ARMArm Holdings plc

ARM — dossier, 2026-09-21

Slot names-ai-capex-2. Model-drafted unattended by the cloud research routine; nothing here is human-verified. Sources and locators are in relationships.yaml; figures below carry theirs inline. Anything from background knowledge is marked [background].

Price reference: $275.61 close, 2026-09-18 (ROIC.ai get_latest_stock_price NASDAQ:ARM; volume 6,946,800 shares, about $1.9bn traded). Shares outstanding 1,064,055,252 at 2026-03-31 (20-F cover, dei EntityCommonStockSharesOutstanding) → roughly $293bn market capitalisation, of which ~13.6% is public float — about 145m shares, ~$40bn.

S1 — Source

watchlist (owner, Apple Stocks, 2026-09-20), cluster ai_semis, coverage sec_fpi, cik 1973239. No Tier 0 promotion. As a foreign private issuer Arm files a 20-F and 6-Ks — no 10-Q, no proxy.

S2 — Kill test

XBRL company facts from the SEC filings (data.sec.gov/api/xbrl/companyfacts, CIK 0001973239); fiscal years end 31 March.

Quarterly (6-K XBRL):

Survives, comfortably — the balance sheet is not the question here. Arm carries no borrowings at all: the XBRL facts contain no debt tag of any kind, only lease liabilities. Cash was $3,058m at 2026-06-30 and equity $8,630m, both rising every quarter in the series. Operating cash flow was $1,524m in FY26.

Three things the kill test does surface:

• Dilution is mild by semiconductor standards — 1,025m diluted shares (FY22) → 1,068m (FY26), +4.2% in four years. That is the one capital-structure line that flatters Arm against its peers.
• R&D is 56% of revenue ($2,776m on $4,920m) and rising; it was 37% in FY22. Operating margin has been sacrificed to it, deliberately.
• GAAP operating income went backwards in Q1 FY27 — $91m against $114m a year earlier on 22% more revenue — because stock compensation was $343m in the quarter, 3.8x operating income and up 42% year on year. Net income of $270m in the same quarter is below the operating line only because interest and tax ran the other way; it is not operating earnings.
• Analyst coverage count: not obtained — see "Evidence that was thin".

S3 — Filings, gate-minimum

Read: FY26 20-F (Item 3.D summary of risk factors and the individual factors on concentration, the PRC, Arm China and controlled-company status; Item 5 results of operations; the auditor's report and its critical audit matter; Item 7.B related-party transactions), the Q1 FY27 6-K (condensed financials, balance sheet related-party lines, Item 5-equivalent MD&A), and the EDGAR form census. No transcript was read. No proxy exists — FPIs do not file one, and the skill's gate-minimum table says to skip it.

The three biggest risks management itself discloses

1. Customer concentration that includes the controlling shareholder. "our top five customers (including Arm China and SoftBank Group) collectively accounted for approximately 57%, 56% and 54% of our total revenue for the fiscal years ended March 31, 2026, 2025 and 2024, respectively" (20-F, Item 3.D). Two of the five top customers are the parent and an entity the parent owns. Separately, Qualcomm — a litigation adversary — "accounted for 9% of our total revenue for the fiscal year ended March 31, 2026".
2. The PRC reaches Arm only through an entity Arm does not control. "Substantially all of our PRC-related revenue is earned through the intellectual property license agreement ("IPLA") with Arm China" (20-F, Item 3.D); PRC revenue was 18% of FY26 total, 19% in FY25, 22% in FY24. Arm transferred its entire equity interest in Arm China to a SoftBank subsidiary on 2022-03-28, has no board seat of its own, and "Arm China's payments due to us are determined based on the financial information that Arm China provides to us". Its former CEO has been litigating against its governance in PRC courts since April 2022.
3. Customers can design around the product they license. "Customers may decide to license our architecture and develop their own processors based on our ISA, rather than utilize our predeveloped products through an implementation license" (20-F, Item 3.D) — and Arm is now competing with those same customers: "some of our customers may face direct competition from us in silicon production products, such as with the Arm AGI CPU."

The finding this run exists to record

Arm reports revenue in two lines on the face of the income statement — external customers, and related parties (SoftBank Group and Arm China):

Management's own MD&A states the split of the growth: "Revenu e from external customers increased $237 million, or 7%... Revenue from related parties increased $676 million, or 82%... primarily driven by a $591 million, or 141%, increase in license and other revenue". Of the $913m of FY26 revenue growth, $676m — 74% — came from the controlling shareholder's own orbit, and most of that was licence revenue rather than royalties. Reported growth was +23%; growth from unrelated customers was +7%.

Two things sharpen it and one softens it.

• Sharpening: the auditor's sole critical audit matter is "Revenue Recognition for Material Long-Term Revenue Contracts", whose description opens "The Company routinely enters into material long-term revenue contracts with customers, including related parties" and whose audit response includes "an evaluation of whether the transaction price and corresponding performance obligations were fairly presented for related party transactions". The auditor flagged the same line.
• Sharpening: the Shareholder Governance Agreement exempts SoftBank transactions existing at the IPO from related-party review altogether, and exempts ordinary-course transactions the audit committee judges to be on third-party terms below a $20m threshold.
• Softening: Q1 FY27 does not extend the pattern. External revenue was $901m, +24%; related parties $388m, +18%; the related-party share held at 30.1% rather than climbing. One quarter is not a trend, but it is the opposite of the FY26 shape.

Balance-sheet exposure to the same counterparties is material and mostly unbilled: contract assets from related parties $577m at 2026-06-30 ($646m at 2026-03-31) and receivables from related parties $281m ($270m) — together about 57% of a full year of related-party revenue.

Red-flag sweep

• Auditor: Deloitte; opinion clean, one critical audit matter as described above. No auditor change and no NT 20-F in the form census since 2023.
• Going concern: not applicable — no debt, $3.1bn cash, positive operating cash flow.
• Related party: this is the related-party story; see above.
• Restatement / comment letters: the census shows 5 UPLOAD and 5 CORRESP items, consistent with the 2023 IPO review period. Not verified by full-text search — see "Evidence that was thin".

S5 — Ownership

Control. "As of May 21, 2026, SoftBank Group beneficially owns approximately 86.4% of our total issued and outstanding share capital, representing a majority of the total voting power" (20-F, Item 3.D). Arm is a Nasdaq controlled company and relies on the exemptions; as an FPI it files no proxy and no domestic Section 16 reports. The public float is about 145m shares.

Institutional register. threespread.py holders ARM returned zero 13F filers for CUSIP 042068 205, so this came from the Equibles fallback (13F report date 2026-06-30, 1,058 rows across 985 institutions, 156,099,393 shares including option notional). Largest common lines: Invesco 12,816,588 (8.21% of institutional 13F shares); DZ Bank 6,504,147; Capital World 4,330,875; Arrowstreet 4,094,696; FMR 4,062,977. Nobody holds a controlling minority stake — SoftBank has it.

The option book is where the opinion is. Five of the top twelve 13F lines are options, and the second-largest line in the whole name is a put: Susquehanna 9,099,400 shares' notional, against a call line of 8,445,200. Jane Street (put 7,079,600 / call 4,311,800), Citadel (put 5,345,100 / call 3,319,900) and Barclays (put 2,967,200) show the same two-sided shape. In a stock with a 13.6% float, the derivative book is a large share of the tradeable opinion.

Short interest (FINRA via Equibles): 16,740,615 shares at 2026-08-31, down 1,523,490 on the settlement. That is ~1.6% of shares outstanding but roughly 11-12% of the public float, at 4.4 days to cover — several times MRVL's 1.4.

Insider transaction cadence: 94 Form 4 filings in the twelve months to 2026-09-21 (Arm files them despite FPI status), latest 2026-09-16, plus 35 Form 144s in the census since 2023. Filings, not transaction detail — a cadence reading only.

What hangs over it. No shelf: the EDGAR census since 2023 shows 20-F, 6-K, one SC 13G (2024), S-8s and the IPO F-1 family, and no F-3 or F-3ASR. There is therefore no registered secondary on file — but SoftBank's 86.4% is a standing overhang of a different kind, and the Shareholder Governance Agreement grants it registration rights.

S8 — Valuation and sizing ⏳ OPEN (user)

Guidance evidence is two rows only (Equibles GetGuidance, from the 2026-07-29 call): Q2 FY27 non-GAAP operating expense $780M and Q2 royalty growth 13%. Arm does not put a revenue guide into the structured record the way Marvell does, so the model below leans on the reported trajectory, and its assumptions are mine, not management's.

For reference, reported FY26 GAAP EPS was $0.85 ($904m / 1,068m), so the stock is on ~325x trailing GAAP earnings, and the gap between the two columns is stock compensation running at ~$1.4bn a year against ~$0.9bn of GAAP operating income.

Bear / base / bull on the FY28 line: bear — royalty growth settles at the guided 13% and licence revenue normalises without the related-party step-up, revenue ~$6.3bn, EPS ~$2.10, 40x gives ~$84; base — the table, EPS $2.81, 60x gives ~$169; bull — AGI CPU and compute-subsystem royalties inflect, revenue ~$8.5bn, EPS ~$3.60, 80x gives ~$288. Only the bull case reaches today's price, and it needs both a 25%+ revenue path and a multiple that itself assumes one.

The honest summary: on any multiple this pipeline can defend, the price already contains the bull case. Arm's business quality is not in question — 97.5% gross margin, no debt, an architecture with no substitute — but at 98x a model FY28 that assumes 25% growth, the margin of safety is the multiple, and there isn't one.

Proposed buy price and proposed size are owner-only and are not stated here. Note for sizing: the 13.6% float, not the $1.9bn of daily turnover, is the real liquidity constraint — an 86.4% holder means the tradeable stock is a small fraction of the market capitalisation, and short interest is already 11-12% of it.

S9 — Monitoring

• Next earnings: Q2 FY27 results, expected early November 2026 — derived from filing cadence (Q2 FY26 results 6-K filed 2025-11-05), not from a calendar feed.
• Filing feed: the Tier 0 EDGAR sweep covers cik 1973239 (watchlist.yaml carries it), so 20-F and every 6-K arrive without further setup. Arm's 6-K stream is the only quarterly disclosure.
• Leading indicator between quarters: the related-party revenue line on the face of the income statement in each 6-K, read as a percentage of total and against external growth. It is one number, disclosed every quarter, and it is what separates a licensing business compounding at 24% from a parent booking licences into a subsidiary at 86.4% ownership.
• Trigger that converts this to a buy: two consecutive quarters of external-customer revenue growth at or above 20% with the related-party share flat or falling, and royalty revenue growing faster than licence revenue, and a price at or below ~60x the model FY28 EPS — about $169 on today's model, roughly 39% below the 2026-09-18 close. Nothing in the business needs to break for that to happen; the multiple simply has to stop assuming the bull case.

Evidence that was thin, and why

• No earnings-call transcript was read, and Arm's structured guidance record holds only two rows. The gate-minimum spec asks for two transcripts and the Q&A; this run traded them for budget across three names. For Arm specifically the cost is higher than for Marvell, because the call is where a controlled company explains a related-party licence.
• The institutional register came from the Equibles fallback, not 3spread — threespread.py holders ARM returned zero filers for the ADS CUSIP (042068 205). The Equibles rows mix common and option positions and are shares-of-13F-total, not shares-of-outstanding, so no ownership percentage of the company can be read off them.
• No efts.sec.gov full-text sweep for restatement or auditor-change language; the auditor and comment-letter reads come from the form census and the opinion instead.
• No analyst coverage count.
• Rene Haas's role is background knowledge, not read from a filing in this run — the node is marked PROPOSED_UNRESOLVED and no edge depends on it.
• Next earnings date is derived from filing cadence, not a calendar feed (ROIC.ai's earnings calendar needs a paid plan on this account).