ARM — Research Report (2026-07-20)

TL;DR

• What it is: ARM designs the CPU blueprints ("architecture") that power nearly every smartphone and a fast-growing share of AI data-center chips — it doesn't make chips, it licenses the plans and collects a royalty on every one shipped.
• My take: the best royalty business in tech, priced for a decade of flawless execution — the fundamentals are excellent, but the stock has almost no room for a stumble.
• One-line educational decision summary: Watch — Medium conviction (educational framework, not advice; see §5g below).
• $4.92B FY2026 revenue, +23% (royalty model scaling with almost no extra cost) — FY2026 results
• ~96% gross margin (this is what a pure IP-licensing business looks like — almost every new dollar is profit)
• ~$285B market cap at ~318x trailing P/E (the market is pricing in years of growth that hasn't happened yet)
• ~87% owned by SoftBank (float is tiny — one shareholder controls ARM's fate, and it's leveraging its stake)
• Strongest bull point: Armv9 + Compute Subsystems roughly double the royalty rate per chip, and Nvidia's new Arm-based AI PC chip line just opened a brand-new market.
• Strongest bear point: at ~120-300x forward earnings depending on basis, even HSBC — a bank that still likes the stock — says the price has "no room for operational errors."
• The one thing to watch next: Q1 FY2027 earnings, July 29, 2026 — first real read on whether AI PC/data-center royalty momentum from Nvidia's Computex announcement shows up in actual numbers.

1. Business overview

So what: ARM doesn't make anything — it sells the intellectual property (IP, the legal blueprint) other companies turn into chips, and gets paid twice: once upfront, once per chip forever after.

ARM Holdings plc, based in Cambridge, UK, designs CPU architectures — the instruction set and blueprint a chip is built around, comparable to a house's structural blueprint rather than the finished house. It sells three things:

The business model: an upfront licensing fee (paid once, or over a multi-year term) plus a royalty — a small percentage of the price of every chip shipped that uses ARM's IP, for as long as that chip design ships. ARM doesn't report by product segment; it reports by revenue type: Royalty revenue and Licensing & other revenue.

Customers: over 500 companies hold active ARM architecture licenses, including Apple, Qualcomm, MediaTek, Samsung, Nvidia (Grace CPU, automotive), Amazon (Graviton), Google (Axion, Tensor), Microsoft (Cobalt), AMD, and Intel — CNBC.

Customer concentration: ARM's own 20-F (annual report filed with the SEC by non-US companies) discloses top-5 customers (including ARM China) at 57% of FY2026 revenue, vs. 56% in FY2025 — StockTitan. Apple, Qualcomm, and MediaTek combined are estimated (not company-disclosed) at roughly 40–50% of royalty revenue — label as estimate.

Geography: China revenue was $749M, ~19% of FY2026 total, run through a joint-venture-like structure (Arm China) that ARM doesn't fully control — a governance and geopolitical wrinkle unique to this market — StockDividendScreener. US + Asia together are estimated at over 80% of revenue.

Fiscal year note: ARM's fiscal year ends March 31. "FY2026" = the year ended March 31, 2026. As a UK-incorporated company, ARM files Form 20-F (the annual report format for foreign private issuers) rather than the US 10-K.

2. Financials — 5 years

So what: revenue nearly doubled in four years while gross margin stayed near-100%, but net income and free cash flow are still tiny relative to the stock's price.

*FY2026 operating income derived from company-reported 18.3% GAAP operating margin × revenue — not independently line-item confirmed. **FY2026 OCF/capex are last-twelve-month figures from a secondary aggregator, not confirmed against the FY2026 20-F cash-flow statement — flagged as approximate, a gap in this pass's research. Sources: FY2024/FY2025 20-F filings, FY2026 results release, StockTitan FY2026 20-F.

Free cash flow: estimated ~$975M LTM (last twelve months); exact FY2026 figure needs primary-source confirmation — a genuine gap after this research pass.

FY2026 detail (year ended March 31, 2026):

Note the large gap between GAAP net income ($904M / $0.85 EPS) and non-GAAP EPS ($1.77) — the difference is mostly stock-based compensation (SBC, paying employees in shares instead of cash), a real economic cost to shareholders (it dilutes ownership) even though it's excluded from non-GAAP figures. SBC of $1,052M is larger than GAAP net income itself.

Net cash: cash of $2.75B against ~$461M debt implies roughly $2.3B net cash; one secondary source cites $3.08B (likely including unconfirmed short-term investments) — use the conservative figure until reconciled against the primary balance sheet.

Q1 FY2027 guidance (quarter ending June 30, 2026): revenue midpoint ~$1.26B; non-GAAP diluted EPS ~$0.40 — Investing.com. Actual results report July 29, 2026 — BusinessWire.

3. Key metrics — and why each matters here

So what: almost every metric here is a royalty-business metric, not a manufacturer's metric — the business itself is close to ideal, the price is the open question.

4. Unit economics — the one number this business runs on

So what: ARM's real growth lever isn't chip volume, it's the royalty rate — and that rate is roughly doubling as the industry migrates to ARM's newest architecture.

The atomic unit is a royalty rate, a percentage of a chip's average selling price (ASP) ARM collects every time that chip ships. The legacy (pre-v9) blended rate is roughly ~1.7% of chip ASP. Armv9, ARM's current-generation architecture, commands roughly double that — ARM's CFO has said "royalty rates are on average, at least double the rates on equivalent Armv8 products" — The Next Platform, TIKR.

Compute Subsystems (CSS) is estimated (not company-disclosed) at an even higher ~8–10% royalty rate. Armv9's share of royalty revenue is cited at roughly ~25%, against a long-term company target of 60–70% attach rate — most of the royalty upside is still ahead, not yet in the numbers.

Volume: Arm-based chips ship at an estimated >7 billion units/quarter across the whole ecosystem — overwhelmingly low-value legacy volume (a few cents of royalty each) — while a small, fast-growing number of high-ASP AI/data-center chips carry the new, much higher rate. This mix-shift, not unit growth, is the real story: FY2026 licensing revenue grew faster (25%) than royalty revenue (21%) — licensing today is tomorrow's royalty base.

5. Value-chain position — bottleneck or commodity?

So what: ARM is close to a monopoly in mobile CPU architecture today, but the two biggest forces reshaping the chip industry — RISC-V and vertical integration — both point at loosening that grip over time.

• ARM's moat: near-total incumbency in smartphone application processors, a 500+ company licensee ecosystem, and genuine expansion into data-center (Neoverse) and automotive.
• RISC-V — an open, royalty-free alternative instruction set — is the clearest long-term threat. Sourcing here is mixed quality: one widely-repeated claim that RISC-V has "~25% of the global processor market" (Jan 2026) is very likely overstated or measuring a narrow subsegment (e.g. microcontrollers) — flagged explicitly as unverified against a primary source (IDC/Omdia) and should not be taken at face value. Better-sourced facts: Qualcomm acquired RISC-V designer Ventana Micro Systems; Google is pushing RISC-V as a "Tier-1" Android architecture, first commercial RISC-V phones expected late 2026; Quintauris (Bosch/Infineon/NXP/STMicroelectronics/Qualcomm JV) is standardizing a RISC-V automotive platform — a direct threat to ARM's automotive plans.
• Vertical integration by ARM's own biggest customers (Apple fully in-house on architecture; Qualcomm's Nuvia-derived custom cores) is a slower-moving version of the same risk.
• The Qualcomm litigation tests ARM's contractual leverage. ARM sued Qualcomm/Nuvia in 2022 alleging its architecture license was breached; a December 2024 jury found no breach, and an October 2025 Delaware federal judge dismissed ARM's last remaining claim, calling it a "complete victory" for Qualcomm — ARM said it would appeal — BusinessWire, The Register. Qualcomm's countersuit was reportedly headed to trial around March 2026 — status should be verified as it predates this report. This loss weakens ARM's ability to police how sophisticated customers use architecture licenses.

Stated TAM (management's own estimate, from ARM's Q4 FY2026 investor materials — treat as company target, not fact): total TAM $535B (FY2026) → >$1.5T by FY2031; cloud AI/data-center TAM $330B → >$1.15T; ARM's own data-center CPU opportunity $2.4B → $100B+ (FY2031), a claimed 40x — Investing.com. Management has also guided to a long-term >$9 EPS by 2031 target.

6. Ownership, management, insider signals

So what: this is not a widely-held public company in any normal sense — one shareholder controls ~87% of the stock, and that shareholder is currently leveraging its stake to fund an unrelated bet on OpenAI.

• SoftBank Group holds 922,733,999 shares (via Kronos II LLC) against 1,064,055,252 total shares outstanding (3/31/26) — ~86.7% by direct calculation; a separately cited 87.1% likely reflects a different date basis — Form 3 via StockTitan. At IPO (Sept 2023) SoftBank's stake was ~90.6%; it has since diluted down purely through new share issuance (SBC), not through selling — SoftBank press release.
• Nvidia fully exited its ARM equity stake — final 1.1M shares ($140M) around February 2026 — while remaining a licensee/partner independent of that relationship — Bloomberg.
• SoftBank leverage overhang: SoftBank has reportedly drawn against a margin loan facility partly collateralized by ARM shares (~$8.5B drawn against a reported ~$20B facility as of late 2025) to help fund its OpenAI commitments — a real, secondary-sourced overhang risk: a leveraged, concentrated 87% holder creates latent forced-seller risk a diversified shareholder base wouldn't.
• Management ownership % and incentive-plan structure: not confirmed in this pass — recommend pulling directly from the 20-F's remuneration report.

7. M&A track record

ARM has not been a significant acquirer recently; its most consequential M&A event was being the target — Nvidia's attempted $40B acquisition of ARM (announced 2020) was blocked by antitrust regulators in 2022, after which SoftBank proceeded with the September 2023 IPO instead.

8. Valuation view

So what: on every multiple, in every source, ARM trades at a large multiple of the semiconductor sector median — the debate isn't whether it's expensive, it's whether the growth ahead justifies it.

Which multiple fits, and why: P/E, since ARM is profitable (unlike a pre-revenue story) and asset-light (unlike a capital-heavy foundry, where EV/EBITDA is the natural lens) — though genuinely hard to pin down since GAAP EPS is so much smaller than non-GAAP EPS (§2), so providers land on very different "forward P/E" numbers. Sources: GuruFocus, Barchart/HSBC.

What a reasonable buyer might pay: no one is acquiring an 87%-SoftBank-controlled company, so the useful question is what return the price implies. At $267/share ($285B market cap), a buyer pays roughly 58x FY2026 revenue and, even on the most conservative forward-EPS basis (~121x), prices in a decade-plus of Armv9/CSS royalty-rate doubling plus meaningful new data-center/AI-PC share. Conclusion: ARM is priced for the bull case to be largely correct; base-case execution likely means the stock is roughly fairly valued to modestly overvalued, not cheap.

9. Bull / base / bear — attributed to named institutions

So what: even ARM's own supporters describe the price as demanding near-perfect execution — this is a stock priced by conviction, not by margin of safety.

• Bull — KeyBanc Capital Markets (John Vinh): raised price target to $430 (from $300), maintained Overweight, July 14, 2026. Bank of America (Vivek Arya) raised its target to $460 (from $335) on June 23, 2026, though BofA's formal rating is Neutral — a bullish target paired with a cautious rating. Core bull driver: Nvidia's June 2026 Computex announcement of a new Arm-based AI/Windows PC chip, extending ARM's royalty exposure beyond mobile — Motley Fool (stock +15% same day).
• Base — Wall Street consensus: average price target $298.96 across 26 analysts (17 Buy / 8 Hold / 1 Sell, "Moderate Buy"), implying ~11.9% upside from $267.19 — MarketBeat, July 20, 2026.
• Bear — HSBC (Frank Lee): downgraded ARM to Hold July 14, 2026, while raising its price target to $315 (~10% upside — a valuation-pause call, not a fundamental-collapse call). Reasons: forward multiple (his calc ~288x) "leaves no room for operational errors or macro slowdowns," a premium even to Nvidia; "the market has pulled years of future growth into the current price"; foundry capacity and memory shortages threaten near-term royalty volumes; execution risk in the data-center silicon pivot — Barchart. No dedicated short-seller/activist report was found (Hindenburg Research shut down January 2025; no other firm has published a targeted ARM thesis as of this research) — stated explicitly rather than fabricated.
• Catalysts: Q1 FY2027 earnings July 29, 2026; Armv9/CSS attach-rate progression toward 60–70%; Qualcomm countersuit resolution; any China export-control change (ARM's Neoverse V-series requires export licenses for Chinese AI/HPC customers, ~19% revenue exposure) — ARM's CEO has argued blanket bans on general-purpose CPUs are impractical ("like banning oil") — SEC 20-F FY2026, Crypto Briefing.
• What would prove the thesis wrong: Armv9/CSS attach rate stalling well below the 60–70% target; RISC-V gaining real (primary-sourced, not promotional) share in automotive or smartphone processors within 1–2 years; a SoftBank forced sale or margin-call event; two straight quarters of decelerating licensing revenue; or the disclosed $2B AI CPU demand pipeline (FY2027–28), of which only $1B currently has secured supply, failing to convert due to physical capacity constraints.

10. PE / quality lens

So what: a private-equity buyer would love the moat and margins, and would be nervous about exactly one thing — that "cheap" and "ARM" cannot currently be used in the same sentence.

Love: ~96%+ gross margin, near-monopoly incumbency in the highest-volume compute category on Earth, a royalty-rate tailwind (Armv9/CSS) requiring no new capex to capture, genuine new-market optionality (AI PC, data center) opening for the first time in 2026. Fear: SBC that exceeds GAAP net income; a controlling shareholder (SoftBank) that is leveraged and pursuing unrelated strategic bets; a valuation multiples above sector median on every measure; a legal loss (Qualcomm) that weakens ARM's ability to enforce licensing terms against sophisticated customers.

11. Practitioner decision layer (educational framework — not investment advice)

Reverse expectations (Mauboussin approach — multiple-and-consensus approximation, not a precise DCF)

At ~$285B market cap and FY2026 revenue of $4.92B, the market prices ARM at roughly 58x trailing sales. Using the forward P/E most consistent with non-GAAP EPS (~121x, GuruFocus) against non-GAAP diluted EPS of $1.77, the implied path requires sustained 20%+ revenue growth for several more years plus material non-GAAP margin expansion or share-count discipline to grow into the multiple within a normal 5-7 year horizon. Base-rate check: few semiconductor-adjacent companies sustain 20%+ revenue growth for 5+ consecutive years once past $5B in revenue (Nvidia is the standout exception, aided by an AI demand shock ARM doesn't yet have at the same scale). Judgment: stretched, not implausible — the Armv9/CSS royalty-doubling mechanism is a real structural tailwind, but the price already assumes it works close to perfectly.

Variant perception

1. Consensus believes: ARM is a high-quality royalty compounder whose Armv9/CSS mix-shift and new AI PC/data-center TAM justify a premium multiple.
2. This report's differentiated view: Consensus agreement — no strongly supported variant perception found; this analysis lands close to the sell-side range (HSBC's own downgrade already captures the "priced for perfection" caution).
3. Evidence: even HSBC (the bank that downgraded ARM) raised its target and stayed constructive long-term; 17/26 analysts remain Buy; SBC/dilution and RISC-V risks are widely discussed already.
4. Why the market may be wrong: the clearest candidate for genuine variant perception is the SoftBank leverage overhang (§6) — a forced-sale/margin-call risk on an 87%-controlling, actively-leveraging shareholder gets little attention relative to product-cycle narratives and isn't fully priced by fundamentals-driven sell-side targets.

Probability-weighted expected value

EV ≈ 0.25×445 + 0.50×299 + 0.25×175 ≈ $305, roughly 14% above the July 17, 2026 price of $267.19. A scenario-discipline exercise, not a price forecast.

Pre-mortem: it's 2028, the stock has halved

1. Multiple compression without a fundamental break: growth merely decelerates to mid-teens (still healthy), but a market that priced in perfection re-rates down to a "normal" 40-60x forward P/E. Early signal: two straight quarters of licensing revenue growth below 20%. Invalidated by: licensing revenue re-accelerating on new CSS/data-center wins.
2. RISC-V erosion becomes real, not promotional: a credible primary source (IDC/Omdia) confirms meaningful automotive or smartphone share loss. Early signal: a top-5 licensee announcing a RISC-V design win in a flagship category ARM currently owns. Invalidated by: RISC-V staying confined to microcontrollers/peripherals.
3. SoftBank forced-sale event: leveraged, ARM-collateralized borrowing against OpenAI commitments triggers a margin call, flooding an unusually thin free float. Early signal: SoftBank Group's leverage ratios deteriorating or its stock falling sharply (precedent: the Dec 2025/Jan 2026 ~20% SoftBank plunge tied partly to Arm earnings). Invalidated by: SoftBank reducing or paying down the margin loan.

KPI driver tree and sensitivity

Royalty revenue = (chips shipped using ARM IP) × (blended royalty rate) × (average chip ASP). Two drivers management can actually move: Armv9/CSS attach rate (currently ~25%, targeting 60-70%) and average royalty rate per chip (roughly doubles from ~1.7% legacy to ~3-4%+ under Armv9, higher under CSS).

• If attach rate moves +10 percentage points faster than expected, royalty revenue growth could plausibly run several points above the ~21% FY2026 pace — a rough sensitivity, not a model; effect depends heavily on which chip categories convert first.
• If chip unit volumes fall -10% in a cyclical downturn, royalty revenue takes a roughly proportional hit unless offset by rate mix-shift — the exact dynamic HSBC flagged with foundry/memory-shortage concerns.
• Licensing revenue (47% of FY2026 total) is the forward indicator for royalty revenue 1-3 years out — a slowdown here is the earliest legitimate warning sign.

Scuttlebutt proxies

• Hiring/employee sentiment (Glassdoor): ARM rates 4.5/5 overall (2,641 reviews), 89% would recommend to a friend, 88% positive business outlook — well above the ~3.7/5 IT-industry average — Glassdoor. A healthy internal-morale signal.
• Customer earnings-call mentions: ARM's own Q3/Q4 FY2026 calls describe Nvidia and Qualcomm "scaling Arm-based platforms in robotics and autonomous systems," with licensing revenue accelerating each quarter (Q3: $505M +25% y/y; Q4: $819M +29% y/y) — Motley Fool transcript. This is ARM's own framing, not independently sourced from the customers' own calls.
• Supplier/capacity signal: ARM discloses AI CPU demand commitments exceeding $2B across FY2027-28 with secured supply-chain capacity covering only the first $1B — a company-sourced data point indicating a real physical constraint on converting demand into revenue.

Decision (Educational — not investment advice)

• Decision: Watch
• Conviction: Medium
• Core assumption: ARM's Armv9/CSS royalty-rate doubling mechanism will keep licensing revenue growing at or above ~20%/year through FY2028, which is required to grow into the current ~120x+ forward multiple within a normal investment horizon.
• Review date: 2026-07-29 (Q1 FY2027 earnings release)
• Controlling question: Did licensing revenue growth accelerate or decelerate versus the ~25% y/y pace set in Q4 FY2026, and did management update the $2B AI CPU demand pipeline's supply-chain conversion?
• Review trigger: Two consecutive quarters of licensing revenue growth below 15% y/y, OR any disclosed SoftBank margin-loan/leverage event involving ARM shares as collateral.

📚 What this company teaches

1. A royalty business scales without capex — but the royalty rate matters more than volume. ARM's real growth engine isn't more phones being sold (unit growth is slow), it's the industry migrating to Armv9/CSS, roughly doubling the take per chip. When evaluating any licensing/royalty model, ask: is growth coming from more volume, or a better rate on the same volume? The latter is usually more durable.
2. GAAP vs non-GAAP earnings can tell two very different stories — and the gap itself is information. ARM's stock-based compensation ($1,052M) exceeds its GAAP net income ($904M). Non-GAAP EPS ($1.77) is more than double GAAP EPS ($0.85). The size of the gap is a signal about real dilution cost to shareholders.
3. Ownership concentration is a risk factor independent of the underlying business. An 87%-controlled company with a leveraged controlling shareholder carries a structural risk (forced-sale/overhang) that has nothing to do with chip royalties — always check who owns the stock and why they might need to sell, not just what the business does.

Self-test: ARM's forward P/E is cited anywhere from ~121x to ~288x depending on the data source. Name two legitimate reasons two reputable data providers could disagree this much on the same company's "forward P/E" in the same week — then explain which reason matters more for deciding whether ARM is actually expensive.

📖 Glossary

• Architecture license: the right to design a fully custom CPU core using ARM's instruction set — the most flexible, highest-tier license ARM sells.
• Instruction set / architecture: the underlying "rulebook" a CPU is built around — like a blueprint's structural code, distinct from the finished building (the chip).
• Royalty rate: the percentage of a chip's selling price ARM collects every time that chip ships, for the life of that chip design.
• Compute Subsystems (CSS): ARM's newest pre-integrated chip design product — commands a higher royalty rate than standard IP licenses.
• P/E (price-to-earnings): share price divided by earnings per share — how many dollars an investor pays for $1 of annual profit.
• Forward P/E: P/E using expected future earnings instead of past earnings — more forward-looking but estimate-based.
• EV/EBITDA: enterprise value (market cap + debt − cash) divided by EBITDA — a debt-neutral way to compare whole-business prices.
• GAAP vs non-GAAP earnings: GAAP = official accounting-rule earnings (includes real costs like stock comp); non-GAAP = a company's own adjusted version, typically excluding items like stock comp.
• Stock-based compensation (SBC): paying employees partly in shares instead of cash — a real dilution cost to existing shareholders.
• Free cash flow (FCF): operating cash flow minus capital expenditures — what's actually left over.
• Net cash / net debt: cash and equivalents minus total debt — positive means the company could pay off all debt today and have cash left.
• RISC-V: an open-source, royalty-free CPU instruction set competing with ARM's proprietary licensing model.
• TAM (total addressable market): the total revenue opportunity a company believes it could capture — a management estimate, not a guarantee.
• 20-F: the annual report SEC form for foreign private issuers (like UK-based ARM), roughly equivalent to a US 10-K.
• Float: the portion of shares actually available for public trading — ARM's float is small because SoftBank holds ~87% of shares outstanding.

Sources: FY2026 results release, FY2026 20-F via StockTitan, FY2025 20-F (SEC EDGAR), MarketBeat, GuruFocus, Barchart/HSBC downgrade, Motley Fool (Nvidia Computex catalyst), Bloomberg (Nvidia stake exit), BusinessWire/The Register (Qualcomm litigation), Glassdoor, The Next Platform. Estimates and dispersed figures are labeled inline; the RISC-V "~25% market share" claim is explicitly flagged as unverified. Educational research framework, not personalized investment advice.

I couldn't save this to investing/arm-2026-07.md — the write needs your explicit permission. Want me to save it there (following the vault's naming convention alongside nvda/tsm/asml)?