20%+ total revenue growth with royalty re-acceleration.
3. The low-float scarcity premium unwinds (SoftBank supply, index free-float reweighting, short-squeeze exhaustion). Falsified by: SoftBank formally committing to a long lock-up and the multiple holding through a growth wobble.
4. A related-party/governance shock (Arm China non-payment, SoftBank revenue proving non-recurring, export-control hit). Falsified by: related-party revenue collecting in cash and proving durable at arm’s-length terms.

15. Source Appendix

See Appendix B below for the full source list. Primary sources: Arm Holdings plc Form 20-F for FY2024/FY2025/FY2026 (filed 2024-05-29, 2025-05-28, 2026-05-26; CIK 0001973239); the FY26 earnings-call transcript series (Q1-FY24 through Q4-FY26), the 2025-03-17 SoftBank special call, and the 2026-03-24 shareholder/analyst call; the Form 3/4 insider corpus (filed Mar–Jun 2026); public market data (accessed 2026-06-09/10); and peer company filings/market data (NVDA, QCOM, AMD, MRVL, INTC, AVGO, TSM, June 2026) for cross-reads. Management forward statements are treated as hypotheses requiring external validation, not as evidence.

APPENDIX A — Standard Diligence Questionnaire

Arm Holdings plc (NASDAQ: ARM) — as of 2026-06-10

Supplemental to the research memo. Answers are grounded in the FY26 20-F, the FY24–FY26 transcript series, the Form 3/4 corpus, and public market data (accessed 2026-06-09/10). Fact/Interpretation/Assumption labels applied where it matters.

General

What thoughtful questions have other investors asked about this company? The recurring institutional debates: (1) Is the ~70x-sales multiple ever justifiable, and how much of it is a low-float (SoftBank ~86%) scarcity premium rather than fundamental value? (2) When — if ever — does operating leverage arrive, given R&D at 56% of revenue and an FY26 margin that fell? (3) Is the v9/CSS royalty-rate escalator durable, or did the Qualcomm/Nuvia loss expose a ceiling on Arm’s pricing power over its biggest customers? (4) Does the AGI-CPU own-silicon pivot strengthen the TAM or destroy the asset-light, ~95%-margin model and alienate licensees? (5) How real and collectible is the ~30% related-party revenue, especially the SoftBank consulting line and Arm China’s self-reported royalties? (6) What is SoftBank’s monetization plan and what happens to the float?

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? (INTERPRETATION) Neither cleanly — Arm is in a secular up-leg (data-center share gains, v9/CSS) layered on a cyclical end-market (smartphone units flat-to-negative). Royalties track licensee shipments, so a broad semiconductor down-cycle would hit revenue directly. Margins are arguably below normalized because of the deliberate R&D ramp, not above. Driven by external environment or internal actions? Both: external = the AI/data-center capex super-cycle; internal = the v9/CSS pricing escalator and share-of-compute wins, which are management’s own execution. How stable are revenues? Royalty (53%) is a high-visibility annuity (~$2.6B FY26 from prior-period obligations); license (47%) is genuinely lumpy and customer/timing-dependent. ~30% is related-party, which lowers auditable stability. Outlook for products/services? Strong secular demand for energy-efficient compute IP across edge, cloud, and physical AI; the key uncertainty is margin conversion, not demand. How big will the market be? Management claims a >$100B data-center CPU TAM by 2030 and a ~$25B revenue ambition by FY31 (ASSUMPTION — company guidance). Global, with a shrinking but still-material (~18%) China exposure.

Business Quality & Competitive Moat

Is the industry getting more or less competitive? More, at the margins — RISC-V is funded by exactly the large customers/states most able to sustain it, and x86 (AMD/Intel) is defending data-center share. But Arm’s core mobile position has been stable for a decade. How profitable is the business (ROIC, ROE)? ROE ~11%, ROIC ~14% (FY26) — low for a 92%-gross-margin IP licensor, dragged by the R&D-consuming model and a bloated post-IPO equity base. Gross margin ~92.5% is world-class; operating margin only ~18.5%. How profitable is the industry / barriers to entry? The CPU-architecture-IP layer is effectively a one-firm category with formidable barriers (ecosystem lock-in + industry-funded R&D scale). RISC-V is the only credible structural challenger, and it is royalty-free by design. Can the business be easily understood? Yes at the model level (license + royalty); no at the valuation/related-party level (SoftBank consulting, Arm China, AGI silicon economics). Undermined by foreign low-cost labor? No — this is IP/engineering, not labor-cost-sensitive manufacturing. Do brands matter? “Arm” is a de facto industry standard/brand among engineers; the real moat is the software ecosystem, not consumer brand. Nature of competition / switching costs? Competition is architectural (RISC-V) and from large customers in-housing. Switching costs are high (recompile/re-validate/re-tape-out an entire software stack) but bounded against sophisticated giants (Apple’s architecture license; hyperscalers’ RISC-V optionality).

Financial Condition & Balance Sheet

Assets not fully recognized on the balance sheet? The Arm software ecosystem and ISA standard — the true source of value — are not capitalized. Conversely, ~$1.6B goodwill is on the books (largely legacy). Off-balance-sheet liabilities? Operating leases $549M (+$76M signed-but-not-commenced, data-center capacity); ordinary contractual commitments. No debt-driven solvency risk. How conservative is the accounting? Mixed. Royalty revenue is estimated from licensee-reported shipments (a critical accounting judgment); license revenue timing is lumpy; the $704M SoftBank consulting line is mostly in contract assets (recognized, uncollected); the $131M Artisan gain is a one-time item. The non-GAAP framing adds back >$1B SBC. (INTERPRETATION) Reported earnings quality is below the gross-margin headline. How CapEx-hungry is the business? Historically near-zero; now rising — capex jumped to $545M FY26 (+$326M YoY) for data-center buildout tied to the AGI-CPU pivot. A structural change to the asset-light thesis.

Capital Allocation & Management

How much FCF, and how is it used? Reported FCF $979M FY26 — but ~nil after the $1,052M SBC that funds it. Used for: R&D, the AGI-silicon/capex pivot, and the $265M DreamBig acquisition. No dividend, no buyback. Significant acquisitions recently? DreamBig Semiconductor (~$265M cash, announced Oct 2025, networking/chiplets). Divestiture: Artisan to Cadence (Aug 2025, $131M gain). Buying back shares? No — and share count is rising (1,044M→1,068M, FY24–26). Issuing large amounts of stock to insiders? Yes — $1.05B/yr SBC (~21% of revenue); senior-management equity awards $91.67M FY26. Compensation policy / motivations? CEO base $1.35M; bonus on Revenue/Non-GAAP Op Income; PSUs up to 14x base; plus a proposed Value Creation Plan tied to $1T/$1.5T/$2T market-cap milestones — share-price-aligned, not return-on-capital-aligned. (INTERPRETATION) Incentives reward momentum and a soaring multiple over disciplined reinvestment economics. Insider behavior? Zero open-market purchases across the Form 4 corpus; all activity is RSU/PSU vest, tax-withholding, and 10b5-1 sales into a stock that roughly doubled. Neutral-to-cautious signal.

Valuation & Market Data

ADR / MLP / K-1? Arm trades as ADSs (American Depositary Shares) of a UK plc; foreign private issuer filing 20-F/6-K. Not an MLP; no K-1. Dividend policy? None; intends to retain earnings indefinitely. English-plc distributable-profits constraint also applies. How profitable? Spectacular gross margin (92.5%), thin operating margin (~18.5%), modest ROE/ROIC (~11%/~14%). Net income vs cash from operations diverging? FY25 was the notable divergence (NI $792M vs CFO $397M) due to a post-IPO RSU-tax/working-capital drain; FY26 reconverged (NI $904M, CFO $1,524M). The structural divergence is SBC: GAAP NI and CFO both benefit from adding back $1.05B of real dilution cost.

Risks & Downside

What would cause the stock to decline? A growth-rate deceleration; failure of operating leverage to appear; a second Qualcomm-type license loss; RISC-V displacing a high-volume design; an Arm China/SoftBank related-party shock; a SoftBank selldown/float normalization; or simply multiple compression from extreme levels (the consensus target is already ~28% below spot). Risk of catastrophic loss? (INTERPRETATION) Low at the business level — net cash, no single-product dependency, durable franchise. High at the multiple level — a reversion toward an optimistic base case implies ~half today’s equity value; coincident growth-wobble + supply event could be worse, amplified by beta ~3.4. Chance of a total loss? Remote for the enterprise (no solvency risk, durable IP). The realistic downside is a large de-rating, not a zero.

Recent News & Events

Has the business environment changed recently? Yes, materially: the AGI-CPU own-silicon pivot (Mar 2026, Meta lead), the Qualcomm/Nuvia JMOL loss (Sep 2025), the Artisan divestiture (Aug 2025), the DreamBig acquisition (Oct 2025), the FY26 reorg into Edge/Cloud/Physical AI, and SoftBank’s Stargate naming (no formal Arm commitments). (Events sourced from the 20-F and earnings-call transcripts.) Significant acquisitions? DreamBig (~$265M). Accounting policy changes? None material flagged beyond ordinary updates. New markets/facilities/management? New: own-silicon (AGI CPU) and data-center facilities/capex; management stable under CEO Rene Haas.

APPENDIX B — Source Appendix

Arm Holdings plc (NASDAQ: ARM) — sources & evidence trail (as of 2026-06-10)

Primary sources prioritized over secondary. Management forward statements are treated as hypotheses requiring validation, not as evidence. All financial figures reconciled to the 20-F where possible.

Primary — SEC filings (CIK 0001973239; foreign private issuer, files 20-F/6-K)

Primary — Earnings & event transcripts (company IR / public transcript services)

Primary — quantitative data

Secondary / cross-read (peer company filings & market data — comps & industry framing)

Public record / context (not independently re-verified)

• Nvidia–Arm acquisition termination (Feb 2022); Arm IPO via ADS (Sep 14, 2023).
• Arm v. Qualcomm/Nuvia: Dec 2024 jury verdict; Sep 30, 2025 judgment as a matter of law for Qualcomm; Arm appeal pending; Qualcomm countersuit trial Q4 CY2026 (per 20-F legal proceedings).

Frameworks applied

• Competition Demystified (Greenwald & Kahn): moat-type taxonomy, market-share-stability & ROIC tests, barriers-to-entry analysis.
• Capital Returns (Marathon): supply-side capital-cycle lens applied to compute IP and the RISC-V challenge.