Summary

  • Arm's results for the quarter ended 30 September 2026 carry a quarterly chip count for the first time by the company's own account: more than 10 billion Arm-based chips shipped, up 1.5 billion year on year, beside royalty revenue of US$670m, up 8%.
  • At those floors, royalty revenue per shipped chip falls by 6.1% to 8.2% — in the quarter Arm credits royalty growth to higher royalty rates per chip. Its Form 20-F says the rate "generally decreases as the volume of sales increases, subject to an agreed minimum royalty per chip".
  • One number is floored and dated by Arm's own account of its reporting chain to the preceding quarter; the other is a current-quarter accrual, trued up later. The divisor arrived; the rate moved further away.

The quarter that printed a divisor

On 4 November 2026 Arm Holdings plc reported total revenue of US$1,443m, up 27%; licence and other revenue of US$773m, up 50%; royalty revenue of US$670m, up 8%; and annualised contract value of US$1,790m, up 12% (Q2 FYE27 shareholder letter). The same release carries the sentence that changes what a reader can attempt: "For the first time, partners reported more than 10 billion Arm-based chips shipped during the quarter, an increase of 1.5 billion chips year on year" (Form 6-K, November 2026). The chief financial officer repeated the milestone on the call (Q2 2027 earnings call transcript).

A royalty line and a chip count in one document look like two halves of a rate. They are not, and the reason is printed beside them. Arm credits the increase to "higher royalty rates per chip such as Armv9 architecture and Arm CSS, and increased deployment of Arm-based chips in data centers, partly offset by memory pricing in the smartphone market" (Q2 FYE27 shareholder letter). A company naming per-chip rates as its driver, in a quarter that prints a chip count, has invited exactly one division.

The two numbers are different vintages

Royalty revenue is not counted when the money lands. It is estimated when the silicon ships. "Royalties are recognized on an accrual basis in the quarter in which the customer ships their products," Arm writes, from historical trends, customer reports, forecasts and third-party data on unit shipments, average selling price and product mix. Then: "Adjustments to revenue are required in subsequent periods to reflect changes in estimates as new information becomes available, primarily resulting from actual amounts subsequently reported by the licensees in the period following the accrual, including royalty audit resolutions" (Q4 FYE26 shareholder letter).

Which quarter the chip count describes is a separate question, and Arm has answered it before. In 2023 it wrote that most customers furnish royalty reports "setting forth the actual number of Arm-based chips they shipped in the immediately preceding quarter" (FYE24 Q2 shareholder letter). A count reported in one quarter is a report about the quarter before it, while the accrual beside it is about the quarter itself and still open to revision. Nothing in the documents read here reconciles the two descriptions: the November 2026 release says the chips shipped "during the quarter", the 2023 letter says the reports cover "the immediately preceding quarter". Both are Arm's wording, and no bridge sentence was found either way.

A counter that is a stock, not a flow

The number a reader is likelier to have seen is cumulative. "Cumulatively, our customers have now shipped more than 350 billion Arm-based chips, powering everything from the tiniest of sensors to the most powerful supercomputers" (Annual report FYE26). Two quarters earlier the same site printed "+325bn Arm-based chips shipped" and called it the "cumulative number of Arm-based chips since 1990" (Q3 FYE26 investor presentation). In 2023 the counter stood at 272.5 billion, with an explicit scope: "The cumulative number of Arm-based chips reported as shipped are from inception to date" (FYE24 Q2 shareholder letter).

Inception to date is a stock. It cannot divide one quarter's revenue, or one year's: a chip shipped in 1994 earns its royalty in 1994. My arithmetic, offered as context and not as a rate: 350 billion chips across the 36 years from 1990 to 31 March 2026 averages about 9.7 billion a year, a little less than the 10 billion reported in one quarter of FYE27.

The rate Arm credits is a schedule, not a scalar

Arm's Form 20-F contains the sentence that reframes the quarter. "Our royalty revenue per chip generally decreases as the volume of sales increases, subject to an agreed minimum royalty per chip" (Form 20-F, FYE26). That is a documented schedule with a floor, not a price. It also means the direction of the aggregate quotient and the direction of the per-design rate are not one fact: a company can raise rates on newer designs, ship far more chips on older ones, and report a lower average revenue per chip with nothing wrong.

The same filing fixes the population. Royalty revenue from mobile applications processors was approximately 43% of Arm's royalty revenue in the year to 31 March 2026 (Form 20-F, FYE26), and disaggregated royalty revenue was US$2,613m against US$2,168m and US$1,802m in the two prior years, of which US$490m came from related parties (Form 20-F revenue table R37); the full-year record is confirmed in the fourth-quarter letter beside US$671m for that quarter (Q4 FYE26 shareholder letter).

Arm's architecture mix is a publisher's estimate and says so: royalty mix by architecture such as Armv9 "is estimated at the SoC level based on the architecture of the primary CPU or an approximation of the IP mix and is subject to change based on the availability of additional product detail", with figures "based on the most recent royalty report data that relates to the prior quarter" (Q1 FYE27 shareholder letter). Across the ten documents listed below, read as text on 19 November 2026, no Arm-authored page prints a numeric Armv9 share of royalty revenue and none prints an average royalty rate per chip. Third-party summaries of earlier slide decks circulate several figures for that share; they disagree and none is relied on here.

The arithmetic, labelled mine

Volume growth is a band, because both printed figures are floors. If more than 10 billion chips shipped this quarter and that is 1.5 billion more than a year earlier, last year's quarter was above 8.5 billion; and since Arm calls this the first quarter above 10 billion, last year's was at or below it. Growth therefore lies between 1.5 ÷ 10 = 15.0% and 1.5 ÷ 8.5 = 17.6%. Against printed royalty growth of 8.0%, the quotient a reader can compute — royalty revenue per shipped chip — moves by 1.08 ÷ 1.15 = 0.939 to 1.08 ÷ 1.1765 = 0.918, a fall of 6.1% to 8.2%.

The direction is fixed by the two printed growth rates; only the magnitude is a band.

The levels are weaker evidence than the direction, and saying so matters. US$670m over 10 billion chips is 6.70 cents; over 11.5 billion it is 5.83 cents. Prior-year revenue implied by the printed 8% is US$620.4m, which over 10 billion chips is 6.20 cents and over 8.5 billion is 7.30 cents. The ranges overlap between 6.20 and 6.70 cents, so the honest claim is a ratio under pressure, not two measured rates.

One perimeter worth drawing. Annualising the quarter's licence-and-other line by four — US$773m × 4 = US$3,092m — and setting it against ACV of US$1,790m compares nothing to nothing. ACV is "the total annualized committed fees, excluding any potential future royalty revenue, for all signed agreements deemed to be active through the last day of each applicable reporting period", and the Form 6-K states its divisor: Total Access and Architecture Licence Agreements over the years in the contract, other licences, including single-use and limited-use licences under Arm Flexible Access or a Technology Licence Agreement, "over three years based on the historical licensing patterns of our customers", and the Arm China fee likewise. ACV is not reconcilable to GAAP revenue (Form 6-K, November 2026). That divisor belongs to Arm.

What a reader can check

What can be recomputed from printed values: the chip count against its own year-earlier floor, the royalty line against its own growth rate, and the quotient between them, which falls. What cannot be recomputed is the number Arm names as its driver, because a per-chip royalty is negotiated design by design, described in the filing as declining with volume subject to a minimum, and never printed as a rate.

That is a documentary gap, not a demonstrated defect. Arm's numbers can all be right and the average can still fall. The point of printing a denominator is that it invites a division; the point of calling the rate a declining schedule is that the division was never the rate. Anyone quoting "higher royalty rates per chip" and "more than 10 billion chips" in one breath should notice the release now makes it possible to divide them, and should say which of the two things they mean.

Sources