Business Overview

Arm Holdings plc (Nasdaq: ARM)

5 September 2026

Evidence base: Arm’s Form 20-F for fiscal 2026 (filed 26 May 2026) and for fiscal 2025 and 2024; quarterly shareholder letters and earnings releases through Q1 FYE27 (quarter ended 30 June 2026); the Q4 FYE26 investor presentation. Industry structure is drawn from independent sources — the UK Competition and Markets Authority’s 2021 report on the proposed Nvidia/Arm transaction, Mercury Research, IDC, IPnest, the Semiconductor Industry Association, CSIS and competitors’ own filings. Arm’s fiscal year ends 31 March.

This is a business and industry overview for long-term investors. It is not a valuation and not a recommendation. No price, multiple or target appears anywhere in this document.

1. Executive Snapshot

Sources as cited in the sections that follow. Royalty per chip is inferred from disclosed royalty revenue and disclosed unit volumes; Arm does not publish a royalty rate.

2. What the Company Does

Every chip needs a processor, and designing one is expensive, slow and mostly undifferentiated. A company building a phone processor, a car’s braking controller or a server CPU does not want to invent an instruction set, write a compiler for it and persuade the world’s software to target it. Arm solves that problem once and rents the answer to everyone. It designs CPU architectures and processor cores, licenses them, and never manufactures anything.

The revenue arrives in two forms, and they behave very differently. A licence gives a customer the right to use Arm’s technology — either an implementation licence for a ready-made Arm core, or an architecture licence permitting the customer to design its own processor compatible with Arm’s instruction set. Licence and other revenue was $2,307m in fiscal 2026, up 25%. Then, for the entire commercial life of every chip built under that licence, Arm collects a royalty: $2,613m in fiscal 2026, up 21% (Form 20-F, FY2026). Total revenue was $4,920m, up 23%.

Tracing one unit

The unit of economics is one chip shipped. Arm’s own description of the mechanics is precise and worth quoting, because it contains the constraint most summaries omit: “Royalties are generally either set as a percentage of the licensee’s average selling price per chip or as a fixed amount per chip. The royalty rates per chip typically reduce over time as the total volume of chips incorporating our products shipped increases; notwithstanding such reductions in royalty rates and fees per chip, license agreements with component manufacturing customers typically include a minimum royalty percentage or fee per chip” (Form 20-F, FY2026).

Follow a single unit. A handset maker licenses an Arm core, pays a fee, and spends two to three years designing a system-on-chip around it. The chip enters production; every unit shipped generates a royalty, reported to Arm and collected in arrears. That royalty is small. Arm does not disclose a rate, but the arithmetic is available for the last year in which it published both numerator and denominator: $2,168m of royalty on 30.6bn chips in fiscal 2025 is approximately 7.1 cents per chip (inferred), against 6.3 cents in fiscal 2024 — a 12% increase in the price of the unit. Because gross margin is 97.5%, essentially all of that seven cents is gross profit. The chip itself may sell for anything from a few cents to several hundred dollars.

That gap is the whole design of the business. Arm’s take is small enough that no licensee has a strong economic reason to remove it, and universal enough that it compounds with the entire industry’s unit growth. Arm captured roughly 1.7% of the total value of chips containing its technology at the time of its 2023 listing (Form F-1). Pricing power is exercised not by raising the rate but by increasing the amount of Arm content in each chip.

Raising the price of the unit

Two mechanisms do that work. The first is architectural: moving customers from Armv8 to Armv9. Arm last quantified the uplift in February 2024, stating that “the royalty rates for Armv9 products are typically at least double the royalty rates for equivalent Armv8 products” (Q3 FYE24 shareholder letter). It has not repeated that quantification since, and stopped disclosing Armv9’s share of royalty revenue after it reached 25% in the December 2024 quarter. The second is scope: Compute Subsystems (CSS) bundle a validated multi-core cluster with interconnect and system IP rather than a single core, so Arm supplies more of the chip and earns a higher rate per unit. Five customers were shipping CSS-based chips as of the December 2025 quarter, including all four leading Android handset makers.

A third line is being added. In March 2026 Arm announced the Arm AGI CPU, its first production silicon product, a data-centre processor co-developed with Meta as lead partner. Arm reports customer demand exceeding $2bn across fiscal 2027 and 2028, and expects fiscal 2028 to be “the first year for meaningful Arm AGI CPU revenue” (Q4 FYE26 investor presentation). Nothing has yet been recognised. This matters to the economics as much as to the strategy: Arm’s own filing warns that production silicon “may have materially different margin profiles,” and a business that buys wafers cannot earn 97.5% gross margin. Section 5 takes up what that does to the blended figure.

3. Industry, Competitive Position and Moat

Semiconductor design IP is a small industry attached to a very large one. IPnest put the design-IP licensing market at $8.5bn in 2024, growing 20%, against global semiconductor sales of $792bn in 2025 (SIA). Arm alone held roughly 64% of design-IP revenue in 2022 (IPnest); the top four vendors took 75% of the 2024 market. The pool is concentrated, and it is concentrated in Arm.

Where the profit sits, and why it is not the licensing model

Comparing operating margins across the value chain settles a common misconception. TSMC earned a 53.6% operating margin in 2025 and Nvidia 60.4% in fiscal 2026; Arm earned 43.0% on a non-GAAP basis and 18.3% on a GAAP basis; Synopsys, which sells IP alongside EDA tools, earned 13.0% GAAP; Ceva, a sub-scale IP licensor with the identical business model to Arm, lost money at a 10.3% negative operating margin; Amkor, in packaging and test, earned 7.0%. The IP licensing model does not confer profitability. Installed base does. Arm and Ceva both license processor IP at near-100% gross margin, and only one of them covers its research budget.

What actually protects Arm

The protection is that an instruction set is a standard rather than a product, and standards are defended by the software written against them, not by patents. Over 22 million developers write for Arm and more than 350 billion Arm-based chips have shipped cumulatively (Form 20-F, FY2026). Switching architecture means recompiling, revalidating and re-certifying an entire software stack, and for regulated end markets such as automotive that revalidation is measured in years.

Independent evidence supports this rather than merely repeating Arm’s claim. Reviewing the proposed Nvidia acquisition in 2021, the UK Competition and Markets Authority found significant barriers to switching CPU IP licensor and concluded that “the constraint posed by current or future alternative suppliers of CPU IP … (such as RISC-V and MIPS) is weak.” It recorded third parties citing “the lack of credible alternatives,” and quoted Arm’s own internal documents describing its ecosystem as “difficult to replicate.” The CMA put Arm at 30–40% of worldwide CPU IP by value, but 70–80% of the market for non-proprietary CPU IP — the distinction that matters, since the remainder is largely Intel and AMD designing for themselves.

Named competition, and what each one actually threatens

• RISC-V. The open instruction set is the most-discussed threat and the most frequently misread. More than 2bn RISC-V chips exist and Omdia projects roughly 25% of the market by 2030. But Nvidia has shipped around 1bn RISC-V cores and Qualcomm more than 650m — inside their Arm-based products, as embedded controllers where no application software runs and the switching cost is therefore near zero. That is where RISC-V wins: microcontrollers, industrial, accelerator control, and sovereignty-driven Chinese designs. CSIS assesses that it “won’t be competitive with x86 and ARM in mainstream consumer electronics applications for at least 10 years.” The near-term damage is to Arm’s embedded royalty base, not to its application-processor franchise.
• Customers designing their own cores. Apple, Qualcomm, Nvidia and every major hyperscaler design custom CPUs. This looks existential and is not, because they do it under architecture licences and still pay Arm a per-unit royalty on substantially all chips shipped. Vertical integration by a customer converts a volume risk into a mix risk: Arm keeps the unit but captures less content on it. The threat is to the seven cents, not to the chip.
• x86. Intel held 66.8% of server CPU units in Q1 2026 and AMD 34.5% in Q2 2026 (Mercury Research). AMD is taking share from Intel considerably faster than Arm is taking share from x86 in general-purpose servers.
• Arm itself. The AGI CPU competes with the products its licensees sell. Arm’s own risk factors concede the point: new products “may be competitive with, or may be perceived as competitive with, products sold by our customers, which could cause existing customers to significantly reduce or terminate their relationships with us and could result in existing and new customers electing to use alternative architectures such as x86 and RISC-V” (Form 20-F, FY2026).

The contested number

Arm’s data-centre position is where independent and company sources diverge most sharply, and both are correct. Mercury Research put Arm-based processors at 13.6% of server CPU shipments in Q2 2026, up 0.5 points year on year, and attributed the gain primarily to Nvidia’s Grace CPU shipping inside Blackwell NVL72 AI racks. Arm states it holds roughly 50% share of CPU compute among the top hyperscalers. The denominators differ — all server CPU units versus internal compute at a handful of buyers — and the reconciliation is instructive: Arm is close to universal where hyperscalers design their own silicon, and marginal in the general-purpose server market that x86 still owns. Recent share gains have come from AI rack host processors rather than from displacing x86 in conventional servers.

The CMA report also cited Arm’s internal projection of above 25% data-centre share by 2028 and above 70% by 2030. Those were Arm’s forecasts, not CMA findings, and at 13.6% in mid-2026 they have materially undershot. That is a useful calibration on how the company’s own long-range targets have travelled.

On the evidence, the businesses that sustain pricing power in this industry own a standard rather than a product, monetise through a small royalty attached to a very large installed base, and remain indifferent to which downstream customer wins. Arm is the clearest example of that type in the industry. The question the rest of this memo takes up is whether the company is choosing to remain that kind of business.

4. Growth Engine

Arm reported 23% revenue growth in fiscal 2026. Decomposing it changes the picture materially, and the decomposition is disclosed rather than estimated: Note 4 of the Form 20-F splits every revenue line between external customers and related parties.

Source: Arm Holdings plc Form 20-F for fiscal 2026, Note 4 “Disaggregation of Revenue” and Note 20 “Related Party Transactions.” Related parties comprise Arm China ($790.6m in FY2026), an unnamed affiliate of SoftBank Group ($704.4m) and Ampere ($3.6m). Percentages calculated from the disclosed figures.

Three things follow. First, the royalty engine is real and strong: royalty revenue from external customers grew 20.4%, driven by rate and content rather than volume. Arm’s own illustration from the prior year is the cleanest statement of the mechanism — smartphone royalty revenue rose approximately 30% on less than 2% unit growth. Arm is not being paid more times; it is being paid more each time.

Second, the third-party licensing book shrank. Arm’s management discussion records “a $123 million, or 9%, decrease in license and other revenue” from external customers. Point-in-time licence recognition fell for a second consecutive year, from $1,372m to $1,227m. Headline licence revenue nonetheless rose 25$591 million, or 141%,” to $1,009m — an increase larger than the entire growth in licence revenue.

Third, almost all of that came from one arrangement. Revenue from an unnamed SoftBank Group affiliate went from $145.5m to $704.4m under a consulting agreement covering, in Arm’s words, “certain technical consultancy and advisory services relating to potential transactions, strategic partnerships, licensing agreements, commercial arrangements or other arrangements involving SoftBank Group or its affiliates.” Related parties supplied 30.5% of fiscal 2026 revenue against 20.5% the year before. Stripping that single line out, revenue grew approximately 9% rather than 23% (inferred from the disclosed figures).

Acquisitions are not the complicating factor they often are. Arm bought DreamBig Semiconductor for $265m in cash, completing on 1 July 2026 — after the fiscal 2026 year end and immaterial to reported revenue. Ampere Computing, frequently associated with Arm, was acquired by SoftBank Group for $6.5bn and does not sit on Arm’s balance sheet. Reported growth and acquired growth are therefore the same thing in fiscal 2026; the distinction that matters at Arm is external versus related-party, not organic versus acquired.

The drivers, ranked

• Royalty rate per chip, through Armv9 and CSS adoption — structural. The single most important driver, and the one that compounds without needing unit growth. Each architectural generation roughly doubled the rate when Arm last quantified it, and CSS raises Arm’s content share further. It works in a flat unit market, which is exactly the market Arm faces.
• Data-centre penetration — structural, currently amplified by a cycle. Data-centre royalty revenue more than doubled year on year in both fiscal 2026 and Q1 FYE27, and cumulative Neoverse cores passed 1.5bn with the most recent 500m shipping in nine months. The share gain is durable; its current rate of growth depends on an AI capex boom.
• The SoftBank consulting arrangement — temporary. It contributed $559m of incremental revenue in fiscal 2026, one statement of work has been restructured into a fixed $300m payment falling in fiscal 2027, and it is not a repeatable commercial franchise.
• Licensing breadth — management-driven, and currently decelerating. Annualised contract value grew 13% year on year in Q1 FYE27, down from 28% two quarters earlier. ACV is the forward-looking measure of the contracted base, and it is now growing at roughly half the rate of recognised licence revenue.
• Edge and automotive share gains — structural but small. Share of automotive and robotics rose from 36% to 44% between fiscal 2022 and 2025, and edge from 12% to 15%, on a combined base of roughly a quarter of revenue.
• Smartphone units — cyclical, and a headwind. IDC forecasts global shipments to decline 0.9% in 2026, revised down from 1.2% growth on memory-cost inflation. Arm’s largest end market contributes no volume growth at all.

5. Margin, Cash and Capital Allocation

Arm’s gross margin was 97.5% in fiscal 2026. There is no cost of goods to speak of: the product is a design file, and shipping it to the ten-millionth chip costs nothing. This has an important consequence — gross margin tells you nothing useful about Arm, because it cannot move. Every question about profitability is a question about operating expense.

Sources: Form 20-F for fiscal 2026 and 2025; Q1 FYE27 shareholder letter. Non-GAAP operating margin excludes share-based compensation and related employer taxes; Arm does not report adjusted EBITDA. Q1 FYE27 external-customer revenue is not disclosed, as the split is given annually only. Fiscal 2024 GAAP operating margin of 3.4% reflects $1,037m of share-based compensation recognised on listing and is not comparable with later years as an operating result.

The gap between the two margin lines is share-based compensation: $1,052m in fiscal 2026, equal to 21.4% of revenue. It is a real cost and it is not falling. Diluted share count rose from 1,068m to 1,078m between the fiscal year end and June 2026, and Arm has never repurchased a share, so the dilution is not offset.

The GAAP margin trajectory is the thing to watch, and it is deteriorating for a specific reason. In fiscal 2026 research and development expense rose 34% to $2,776m while revenue rose 23%, so GAAP operating income grew only 8% to $900m. In Q1 FYE27 total operating expense grew 28% against 22% revenue growth, and GAAP operating margin fell to 7.1% from 10.8%. Arm is deliberately spending ahead of revenue to build a silicon business, and because it has no variable cost to flex, that spending falls straight through to operating profit.

Cash generation improved sharply. Operating cash flow was $1,524m in fiscal 2026 against capital expenditure of $545m, giving non-GAAP free cash flow of $882m compared with $99m the prior year; trailing-twelve-month free cash flow reached $1,397m by June 2026. One caution on quality: $645.8m of contract assets were outstanding against the SoftBank affiliate at 31 March 2026, meaning roughly 92% of that $704m of revenue had not been invoiced at the year end. Arm China owed $276.2m, against which Arm carried a $28.3m expected-credit-loss allowance and charged $12.3m to earnings in fiscal 2026.

Where the cash goes

Ranked by fiscal 2026 outflow: research and development $2,776m; capital expenditure $545m; acquisitions $265m (DreamBig, closing after year end); dividends nil; buybacks nil. Arm holds $3,888m of cash and short-term investments as at 30 June 2026 and carries no borrowings — its only fixed obligations are $62m of finance leases and $549m of operating leases.

The ranking says something unambiguous about how management thinks. Every dollar is being reinvested, overwhelmingly into engineering headcount — 9,584 employees at the fiscal year end, 84% of them engineers, up 15% in a year — and none is being returned. Management has set fiscal 2031 targets of $10bn of revenue from the IP and CSS business and $15bn from the silicon business, with non-GAAP earnings per share above $9 and royalty revenue compounding at 20% a year from fiscal 2026 (Q4 FYE26 investor presentation). Those targets appear in an investor presentation and not in the Form 20-F or the shareholder letters.

The structural consequence deserves stating plainly. If that plan is delivered, roughly 60% of Arm’s revenue in fiscal 2031 comes from selling chips, a business management targets at above 30% operating margin, against above 65% for the IP business. A company whose defining financial characteristic is a 97.5% gross margin is choosing to make most of its future revenue in a form that cannot carry one. That may well be the right decision — the addressable revenue is far larger — but it is a deliberate exchange of margin quality for scale, and it is the most important thing about Arm today.

6. Cyclicality, Constraints and What to Monitor

Arm has no factories, no inventory and no capacity constraint, so it cannot suffer the classic semiconductor downturn of unsold stock and idle fabs. Its cyclicality is second-hand: it arrives through its licensees’ unit shipments and through their willingness to start new design programmes. Royalty revenue tracks the former with a lag of a quarter; licence revenue tracks the latter with a lag of years, because a licence signed today produces royalties on chips shipping in 2029.

End-market revenue mix is the most recent full disclosure, for fiscal 2024, from the Q4 FYE25 investor presentation; Arm has not published an updated revenue split by end market since. Share figures are from the Q4 FYE26 investor presentation on Arm’s current segment definitions, which were redefined in fiscal 2026 and are not comparable with the earlier categories. Conditions are sourced as noted in the text.

Where the business sits in its cycle

Three readings, and they point in different directions. Royalty revenue is at a record and rose 21% in fiscal 2026. Non-GAAP operating margin, however, is past its own peak — 46.7% in fiscal 2025, 43.0% in fiscal 2026, 41.2% in the June 2026 quarter — because opex is growing faster than revenue by management’s choice. And the end markets are at opposite extremes: smartphones, at roughly 47% of revenue including other mobile, face a forecast unit decline in 2026, while data centre is being pulled by hyperscaler capital expenditure of approximately $725bn in 2026 against $410bn in 2025, a 77% increase. The embedded market has completed its inventory correction and turned tight, with STMicroelectronics reporting a book-to-bill near 2 and extending microcontroller lead times in its June 2026 quarter.

So Arm is earning record royalties, at a declining margin, with its largest market flat and its fastest-growing market at an all-time capex peak. A record royalty figure at a capex peak and the same figure at a trough are opposite facts, and this one is a peak reading. The downside case is not that Arm loses share — it is that hyperscaler capex normalises. Data-centre royalties more than doubled in each of the last two reported periods; a flat capex year would not reduce Arm’s installed base, but it would remove the growth that is currently offsetting a shrinking third-party licence book and a flat handset market.

7. Risks, Unknowns and Questions for Deeper Work

Cyclicality is covered above. What follows are the risks that a normal cycle does not describe, ordered by how badly they compound with one another.

• Revenue quality is deteriorating beneath a growing headline. Third-party licence revenue fell 9% in fiscal 2026 while reported licence revenue rose 25%; ACV growth halved to 13%; point-in-time licence recognition declined for a second year; and the gap was filled by a parent-affiliate arrangement of which 92% was unbilled at the year end. Each of these on its own is explicable. Together they mean the disclosed measures of the contracted third-party base are all pointing down while reported revenue accelerates, and an investor forming a view on royalty growth in 2029 is relying on a licence book that shrank in 2026.
• Arm does not control 16% of its revenue. Arm China contributed $790.6m in fiscal 2026, and the relevant risk factor is stated without hedging: “Neither we nor SoftBank Group control the operations of Arm China, which operates independently of us.” Arm transferred its entire equity interest in 2022 and now holds an indirect 4.8% non-voting economic interest. Substantially all PRC-related revenue is earned through a single licence agreement with an entity Arm cannot direct, cannot audit and would struggle to replace, in a jurisdiction actively promoting RISC-V as a sovereign alternative.
• Selling chips against the customers who pay the royalties. The AGI CPU makes Arm a competitor to the licensees that generate its royalty stream, in the one end market where those licensees are investing most heavily. The mechanism of harm is not a lost sale but a lost renewal: a hyperscaler or merchant silicon vendor choosing an architecture for a 2030 product now has a reason to weight RISC-V more heavily in that decision. Arm’s own filing names this outcome, and its long-range plan depends on the licensees whose business it is entering.
• The Qualcomm judgment narrows Arm’s ability to reprice. On 30 September 2025 the court ruled that Nuvia did not breach its architecture licence and that Qualcomm’s acquired custom cores were licensed under Qualcomm’s own agreement. Arm has appealed to the Third Circuit and the appeal is pending; Qualcomm’s counterclaim was consolidated in March 2026 with trial expected in the fourth calendar quarter of 2026. Qualcomm is separately disclosed at 9% of fiscal 2026 revenue. The precedent means a licensee can acquire a CPU design team and carry the resulting cores under its existing terms — removing an acquisition as a moment at which Arm can renegotiate rates.
• 72% of the company is pledged collateral. SoftBank Group beneficially owned 86.4% of Arm as at 21 May 2026, and 769,029,000 shares — a 72.0% equity interest — are pledged as security under the SoftBank Group Facility, with prepayment triggers on a change of control or if the ADS trading price falls below certain thresholds. This is a financing structure at the parent that references Arm’s own share price, and it sits above a minority free float.
• The disclosure set has been narrowing while the questions have been widening. In eighteen months Arm has stopped publishing annual chip shipments, quarterly headcount, remaining performance obligations, and the number of Arm Total Access and Flexible Access licensees; it last quantified the Armv9 royalty premium in February 2024 and last gave Armv9’s share of royalties in the December 2024 quarter. Management’s stated reason is that these metrics “become less relevant to our growth with the extension of our business into production silicon.” The effect is that the denominator behind royalty per chip, the forward book, and the architectural mix have all become unobservable at once — and remaining performance obligations were declining 7% year on year when they were withdrawn.

What the sources could not answer

These are the gaps a deeper piece of work would need to close, and each is a genuine absence rather than an item not yet looked for.

• Annual chips shipped in fiscal 2026. Not disclosed, so royalty per chip — the price of Arm’s unit — cannot be calculated for the most recent year.
• The average royalty rate, or royalty as a percentage of chip average selling price. Never disclosed in any filing.
• Armv9’s current share of royalty revenue, and the CSS royalty premium. Last given as 25% in the December 2024 quarter; the CSS multiple appears in no filing.
• Royalty revenue in dollars by end market. Arm publishes growth rates and percentage share only, so the true revenue exposure to smartphones versus data centre cannot be measured.
• The identity of the SoftBank affiliate counterparty, the duration of the consulting agreement, and whether the fiscal 2026 revenue recurs beyond the fixed $300m falling in fiscal 2027.
• The commercial terms behind the “$2bn of customer demand” for the AGI CPU — whether these are binding orders, letters of intent, or design wins — and the product’s gross margin, on which the filings are silent.
• The outcome of Arm’s Third Circuit appeal and Qualcomm’s counterclaim trial, both unresolved.
• Whether the fiscal 2031 targets in the investor presentation carry the operating margins to the segments as this memo reads them; one rendering of the slide reverses the two figures, and the mapping used here is the economically coherent one rather than an independently confirmed one.

8. Investor Takeaways

• What this business really is: a toll on the world’s chip production, collected at roughly seven cents a unit across more than thirty billion units a year, protected not by patents but by the software written against its instruction set.
• The core economic engine: raising the price of the unit rather than the number of units. Royalty revenue grew 20% from external customers in a year when its largest end market shipped no additional volume.
• The main growth lever: data centre, where Arm is close to universal in hyperscalers’ own silicon and marginal in general-purpose servers — and where its growth is presently levered to a 77% increase in hyperscaler capital expenditure.
• What could break the story: not RISC-V displacing Arm in phones, but the combination of a shrinking third-party licence book, a headline flattered by parent-affiliate revenue, and a deliberate move into 30%-margin silicon that puts Arm in competition with the licensees who pay its royalties.
• What to monitor: the external-versus-related-party revenue split in each Form 20-F, ACV growth in each shareholder letter, and operating expense growth against revenue growth — the one lever that sets margin in a business with no cost of goods.