Thesis: Datacenter-CPU royalty ramp + hyperscaler captive build-out is structural; ARM is the IP-layer beneficiary of every Grace/Graviton/Cobalt/Trainium chip shipped, but valuation (P/E forward 100x, P/S 66x) prices in flawless FY27 datacenter mix-shift execution.

User request 2026-05-06 — six specific questions answered below rather than the generic template. ARM is structurally adjacent to ai-power-bottleneck (cited in that perspective's watch_for as "ARM positioning at Q2 13F — chip-IP royalty pricing-in cross-check"), already covered in ai-infra + semis + mag7-adjacent + focus. No watchlist gap.

Tape (2026-05-22 close)

The tape geometry is the dominant fact: ARM has roughly doubled in 3 months, is two days off the all-time high, and trades at forward P/E 100x. This deep-dive is not an entry recommendation at current tape; it grades the thesis (medium) separately from entry (deferred — see Setup).

Q-by-Q financials (ARM fiscal year ends 31 March)

Annual context: FY25 revenue $4,007M, op income $831M (20.7%), net $792M, EPS $0.61. FY24 revenue $3,233M, op income $117M (3.6%), net $306M, EPS $0.29. Revenue growth FY25 vs FY24 = +24.0%; FY26 4-quarter EPS sum = ~$0.84 (consistent with the annual $0.85 figure); FY26 annual revenue = Q1–Q4 sum $4,920M (matches the $4.92B press-release figure) = +22.8% revenue growth FY26 vs FY25, only a mild deceleration from FY25's +24% (2026-06-05 re-fetch filled the real Q4; the earlier "~$4.7B / +17-18%" was the pre-Q4-data placeholder).

Quality-of-earnings notes: Q4 FY25 op-margin spike to 33% reflects a backlog/license-revenue catch-up — non-recurring; FY26 quarterly op-margin running 11-15% reflects elevated R&D + ramp investment. Gross margin pinned ~97% confirms the IP-licensing model (no inventory cost). Net cash position: $3.6B cash vs $432M total debt = $3.2B net cash. Trailing FCF ~$1B against $326B market cap.

Six questions (the actual ask)

(a) Stack position — does ARM belong in ai-power / mag7 / a new bucket?

ARM is the license-IP layer above silicon: NVIDIA Grace CPU, AWS Graviton (4 generations), Apple Silicon (M-series, A-series), Microsoft Cobalt 100, Google Axion, Tesla Dojo control plane — all ARM-architected. It is supplier to the hyperscalers, not a hyperscaler. Therefore:

• Not mag7 — that bucket is the seven large-cap consumer/cloud names; ARM is a different shape (sole-source IP vendor, no consumer-facing platform, no advertising/cloud revenue).
• Already in ai-infra + semis + focus + insider-universe + ai-scan + tech-insider-buys — coverage is correct.
• Already cross-cited in ai-power-bottleneck perspective watch_for as the chip-IP royalty pricing-in cross-check.
• No new watchlist warranted. ARM is the only public "license-IP-only" name at scale (Cadence + Synopsys are EDA tooling, structurally different). A "license-IP" cohort of one is not a cohort.

(b) Royalty math — v9 uplift trajectory + datacenter-CPU per-chip economics

Two royalty streams: architectural licenses (Apple, Qualcomm, AWS, NVIDIA pay big upfront fees + lower per-chip rate to modify the architecture) and off-the-shelf designs (Cortex/Neoverse cores pay higher per-chip rate but no upfront). v9 ISA launched 2021; ARM's public commentary at FY24/FY25 calls suggested v9 royalty rates ~2x v8. v9 adoption mix has been a key disclosed metric — moving from ~10% of royalty revenue in mid-FY24 toward 25-30% by mid-FY26 (verify against most recent shareholder letter).

Per-chip royalty rate (order-of-magnitude estimate from industry channel checks, not from ARM disclosure):

• Smartphone SoC (v8/v9 Cortex-A): $0.30–$0.50/chip (refined 2026-05-28 → $0.40–0.80; Apple <$0.30 on legacy ALA, Android/Qualcomm $0.40–0.80)
• PC/laptop CPU (M-series, Cobalt): $1–$5/chip (refined → $1–4)
• Datacenter CPU (Neoverse-based Grace/Graviton/Cobalt server parts): $20–$50/chip (refined → least-grounded number here; $15–50 at list rate, but architecture-license discounts à la Apple's ~85%-below-list could pull effective $/chip to $5–25 — see per-chip empirical addendum)

What ARM actually discloses is a rate, not a $/chip (empirically confirmed 2026-05-28): royalty ≈ 1.7% of chip TAM value; "typical" rate 1–2% of ASP. Per-chip dollars = that rate × the chip's ASP. Blended across all Arm chips: ~$0.055/chip (FY23, $1,675M ÷ 30.6B) rising to ~$0.077 (FY26) — the rise is the mix-shift thesis in one number (not a rate hike; more expensive, more core-dense chips).

Datacenter mix-shift math: 1 server CPU = ~50-100 smartphone SoCs in royalty revenue. Hyperscaler datacenter CPU TAM growing 30-50% per year; smartphone CPU shipments flat-to-down. The arithmetic argues for revenue acceleration as datacenter mix rises — but the inflection requires datacenter unit volumes that haven't yet materialized at scale (Graviton is the most-deployed by AWS, ~20% of new EC2 launches per AWS Q1 2026 commentary; Cobalt and Axion are still ramping).

(c) Arm China overhang — Allen Wu litigation status

Allen Wu was forced out of Arm China in 2020; the ownership/governance dispute structurally overhung ARM's IPO prep through 2022-23. Resolution path: a new ownership structure was established under which Arm Holdings via subsidiary owns approximately 47% of Arm China, with the remainder held by Chinese investors. Arm China operates as an independent entity that licenses ARM IP to Chinese customers under a defined arrangement.

Revenue at risk: Arm China historically contributed ~20-25% of ARM consolidated revenue; that contribution is now recognized via the license arrangement rather than direct subsidiary consolidation in the same form. Residual risk is two-fold: (1) any further deterioration in US-China tech export controls could restrict licensing of advanced (v9, Neoverse) cores to Chinese customers; (2) Arm China is structurally outside ARM's direct operational control — IP enforcement and revenue collection rely on the bilateral arrangement holding. Not a current-quarter catalyst, but a recurring tail-risk that flares with each new export-control headline.

Correction (2026-05-28): ARM does not own ~47% of Arm China. It transferred its entire equity interest to SoftBank's Acetone Limited in 2022 and now holds only a ~4.8% indirect, non-voting interest with no board or management rights. Current revenue share is ~17% (FY2025) and falling (24%→21%→17% over FY23-FY25). The ~20-25% estimate above is stale. See the Disclosure addendum (2026-05-28).

(d) Hyperscaler captive-CPU pivot — friend or foe?

Friend. Every captive CPU shipped (Graviton, Cobalt, Grace, Axion, Trainium-control-plane) pays ARM a royalty. The risk people sometimes voice — "captive CPUs squeeze ARM royalty economics" — confuses architectural-license terms with off-the-shelf-design terms. Architectural licensees pay larger upfront fees in exchange for lower per-chip rates; but the per-chip absolute royalty on a datacenter CPU is still meaningfully positive (estimated $20-50/chip) AND the volume curve is growing 30%+ annually.

Refinement (2026-05-28): the ALA discount can be large — Apple, ARM's biggest unit customer, pays <$0.30 on a ~$130 chip (~0.23% effective, ~85% below the 1–2% list rate) on its legacy architecture license. The hyperscalers (AWS/NVIDIA/Microsoft) are also architecture licensees and could hold similarly steep discounts, which is why the $20–50 datacenter estimate is the softest number in this deep-dive. The "friend" conclusion still holds — server ASPs of $2–5K keep the absolute dollars large even at a deeply discounted rate — but the per-chip magnitude has a wide confidence band. See per-chip empirical addendum.

The only scenario in which captive CPUs become "foe": if a hyperscaler moves to RISC-V to escape ARM royalties altogether. This is the bear vector — see (f). Today, none of NVIDIA / Amazon / Microsoft / Google / Apple has announced a meaningful RISC-V datacenter program at production scale.

(e) Valuation comp set

ARM is the richest valuation in the AI-infra cohort by a wide margin. The premium is structurally defensible on:

• Highest gross margin (97% vs 75% NVDA/AVGO) — no silicon manufacturing cost, pure IP licensing economics
• No capex risk — ARM's investing cash flow is R&D-funded, not fab-CapEx-funded
• Royalty annuity model — once a chip ships, ARM gets paid for its full life
• Mix-shift optionality from mobile → datacenter

But premium is extreme even with all that. P/S 66x against +20% revenue growth = a PEG-on-sales of 3.3x, which is the kind of number you see at the late-cycle FOMO peak. Any quarter where datacenter mix-shift fails to accelerate visibly will compress the multiple sharply. Forward P/E 100x is the multiple for a flawless multi-year datacenter ramp — not for in-line execution.

(f) Competitive moat / RISC-V threat

RISC-V is a real open-source ISA with growing momentum in:

• Embedded / micro-controllers — Espressif ESP32-C series, SiFive cores, Bouffalo, many Chinese vendors. ARM's Cortex-M franchise is already losing some share here on price.
• Specialty accelerators — Tenstorrent (Jim Keller's company), Ventana Micro Veyron-V2 server cores. Not yet in production at hyperscaler scale.
• Datacenter CPUs — no production-scale deployments yet (May 2026). Ventana, Tenstorrent, SiFive are years from competing with Neoverse on software-ecosystem maturity.

Realistic RISC-V threat timeline:

• Embedded: continued share erosion (already happening, low-margin for ARM)
• Mobile: structurally captive (Apple, Qualcomm, MediaTek all architectural-license, would not switch ISAs)
• Datacenter: a 5-10 year story; meaningful RISC-V share would require a hyperscaler-scale software ecosystem (compiler, toolchain, kernel support, every relevant Python/Java/Go runtime) that does not yet exist
• ML accelerators: less directly relevant — accelerators are mostly post-ISA (CUDA, ROCm, custom)

ARM's moat is the software ecosystem more than the ISA itself. Every datacenter Linux distribution, every cloud runtime, every Android SDK targets ARM as a first-class platform. Replacing that takes a decade of corporate-grade investment.

Setup

Entry zone: $240–270 (-12% to -22% from current). The current tape (+147% 3M, RSI 76, two days off ATH) is mathematically not an entry — even a constructive bull would want to wait for a 15%+ digestion pullback to the SMA50 region ($182) or at least to mid-cycle ($240-270 = the price band from mid-April). Stop: $200 (~ -10% from low end of entry, below the FY26 Q3 print breakout shelf and approximating SMA50). Target (base): $360 (+18% from current; FY27 datacenter mix-shift visible, P/E rerates modestly higher on confirmed datacenter inflection). Target (bull): $420 (+37%; Graviton/Cobalt/Axion volumes step-function higher, v9 royalty mix 40%+, multiple stays at 100x forward on accelerating fundamentals). Conviction: medium — thesis is structurally sound and supported by 97% gross margins + net-cash balance sheet + clearly-defined datacenter mix-shift, but the entry geometry is hostile (P/S 66x, RSI 76, parabolic 3-month run) and there is no quarter visible in the financial statements yet where datacenter mix-shift has materially re-rated revenue growth. Buy on disappointment, not on confirmation.

Bull case

• Datacenter inflection visible by FY27 (April'26-March'27). Graviton4 ramp + Cobalt 100 production scale + Google Axion expansion + NVIDIA Grace Hopper installed base growing. Combined hyperscaler ARM-CPU unit volume doubles or more over the next 18 months; royalty mix from datacenter rises from current ~15% toward 25-30%.
• v9 royalty rate uplift mechanically compounds revenue. As v9 mix moves from ~25% to 50%+ over FY26-FY28, the implied royalty rate per chip rises ~30-40% (blended), independent of unit growth.
• Pure-IP business model: incremental margin is enormous. Gross margin 97.5% means almost every dollar of incremental revenue flows to op income absent R&D step-ups. Operating leverage on a $4B → $6B revenue ramp is non-linear.
• Net cash balance sheet ($3.2B net cash, no leverage risk) — survivable through any cyclical pause without dilution risk.
• Software ecosystem lock-in insulates against RISC-V displacement for 5-10 years in datacenter and indefinitely in smartphone.

Bear case

• Valuation prices in flawless execution. Forward P/E 100x, P/S 66x with revenue growth decelerating from FY25's +24% to FY26's est. +17-18%. The decel is the bear's strongest single number. If datacenter mix-shift doesn't visibly re-accelerate growth back above 25% by mid-FY27, the multiple compresses 30-50% just on rerate. (Refined 2026-05-28: FY26 actual revenue $4.92B = +~23% over FY25's $4.007B, not +17-18% — the deceleration is mild, 24%→~23%, which materially softens this bear point. See Disclosure addendum.)
• The +147% 3-month tape is rotation/momentum-driven, not fundamentally-driven — the FY26 Q1-Q3 prints in front of you don't yet show the datacenter inflection. The market has front-run the narrative.
• Captive CPU royalty per-chip ASP is opaque. ARM has not disclosed datacenter-specific royalty in detail; what looks like a sequential ramp could in fact be lower per-chip than the embedded base assumption.
• Arm China structural risk (~20-25% of revenue dependency) — any escalation in US-China tech export controls would directly impair revenue.
• RISC-V is a 5-10 year multi-step bear thesis, not a current-quarter risk — but at 100x P/E, a 5-year horizon view matters. If even 10-20% of incremental datacenter CPU share moves to RISC-V over FY28-FY30, the terminal-value math degrades materially.
• Optical/silicon-photonics catch-up risk — if compute consolidates further onto NVIDIA-style monolithic GPU+CPU dies (Grace Hopper as the template), the CPU-as-separate-silicon premise erodes, reducing total addressable royalty units.

Catalysts

• Q4 FY26 earnings (already printed mid-May 2026, reflected in current $306 tape) — EPS $0.29 implied from the annual sum. Full income statement needs to land with the financials provider for proper attribution. Next forward catalyst is Q1 FY27 (~August 2026) which is the first quarter post-Q4-blowout and will show whether momentum persists.
• v9 royalty mix disclosure at each quarterly print — the most important single forward metric.
• NVIDIA Q1 FY27 (2026-05-28, three days from now) — power-infra + Grace-CPU commentary is a cross-read for ARM datacenter momentum. Already flagged in ai-power-bottleneck watch_for. (Correction 2026-06-11: the 2026-05-28 date was a cadence estimate — NVDA's official IR date was 2026-05-20, so the print had already landed when this was written; see the May 31 calendar-correction note.)
• Hyperscaler capex commentary — AMZN/MSFT/GOOGL/META quarterly capex prints. Each $10B step-up in capex correlates ~1:1 with ARM-CPU royalty volume uplift (via Graviton/Cobalt/Axion).
• Any RISC-V hyperscaler announcement — currently zero; would be the single most material bear catalyst if one of AMZN/MSFT/GOOGL commits to a RISC-V production CPU.
• US-China export-control escalation affecting v9 Neoverse licensing to China — recurring tail-risk.

Open questions / follow-ups

• Q4 FY26 full income statement — the data has EPS but not revenue/op-income for the Mar'26 quarter as of the 2026-05-25 pull; needs a re-check in 1-2 weeks once the 20-F/6-K is filed (ARM as FPI files 20-F annually + 6-K quarterly). File as follow-up below.
• v9 royalty mix as % of royalty revenue — most recent management disclosure. Pull from latest shareholder letter / earnings call transcript.
• Datacenter royalty revenue mix — has ARM disclosed datacenter-specific royalty in dollar terms? If so, the multi-quarter trajectory is the single most valuable validation of the thesis.
• Arm China revenue share (current) — get latest 20-F disclosure on Arm China revenue contribution and the licensing-arrangement structure.
• Per-chip royalty empirical estimate — channel checks vs Apple/Qualcomm/AWS historical disclosures. Closed 2026-05-28 — see the May 28 per-chip-royalty investigation; bounded ranges folded into section (b)/(d) + addendum.

Disclosure addendum (2026-05-28)

Closes three disclosure-check follow-ups filed 2026-05-25 (v9 royalty mix, datacenter royalty $-mix, Arm China current revenue share). Verified against ARM's primary SEC filings — the Q4 FYE26 6-K shareholder letter (filed 2026-05-06) and the FY2025 20-F (filed July 2025) — plus analyst reads of the investor deck for the v9 chart figure ARM no longer prints in prose. Net: thesis holds; two body characterizations need correcting — Arm China ownership (material error) and the FY26 deceleration assumption (the bear's strongest number is softer than modeled).

v9 royalty mix — ~31% at last clean disclosure; target 67-70% (TASKS-RESEARCH.md:27)

• ARM no longer states an explicit Armv9-% in the earnings-release prose or slide text. It now frames the figure as an estimated "royalty technology mix… estimated at the SoC level based on the architecture of the primary CPU" (Q4 FYE26 6-K). The number lives only in a deck chart — each quarter now requires reading the presentation, not the press release.
• Last clean data point: Q4 FY2025 (Mar 2025), Armv9 ≈ 31% of royalties (Armv7 down to ~25%), per analyst reads of the deck. v9 had plateaued near ~25% for several quarters through FY25 (a documented investor concern), then stepped to "more than 30%."
• ARM's stated long-term attach target: v9 tops out around 67-70% of royalties over time.
• vs deep-dive: the body estimated "~25-30%, bull case needs v9 → 50%+ by FY28." Verified ~31% confirms the estimate was right-to-slightly-low; the bull-case 50%+ is roughly the midpoint of ARM's own terminal 67-70% target — plausible but multi-year. No thesis change.

Datacenter royalty $-mix — NOT disclosed in dollars; opacity confirmed (TASKS-RESEARCH.md:29)

• ARM does not break out datacenter / cloud-compute royalty in dollar terms. This is the gap the follow-up asked about — confirmed as a deliberate non-disclosure. Filed as known reporting opacity. The bear-case point "captive CPU royalty per-chip ASP is opaque" stands: $/chip and datacenter-$ both remain undisclosed.
• What ARM does disclose, in order of usefulness as a trajectory proxy:
  • Cloud-compute share of royalty revenue: 9% (FY2022) → 20% (FY2025) — the single best disclosed datacenter-trajectory metric (a %-of-royalty mix figure, not a dollar line).
  • Q4 FYE26: "data center royalty more than doubled year-over-year" (qualitative growth, no base disclosed).
  • ~50% share among top hyperscalers (Q4 FYE26 letter) — share, not revenue.
  • "Soon the data center will be Arm's largest business" (CEO Rene Haas, Q4 FYE26).
  • Arm AGI CPU launched — ARM's first in-house data-center chip; demand "exceeded expectations." Deck's FY2031 framing targets AGI CPU revenue ~$15B + IP/CSS ~$10B (aspirational long-term, secondary-source summary — treat as a target, not a model input).
• vs deep-dive: the body's "(b)" assumed datacenter royalty mix ~15%; ARM's disclosed cloud-compute mix is ~20% of royalty as of FY2025 — slightly higher than assumed, directionally confirming the structural mix-shift.

Arm China — 17% of revenue and falling; ownership ~4.8% indirect non-voting (TASKS-RESEARCH.md:31)

• Revenue share (FY2025 20-F): Arm China ≈ 17% (FY2025), 21% (FY2024), 24% (FY2023) of total revenue — the largest single customer, but a clearly declining concentration (24→21→17%). Top-five customers (incl. Arm China) ≈ 56% of FY2025 revenue.
• Tail-risk magnitude is shrinking, not growing. The body used a "~20-25% historical" estimate framed as a recurring tail risk; the current ~17% (down 7 points in two years) means the export-control / counterparty exposure is materially smaller than the body implied and trending down.
• Ownership — correction to body section (c). The body states "Arm Holdings via subsidiary owns approximately 47% of Arm China." This is wrong. Per the FY2025 20-F: ARM transferred its entire equity interest in Arm China to a SoftBank Group subsidiary (Acetone Limited) on 2022-03-28. ARM now holds only a 10% non-voting interest in Acetone, which the filing states "represents an approximate 4.8% indirect ownership interest in Arm China." Acetone (SoftBank-controlled) holds ~48% of Arm China; HOPU Investment Management ~35%; other Chinese parties ~17%. ARM has no board representation and no management rights; "neither we nor SoftBank Group control the operations of Arm China, which operates independently of us." The ~47-48% the body cited is Acetone/SoftBank's stake, not ARM's.
• What this changes: Arm China is a commercial IP-license counterparty (the IPLA, under which Arm China sublicenses ARM IP to PRC customers), not a consolidated or controlled subsidiary. The risk is "concentrated customer + export-control conduit," not "governance of a half-owned entity." Combined with the declining 17% share, the net tail risk is lower and better-characterized than the body's framing.

Q4 FYE26 / FY2026 actuals now public (refines the bear-case decel point)

Body flagged Q4 FY26 as "revenue not yet available." Press-release topline (Q4 FYE26 6-K, 2026-05-06): revenue $1.49B record; royalty $671M (+11% YoY); license $819M (+29% YoY); non-GAAP EPS $0.60; non-GAAP op income $731M (49% non-GAAP op margin). Full year FY2026: revenue $4.92B; royalty $2.613B (+21%); license $2.307B (+25%).

• FY2026 total revenue grew ~23% over FY2025's $4.007B — versus the body's estimate of "+17-18%, decelerating from FY25's +24%." The actual deceleration (24% → ~23%) is mild, which softens the bear's single strongest number.
• GAAP per-quarter Q4 line now folded (2026-06-05 re-fetch). Q4 FY26 GAAP: revenue $1.490B (= the press-release record), gross_profit $1.458B (97.9% GM), operating_income $440M GAAP (29.5% GAAP op margin) vs the $731M / 49% non-GAAP above — the ~$291M gap is stock-based comp + acquired-intangible amortization. Net income $313M GAAP; GAAP EPS $0.29 (vs $0.60 non-GAAP). OCF $260M / FCF $170M / capex -$90M. Folded into the Q-by-Q table above; Massive re-checked 2026-06-05, still no ARM coverage (FPI).

Per-chip royalty empirical — bounded (2026-05-28)

Closes the last open follow-up. Full work in the May 28 per-chip-royalty investigation. The deep-dive's channel-check ranges survive empirical bounding against the ARM F-1, the Apple-royalty report, and the ARM v Qualcomm trial — with two refinements:

• ARM discloses a rate, not a $/chip. F-1: royalty ≈ 1.7% of chip TAM value; "typical" 1–2% of ASP; set as % of ASP or fixed per chip, declining with volume, with a contractual minimum. Every per-chip dollar is rate × ASP. Blended: ~$0.055/chip (FY23 = $1,675M royalty ÷ >30.6B chips) → ~$0.077 (FY26 = $2.613B ÷ ~33B). The rise is the mix-shift thesis (higher-ASP, more core-dense chips), not a rate hike.
• Smartphone refined up to $0.40–0.80 — Apple <$0.30 (legacy ALA) anchors the floor; Android/Qualcomm ~$0.40–0.80 (1–2% × ~$40 ASP, cross-checked by Qualcomm's ~$300M/yr total payment to ARM ÷ ~500M chips ≈ $0.60).
• Datacenter $20–50 is the least-grounded number and now flagged as such: ARM doesn't disclose datacenter royalty in dollars (confirmed above), and the Apple precedent shows ALA customers can pay ~85% below list (~0.23% effective) — so if AWS/NVIDIA/Microsoft hold Apple-like terms, effective datacenter $/chip could be $5–25, not $20–50. Server ASPs of $2–5K keep absolute dollars large either way, so the "1 server CPU ≈ 50–100 phones" arithmetic holds directionally but with a wide band. No thesis change — the epistemic gain is knowing which number is soft and why (undisclosed + ALA-discountable), which makes any future ARM datacenter-$ disclosure or hyperscaler ALA-terms leak the highest-value datapoint to watch.

Sources

• Tape: ai-infra (2026-05-22 close), ohlc
• Fundamentals: income-statements, balance-sheets.json, cash-flow-statements.json (yfinance fallback, source documented in _source.json — Massive/Polygon returned no quarterly data for ARM as FPI, fell back to yfinance via python/)
• Company: company (Polygon — name, ADR mechanics, employee count, IPO date)
• SEC filings (6-K, 20-F): sec-filings
• Perspective context: ACTIVE (ai-power-bottleneck watch_for cites ARM positioning at Q2 13F as chip-IP cross-check)
• Prior coverage: 2026-04-10-arm-deep-dive (refresh-template, no thesis depth; this deep-dive is the substantive version)
• Cohort comp valuations: yfinance info dict (forward P/E, P/S TTM, margins, growth) for ARM/NVDA/AVGO/MRVL/QCOM as of 2026-05-22
• Disclosure addendum (2026-05-28) primary sources:
  • Q4 FYE26 6-K shareholder letter (filed 2026-05-06): SEC EDGAR Archives/edgar/data/0001973239/000197323926000062/exhibit992fye26q431-marx26.htm — royalty/license splits, "data center royalty more than doubled YoY," "~50% share among top hyperscalers," Arm AGI CPU, "royalty technology mix… estimated at the SoC level"
  • Q3 FYE26 6-K (filed Feb 2026): …000197323926000005/exhibit992fye26q331-decx25.htm — Q3 royalty $737M (+27%)
  • FY2025 20-F (filed July 2025): Archives/edgar/data/0001973239/000197323925000016/arm-20250331.htm — Arm China revenue share (17%/21%/24% FY25/24/23), Acetone Limited ownership structure (ARM 10% non-voting = ~4.8% indirect), IPLA, top-5 customer concentration (~56% FY25)
  • v9 royalty mix chart figure (~31% Q4 FY25, target 67-70%): More Than Moore (morethanmoore.substack.com) + Investing.com analyst reads of the ARM investor deck (ARM no longer prints the % in release prose)
  • Cloud-compute royalty share 9% (FY22) → 20% (FY25): Investing.com summary of ARM Q4 FY26 investor presentation
  • Per-chip royalty empirical (2026-05-28): ARM F-1 (FY2023) SEC EDGAR Archives/edgar/data/0001973239/000119312523216983/d393891df1.htm — royalty revenue $1,675M FY23, >30.6B chips, "1.7% of industry TAM", 1–2%-of-ASP rate structure; Apple <$0.30/chip (MacRumors / Tom's Hardware / Gigazine, Nov 2023); ARM v Qualcomm trial ~$300M/yr (Tom's Hardware, Dec 2024). Full work: 2026-05-28-arm-per-chip-royalty-empirical

Note on Massive fallback: ARM Holdings plc is a UK-incorporated foreign private issuer; SEC filings are 6-K (quarterly) and 20-F (annual), not 10-K/10-Q. Massive/Polygon's quarterly-statements endpoint is built off domestic-issuer XBRL filings and returns "No results found" for ARM. yfinance pulls direct from FPI-compatible sources and returned 5 complete quarters (Q3 FY25 through Q3 FY26) plus partial Q4 FY26 EPS. Saved to massive with a _source.json provenance note so the deep-dive-financials-lint can resolve the paths and so future agents can see the fallback was deliberate.