Arm Holdings plc American Depositary Shares ARM

Reference table

1. What the business is, and how the disclosure actually works

Start with the disclosure regime, because it changes what is knowable. Arm Holdings plc is incorporated in England and Wales, headquartered at 110 Fulbourn Road, Cambridge, and listed on Nasdaq Global Select as American Depositary Shares since 2023-09-14. It is a FOREIGN PRIVATE ISSUER. It files a Form 20-F annual report, not a 10-K, and Form 6-K current reports instead of 10-Qs and 8-Ks. There is no 10-K and no 10-Q in existence for this company, and the 20-F says why in its own risk factors: "we are not required to file periodic reports and financial statements with the SEC as frequently or as promptly as U.S. public companies... our shareholders will not be afforded the same protections or information than what is accorded to investors holding shares in public companies organized in the U.S."

The archive we hold is exactly three documents: the 20-F filed 2026-05-26 covering the fiscal year ended 2026-03-31 (813KB of text), and two 6-Ks of 691 bytes (2026-05-06) and 692 bytes (2026-07-29). Both 6-Ks are cover pages and nothing else. Each consists of the registrant name, commission file number 001-41800, the Form 20-F/40-F checkbox, an "EXHIBIT INDEX" heading and a "SIGNATURES" heading. Neither contains a single financial figure. The 2026-05-06 stub is tagged to the period 2026-01-01 to 2026-03-31 (the FY2026 fourth quarter and full year) and the 2026-07-29 stub to 2026-04-01 to 2026-06-30 (Q1 FY2027). The earnings releases those 6-Ks furnished are in absent exhibits. This is the known extraction limit stated in the data contract, and on ARM it bites harder than anywhere else in the batch, because with no 10-Q there is no other filing that could cover an interim period.

Consequence, stated plainly so no reader mistakes it: every June-2026-quarter figure in this dive — revenue $1,289M, operating income $98M, net income $270M, GAAP diluted EPS $0.25 — comes from the VENDOR payload, cross-checked against the earnings-calendar entry for 2026-07-29. The most recent filing-verified financial statements in this dive are as of 2026-03-31.

The business itself

Arm describes itself in the 20-F as "a global leader in the semiconductor industry" whose "principal operations and activities are the licensing, marketing, research and development of central processing unit ('CPU') IP, graphics processing unit IP, systems IP, compute subsystems ('CSS'), and associated software, tools and other related services."

Chief executive Rene Haas. 9,584 employees as of 2026-03-31 across 19 countries, of whom approximately 84% are engaged in engineering activities — plus an average of 2,820 temporary employees during FY2026. Approximately 8,600 issued patents and 2,750 applications pending. Founded 1990 as a joint venture of Acorn Computers, Apple Computer and VLSI Technology; listed in London and on Nasdaq from 1998 until SoftBank Group took it private in 2016; re-listed via ADSs in September 2023 at $51 per ADS, in an offering from which Arm itself received no proceeds (it was a secondary sale of 102.5M ADSs by a shareholder).

How Arm gets paid — two lines, and the split matters

The 20-F disaggregates revenue into exactly two categories for major product offerings, and the vendor's seg_prod block reproduces them faithfully:

The reconciliation is exact in all three years — $2,613M + $2,307M = $4,920M, matching the 20-F's stated "Total revenue increased $913 million, or 23%, to $4,920 million" to the dollar. This is one of the cleanest vendor-to-filing segment ties in the batch and it is worth saying so.

Royalty is the annuity. Per the 20-F: "Royalties are generally either set as a percentage of the licensee's average selling price per chip or as a fixed amount per chip... Royalty revenue is recognized on an accrual basis in the quarter in which the customers ship chips containing our products." Critically, it is accrued on estimates: "using estimates from sales trends and judgment for several key attributes, including industry estimates of expected shipments, the mix of products sold, the percentage of markets using our products, and average selling price." The company then trues up in the following quarter. That is a revenue line built on management judgement about other companies' shipments, and the 20-F says so.

Licence is the lumpy line. "Over the term of a license, contractual payments can generally range from hundreds of thousands of dollars to hundreds of millions of dollars" and "We generate a significant proportion of our Licensing and other revenue from a relatively small number of customers." The FY2026 increase is attributed to "fluctuation in timing and size of multiple high-value license agreements and contributions from backlog into the current period from arrangements entered in prior periods" — which is the company telling you the licence line is timing-driven and should not be extrapolated quarter to quarter.

Two structured licence products sit alongside the bespoke agreements: Arm Flexible Access (annual portfolio fee plus a tape-out fee, aimed at smaller companies) and Arm Total Access (a periodic portfolio fee with manufacturing rights included and no tape-out fee, aimed at large established semiconductor companies). And Architecture Licence Agreements, under which a licensee designs its own Arm-compatible CPU for a fixed fee — the arrangement at the centre of the Qualcomm/Nuvia litigation described in Section 8.

Compute Subsystems and the March 2026 pivot into silicon

The growth mechanic the 20-F names most explicitly is moving up the value stack. Compute Subsystems (CSS) are "Arm's CPU, GPU, and System IP products integrated into a foundational compute platform optimized for a specific end market. These CSSs are pre-integrated and pre-verified configurations of Arm technology that deliver significantly higher value to customers by reducing development costs and time-to-market." The value-capture logic follows in the trend section: "For chips where our products have provided more value, we will typically receive higher royalty revenue per chip."

And in March 2026 the company crossed a line it had never crossed in thirty-six years. The 20-F states: "In March 2026, the Company announced it is expanding its compute platform into production silicon products with the Arm AGI CPU" — Arm's own branded chip. It also states, twice, that this "did not have a material impact to our revenue for the fiscal year ended March 31, 2026."

Read the associated risk factor, because it is the honest counterweight to the enthusiasm. The 20-F warns that production silicon "may subject us to new or enhanced competitive, brand, technological, regulatory, operational and financial risks" and, in the trend section, that "our expansion into production silicon may have materially different margin profiles, revenue recognition characteristics, and sales cycles compared to our IP licensing business." That is management stating, in its own filing, that the 95% gross margin is not guaranteed to survive the pivot. Half-weighted as management framing, but it points against the company's own interest, which makes it more credible rather than less.

Geography — and a trap in how it is measured

The vendor's seg_geo block ties exactly to reported revenue in all three years ($4,920M / $4,007M / $3,233M), which is a clean pass. But the 20-F attaches a caveat that changes the meaning: "Revenue by geographic region is allocated to individual countries based on the principal headquarters of the customers. The geographical locations are not necessarily indicative of the country in which the customer sells products containing the Company's technology IP."

Two things to flag. First, Japan appears as a separate line for the first time in FY2026 at $825M, 16.8% of revenue; in FY2025 and FY2024 it sat inside "Other countries", which is why that line falls from $589M to $373M. The year-on-year "Other countries" comparison is therefore broken and no conclusion is drawn from it. Second — and this is the one that matters — Japan at 16.8% of revenue is not a coincidence. SoftBank Group is headquartered in Tokyo, and revenue from entities under common control with SoftBank is attributed to the customer's headquarters. The geographic table and the related-party disclosure in Section 2 are describing overlapping money.

The 20-F also states directly that "During the fiscal years ended March 31, 2026 and 2025, revenue from sales to customers outside of the U.S. accounted for approximately 64% and approximately 57% of total revenue, respectively" — matching the vendor table to within a point (100% − 35.8% = 64.2%).

Customer concentration — high, but improving

From the 20-F's segment note, verbatim structure:

The largest customer has fallen from 21% (FY2024) to 17% (FY2025) to 16% (FY2026). That is the single concentration metric in this dive moving in the right direction, and it deserves to be said in a document that is otherwise hard on the revenue quality.

Competition, per the filing

The 20-F does not publish a named competitor list in the manner a 10-K would. What it does name, repeatedly and by name, is RISC-V — an open-source instruction set architecture — as the ecosystem customers and partners might be pushed toward, and it identifies the PRC's "15th Five-Year Plan and related initiatives" emphasis on semiconductor self-sufficiency as a policy vector that "could encourage financing opportunities to our competitors in the PRC on favorable terms, or influence major PRC customers to favor adoption of IP of our competitors in the PRC over our own IP." The knowledge base independently carries both the bull and bear sides of exactly this question; see Section 7.

2. The single most important thing to understand about these numbers: 74% of last year's growth came from related parties

Everything else in this dive is subordinate to one paragraph in the 20-F's management discussion, and the vendor payload does not carry it in any form.

Here is the sentence, verbatim:

> "Revenue from external customers increased $237 million, or 7%, during the fiscal year ended March 31, 2026, as compared to the fiscal year ended March 31, 2025, driven by a $360 million, or 20%, increase in royalty revenue, offset by a $123 million, or 9%, decrease in license and other revenue. Revenue from related parties increased $676 million, or 82%, during the fiscal year ended March 31, 2026 as compared to the fiscal year ended March 31, 2025, primarily driven by a $591 million, or 141%, increase in license and other revenue and a $85 million, or 21%, increase in royalty revenue."

And here is who "related parties" means, also verbatim:

> "Revenue from related parties is derived from Arm China, customers in which we have an equity method investment, and other entities related to us by virtue of common control by SoftBank Group."

The percentages in that paragraph are sufficient to rebuild the split exactly, and we did. (Numeric tables are stripped from the current filing extraction, so this is derived from the filing's own prose percentages and the reported totals — every figure below reconciles to the audited $4,920M and $4,007M.)

Three conclusions follow, and they are the whole of the analytical case in this dive.

First: 74.0% of the year's revenue growth came from related parties. $676M of $913M. The headline "+23%" that every screen prints, and that the forward multiple is being paid for, decomposes into +7.4% from arms-length customers and the rest from inside the group.

Second: external licence revenue is DECLINING. Down 8.7%, from $1,420M to $1,297M. This matters far more than its size suggests, because of a timeline the 20-F states in its own words: "it can take two to three years to design a new Arm processor, and it can take another two to three years to develop a chip." Licences are the leading indicator of royalties four-to-six years out. A declining external licence line while the reported total grows 23% is exactly the shape of a company whose forward royalty pipeline is being flattered in the present by a related-party licence transaction.

Third: the related-party growth is overwhelmingly LICENCE, not royalty. Related-party licence revenue rose 141%, from $419M to $1,010M — a $591M increase inside a single year. Royalties, which require chips to actually ship, rose only 21% on the related-party side. Licence revenue is recognised on delivery of a design and is negotiated between counterparties; royalty revenue requires the physical world to cooperate. The half of the related-party number that grew fastest is the half that is easiest to time and to price.

None of this is an allegation. The 20-F discloses it, an audit committee approves related-party transactions under a written policy, and the auditors expressed an unqualified opinion. The point is narrower and it is a valuation point: a dollar of revenue negotiated between two entities under common control with an 86.4% shareholder deserves a lower multiple than a dollar won from an unrelated customer in a competitive process, and at 125.9x forward earnings the market is applying no such discount at all.

The Arm China layer makes it structural rather than incidental. From the trend section: "We utilize our commercial relationship with Arm China to access the PRC market for IP revenue, and substantially all of our PRC-related revenue is generated through the IPLA with Arm China, a related party... Our revenue is calculated as a percentage of license and royalty fees earned by Arm China from sub-license arrangements entered into with its end customers." And critically: "Despite our reliance on Arm China through our commercial relationship with it, both as a source of revenue and a conduit to the important PRC market, Arm China operates independently of us. Under the IPLA, Arm China's payments due to us are determined based on the financial information that Arm China provides to us."

Arm neither controls nor consolidates the entity through which substantially all of its China revenue flows, and depends on that entity to self-report the numbers. The risk factor is blunter still: "In the past, we have had issues obtaining timely and accurate information from Arm China, as well as enforcing our right to conduct an audit. We believe the underlying problems causing our past inability to obtain such information have been resolved, but we can provide no assurances..."

3. Growth and margin trajectory

Annual, from the vendor income series, cross-checked to the 20-F's stated totals:

Five-year revenue compound growth: 19.4%. Steady, high, and not obviously cyclical — which is the structural argument for Arm over the equipment names in this batch: royalties accrue on chips shipped, not on fab capital budgets.

Two things in that table need explaining before any of it is usable.

The FY2024 operating margin of 3.1% is real and it is share-based compensation. Research and development jumped from $1,080M to $1,932M in a single year — 40.3% to 59.8% of revenue — because the September-2023 IPO converted a large block of cash-settled awards into equity-settled ones and recognised the cost. The cash-flow statement shows share-based compensation going from $79M (FY2023) to $1,037M (FY2024). FY2024 is an accounting event, not an operating collapse, and the four-year operating-margin series should be read as roughly 25% → event → 21% → 18.5%, i.e. compressing, not recovering.

The FY2026 gross margin of 92.5% is rejected. The vendor's annual costOfRevenue of $370M is contradicted three ways: by the sum of its own four FY2026 quarters ($264M of cost of sales, a 94.6% margin); by its own ratios_ttm.grossProfitMarginTTM of 95.35%; and by the 20-F's explicit statement that "Cost of sales remained flat during the fiscal year ended March 31, 2026, as compared to the fiscal year ended March 31, 2025" — a year in which the vendor itself reports cost of sales of $206M. The filing wins. We use a gross margin of approximately 94.6% for FY2026 and treat the annual 92.5% as a vendor artefact (detailed in §8, finding 6). Operating income of $908M is unaffected and is consistent across both the annual and quarterly series to within $7M.

Quarterly, cross-checked against the earnings calendar:

Read the operating-margin column, not the revenue column. Arm's fiscal fourth quarter is structurally the big one — licence agreements land there — and the first quarter is structurally the weak one. But Q1 FY2027's 7.6% GAAP operating margin is the second-lowest in nine quarters, and it is not a revenue problem: revenue grew 22.4% year on year. Research and development jumped from $698M to $838M sequentially, a $140M step in one quarter, annualising to roughly $3.35B against FY2026's $2,776M. The 20-F attributes the FY2026 R&D increase to "investments in next generation products, such as the Arm AGI CPU", and the June quarter looks like that spend accelerating.

And note how the June quarter's net income was made. Operating income $98M; pre-tax income $253M; net income $270M — the net figure exceeds pre-tax because of a $17M tax BENEFIT. So $155M of the $253M of pre-tax income, 61%, came from below the operating line: interest income, equity-investment marks and other non-operating items. A quarter in which most of the profit is non-operating and the tax line is negative is not a quarter you capitalise at 126x.

Five consecutive non-GAAP beats, all modest to large. Q1 FY2026 +0.3%, Q2 +18.6%, Q3 +4.9%, Q4 +3.4%, Q1 FY2027 +12.5% on EPS, with revenue beats of −0.8%, +7.0%, +1.2%, +1.1% and +2.1%. And the market sold the last one anyway — the ADSs closed 2026-08-04 at $280.56 against a 50-day average of $323.94, so the six-day window since the 2026-07-29 print has been a de-rating. We cannot see why, because the 6-K is a cover page and the guidance is in an absent exhibit. That gap is stated rather than filled with speculation.

Research and development is the defining expense line and it is rising as a share of revenue. $2,776M in FY2026, 56.4% of revenue, against 50.1% in FY2025 and 40.3% in FY2023 (pre-IPO, pre-equity-conversion). The 20-F frames it as deliberate: "each year we increase our research and development investment in line with the increased development needs of the next generation of products." Selling, general and administrative went the other way — 24.1% to 17.6% of revenue — so total operating leverage is roughly flat and the operating margin decline is entirely an R&D decision. That is a defensible decision for a company building its first silicon product. It is also the reason the GAAP earnings the price is measured against are as small as they are.

4. Balance sheet and the enterprise-value rebuild

As at 2026-03-31, from the vendor payload and reconciled against the 20-F. This IS a filing-verified date — the 20-F covers exactly this balance sheet.

Correction 1 — netDebt omits short-term investments. The vendor's −$2,294M is total "debt" of $457M less cash and equivalents of $2,751M only. It ignores $850M of short-term investments. The vendor reports net cash of $2,294M; the balance sheet shows $3,144M; we use $3,144M.

Correction 2 — and this one is larger than it looks: Arm has NO BORROWINGS AT ALL. The vendor's totalDebt of $457M is entirely the capitalLeaseObligations field; longTermDebt is $0 in all three years. The 20-F puts the lease position beyond doubt: "As of March 31, 2026, we had operating and finance lease payment obligations of $549 million and $62 million, respectively, with $50 million and $26 million, respectively, payable within twelve months of March 31, 2026."

Finance-lease obligations are $62M of undiscounted payments. The vendor's $457M is therefore predominantly OPERATING lease liabilities folded into a field labelled "capital lease obligations" and then into "total debt." Operating lease liabilities are not borrowings and never have been. This is the lease double-counting defect class, and the vendor's own file contradicts itself on it: debtToMarketCapTTM is 0, longTermDebtToCapitalRatioTTM is 0, interestCoverageRatioTTM is 0 and debtServiceCoverageRatioTTM is 0 — all consistent with a company that has no borrowings, and all inconsistent with the $457M of "total debt" sitting three fields away.

There is no revolving credit facility, no senior notes, no aggregate principal amount and no maturity schedule disclosed anywhere in the 20-F. Searches on "revolving credit", "aggregate principal", "senior notes" and "indebtedness" return only a table-of-contents heading and the SoftBank margin facility, which is SoftBank's borrowing, not Arm's.

The rebuilt enterprise value:

> Market cap $299.638B (1,068,078,760 ordinary shares/ADSs × $280.56)

> + lease liabilities $0.457B (conservatively treated as debt)

> − cash and equivalents $2.751B

> − short-term investments $0.850B

> = corrected enterprise value $296.494B

The vendor reports $297.044B. The overstatement is $550M, or 0.19% — trivial in relative terms because equity value swamps everything at this valuation, but the vendor's implied net cash of $2,594M matches no balance sheet in the file (the 2026-03-31 sheet gives $3,144M with leases, $3,601M without). We cannot reconcile the vendor's EV to any statement in the payload, most likely because it was struck on a June-2026 balance sheet that bal_a does not expose. We use our own build and say so.

If you exclude lease liabilities entirely — which is the technically correct treatment, since operating leases are not debt — net cash is $3,601M and enterprise value is $296.037B. Every EV multiple in this dive uses the conservative $296.494B; the difference is 0.15% and changes nothing.

Correction 3 — trailing EBITDA is overstated by the vendor and the correction makes the multiple WORSE. ratios_ttm.ebitdaMarginTTM of 26.629% on TTM revenue of $5,156M implies EBITDA of $1,373M, which is exactly the figure evToEBITDATTM of 216.35x resolves to. But building it from the four quarters to 2026-06-30 gives operating income of $892M plus depreciation and amortisation of $264M = $1,156M, a 22.4% margin. The gap is $217M and it runs in the company's favour in the vendor file. We use $1,156M, and EV/EBITDA is therefore 256.5x, not 216.3x.

What the corrected balance sheet actually says. Arm has $3,601M of cash and short-term investments, no borrowings, a current ratio of 5.25x and $10,703M of total assets against $2,417M of total liabilities. It is, financially, unbreakable. There is no scenario in this dive in which the balance sheet is the problem — and that is precisely why the risk score of 8 is driven by valuation, float structure, ownership and revenue composition rather than by leverage.

Two balance-sheet items that are worth naming.

Receivables are $2,466M — 50.1% of annual revenue, on 141.6 days of sales outstanding. For a company with no inventory and no payables cycle to speak of, the entire cash conversion cycle is 141.6 days of receivables. Receivables grew 29.9% year on year against 22.8% revenue growth, so collection is lengthening. This is the balance-sheet footprint of licence agreements invoiced on milestone schedules and of related-party balances, and it is the line to watch if the related-party revenue question ever becomes a collection question.

Purchase obligations of approximately $1,057M with a remaining term of twelve months or longer through 2036, for "cloud computing web services, data centers, software, license and services" — against $4,920M of revenue. That is the cost of the compute needed to design chips, contractually committed.

5. Cash flow and capital returns — and the number that matters most

This is the table that should change how the multiple is read, and it is worth being slow about it.

Free cash flow of $979M in FY2026 looks respectable — 19.9% of revenue, on a $299.6B market capitalisation, which is 306x. But that $979M is arrived at after adding back $1,052M of share-based compensation as a non-cash item. Stock compensation is not cash out of the door in the period, but it is unambiguously a cost: it is how Arm paid a workforce that is 84% engineers, and the shares it issues to do so are shares the existing holder does not own. Subtract it, and free cash flow was NEGATIVE $73M in FY2026, NEGATIVE $642M in FY2025 and NEGATIVE $90M in FY2024. Three consecutive years.

Share-based compensation was 21.4% of FY2026 revenue (and 20.5% in FY2025). For scale: it is larger than the entire company's operating income of $908M, and equal to 51.9% of Arm's entire FY2021 revenue of $2,027M. This is not a rounding item to be discussed in a footnote; it is the second-largest economic cost in the business after engineers' cash salaries, and it is the difference between the GAAP EPS of $0.846 and the non-GAAP EPS of $1.77 that the entire street multiple is struck on.

The dilution is real but modest, which is the one mitigating fact. Diluted shares went from 1,026.1M (FY2021-FY2023, pre-IPO), to 1,044.5M (FY2024), 1,063M (FY2025), 1,068M (FY2026) and 1,078M in the June-2026 quarter — roughly 1.0-1.4% a year. The reason the dilution is smaller than the compensation expense implies is disclosed in the cash-flow discussion: FY2026 financing outflows rose $409M because of "payments of withholding taxes on vested share-based awards resulting from the shift to a full withhold-to-cover method in satisfaction of tax obligations." Arm is buying back the tax-withholding shares with cash instead of selling them into the market. That is better for the share count and worse for the cash flow, and it is the honest explanation of why operating cash flow and dilution both look better than the compensation number would suggest.

Capital returns: there are none, and the filing is explicit. No buyback programme is disclosed anywhere in the 20-F. On dividends: "We intend to retain any earnings for use in our business and do not currently intend to pay dividends on our ordinary shares or ADSs", and in the risk factors: "capital appreciation, if any, on our ADSs will be your sole source of gains for the foreseeable future." lastDividend: 0, dividendYieldTTM: 0 and dividendPayoutRatioTTM: 0 in the vendor file are all correct, not zeroed fields. Total shareholder yield is 0.00% and every dollar of return must come from the multiple or from earnings growth. Holders of ADSs additionally bear depositary service fees, which the 20-F discloses the depositary is entitled to charge — a small, permanent, negative carry that a direct holder of ordinary shares does not pay.

Capital expenditure is rising sharply: $93M → $143M → $219M → $545M, which the 20-F attributes to "data center and office expansions and computer hardware purchases." At 11.1% of revenue it is no longer negligible for a company whose entire pitch is that it is asset-light. Note the internal inconsistency in the vendor's own cash-flow record: investmentsInPropertyPlantAndEquipment is −$575M while capitalExpenditure is −$545M and free cash flow is computed on the $545M. We use $545M and flag the $30M discrepancy (§8, finding 11).

6. Valuation — priced in or room?

At $280.56 (market cap $299.638B on 1,068,078,760 shares, corrected enterprise value $296.494B, net cash $3,144M):

Estimate coverage is good where it matters and thin where it does not. FY2028 carries 21 revenue and 22 EPS analysts — the best-covered forward year in the file and the one our fair value anchors on. FY2027 has 19 and 16. FY2030 (10 analysts) and FY2031 (12 revenue / 19 EPS) are EXCLUDED from every conclusion in this dive, because they show revenue growth accelerating to +46.3% and +76.1% five and six years out, with non-monotonic analyst counts — a fabrication signature, not a forecast (§8, finding 8).

Per the data contract, est.ebitdaAvg and est.ebitAvg are not used, and on ARM they fail spectacularly. Every one of the six forward years carries an EBITDA margin of exactly 25.9% and an EBIT margin of exactly 20.4% — a constant-margin signature, the same class of defect the data contract records on other names — while netIncomeAvg is 1.87x to 1.95x ebitAvg in every single year. Net income cannot be double EBIT year after year on any consistent basis. What has happened is that netIncomeAvg and epsAvg are struck on a NON-GAAP basis that adds back share-based compensation, while ebitAvg/ebitdaAvg are on something closer to GAAP. The rows are internally mixed-basis and are discarded.

epsAvg and revenueAvg, by contrast, verify cleanly against FY2026 actuals: revenueAvg of $4,903.2M against the reported $4,920M (−0.34%), and epsAvg of $1.75281 against the four reported non-GAAP quarters summing to $1.77 (−0.97%). This confirms two things at once: the estimate EPS row is reliable, and it is NON-GAAP. All forward valuation here runs on those two rows.

The GAAP-versus-non-GAAP gap on this name is the largest in the batch and it is not a technicality.

The $986M difference is almost exactly the $1,052M of share-based compensation, less tax. Consensus, the price targets, and every forward multiple you will see quoted anywhere are struck on the $1.77 number. We show both and use non-GAAP for forward multiples, with the GAAP figure stated alongside, because that is the basis consensus is struck on and comparing our fair value to a differently-based street number would be dishonest. But the reader should hold one fact steady: the "earnings" in 125.9x forward earnings are earnings that exclude a fifth of revenue in employee compensation.

Peer context. The vendor peer set is largely unusable for this purpose: Applied Materials, Lam Research, KLA and Intel are capital-equipment and manufacturing businesses with entirely different economics; Amphenol, Intuit, ServiceNow and Sony are not comparables at all. The only structurally relevant peer in the set is Qualcomm ($170.8B), which is both a licensee and a litigation counterparty, and no peer multiples are supplied in the file, so no peer-multiple comparison is drawn. What can be said from the batch table is context: KLAC trades at 35.6x FY2027 consensus, TXN and the equipment complex in a similar band, and ARM at 125.9x is roughly 3.5x the forward multiple of the most expensive equipment name in the batch.

6a. What today's price assumes (the inversion)

At $280.56 — 125.9x FY2027 consensus non-GAAP EPS and 92.6x FY2028 — the price embeds the following falsifiable claims, each with a number and a date:

• Revenue rises 23.0% in the twelve months to 2027-03-31, to $6,053M, and then ACCELERATES to +35.1% in FY2028, to $8,180M. (Consensus-derived; 19 and 21 analysts.) The acceleration is the fragile part, not the level. Arm has grown 22.8%, 23.9% and 20.7% in the last three fiscal years — a remarkably stable low-twenties. Consensus needs that to step up by twelve percentage points in the second year. Nearest check: the 2026-11-04 print, where consensus revenue of $1,376.8M would be +21.3% year on year — i.e. still in the low twenties, so FY2028's acceleration is entirely back-loaded and untested.
• Non-GAAP EPS reaches $3.031 by FY2028, up 71.2% from FY2026's $1.77, which requires net margin to expand from roughly 38.4% to 39.6% on the consensus basis while revenue grows 66%. (Consensus-derived; 22 analysts.) This is the most fragile assumption in the price, and the reason is in Section 3: GAAP operating margin was 7.6% in the most recent quarter on a research-and-development run rate of $838M, annualising 21% above FY2026's total. Arm is spending into a silicon product that the filing says contributed nothing to FY2026 revenue and may carry "materially different margin profiles." Consensus needs margin expansion out of the exact period in which Arm has told you it is investing hardest.
• The revenue growth is arms-length. (Our number, derived from the 20-F's own prose.) External revenue grew 7.4% in FY2026 while total revenue grew 22.8%. For the FY2028 consensus of $8,180M to be a market outcome rather than a group outcome, external revenue must roughly double from $3,420M in three years while related-party revenue holds flat. If instead the related-party share keeps rising from 30.5%, the multiple should compress regardless of whether the revenue arrives.
• The market keeps paying roughly 90x or better on the two-year-forward year at every date between now and then. (Our number.) At 70x FY2028E the ADSs are $212; at 60x they are $182; at today's 92.6x they are $281. In the last twelve months this stock has traded at $104.55 and at $439.46 — a band of 34x to 145x FY2028E. The multiple, not the earnings, is by far the widest source of variance in the outcome, and there is no anchor for it.
• The 13.6% free float and the 86.4% SoftBank stake remain undisturbed. (Filing-confirmed structure; our assumption on the direction.) 769,029,000 shares — 72.0% of the company — are pledged as collateral on the SoftBank Group Facility, and the 20-F states that in a default "the providers of the SoftBank Group Facility may... exercise their rights to foreclose on and sell or cause the sale of our shares." The filing also flags that "future exercise of registration rights by SoftBank Group may materially and adversely affect the market price of the ADSs." The price assumes the float stays scarce. Any change to that assumption is a supply event of a size the float cannot absorb.

6b. The return bridge (why the multiple moves)

Expected return over the next twelve months decomposes as: EPS growth (+36.0%, from FY2027E $2.229 to FY2028E $3.031) + multiple drift (COMPRESSION, from 92.6x to roughly 75x on the respective forward year, −19.0%) + shareholder yield (0.00% — there is no dividend and no buyback) ≈ +10% at best, and negative on our own base case.

Our base case assumes multiple COMPRESSION, and it is not a rolldown artefact. Today the market pays 92.6x on the fiscal year ending 2028-03-31. Our base pays 78x on that same fiscal year. That is a genuine 15.8% de-rating on an identical earnings line, not the mechanical effect of the calendar advancing, and we assume it for three stated reasons: first, the composition of growth — 74.0% of FY2026's increment came from related parties, and multiples for related-party-dependent revenue are structurally lower; second, the basis — a multiple on non-GAAP earnings that exclude 21.4% of revenue in stock compensation should carry a discount to a multiple on clean earnings, and today's does not; third, growth maturation on the consensus path itself — EPS +36% in FY2028, +33% in FY2029, and the market normally begins discounting deceleration a year ahead of it.

There is no shareholder yield leg at all. Unlike every other name in this batch, 100% of the return must come from earnings growth and the multiple. No dividend, no buyback, and roughly 1.0-1.4% of annual dilution running the other way. The return bridge for ARM is therefore: EPS growth minus dilution, times whatever the multiple does.

And the multiple is doing most of the work in every scenario. If the multiple simply HELD at 92.6x on FY2028E the price would be $281 — exactly spot, no return at all, after a 36% earnings increase. That is the arithmetic that makes this an Avoid rather than a Watch: at this multiple, delivering the consensus earnings ramp in full produces zero return. If it compressed to 60x the price would be $182 (−35%); at 105x it would be $318 (+13%). The entire spread between a good and a bad outcome sits in a multiple we do not control, have no edge on, and which has ranged from 34x to 145x on the same estimate line within twelve months.

The bull case at $322 does not require multiple expansion at all — it holds today's 92.6x and applies it to the FY2028 consensus high of $3.477. That is a deliberately conservative construction: the bull case is an earnings beat, not a re-rating. The bear case at $155 is the mirror image: a de-rating to 55x applied to the FY2028 consensus low of $2.813.

6c. Variant perception (where we differ, what would surprise)

• We differ from the street on the composition of growth, and this is the whole edge in this dive. The street's $302.14 consensus target, and every forward multiple quoted anywhere, treats FY2026's +22.8% as one number. The 20-F says it is two numbers: +7.4% external and +82% related-party, with external LICENCE revenue down 8.7%. We have found no evidence that the consensus targets discount for it. Watchable number: the related-party share of revenue in the FY2027 20-F, due around May 2027. If it rises above roughly 35%, the growth story is materially a group-internal story and the multiple should halve. If it falls back toward 20% with total growth intact, we are wrong and the thesis inverts.
• We differ on the honest earnings basis. Consensus is non-GAAP, and on ARM the non-GAAP adjustment is essentially one item: $1,052M of share-based compensation, 21.4% of revenue. Free cash flow less stock compensation has been negative for three straight years (−$90M, −$642M, −$73M). Nobody's published multiple reflects that, and it is the difference between "a 20% free-cash-flow margin" and "a business that has not yet generated surplus cash after paying its people." Watchable number: share-based compensation as a percentage of revenue in the FY2027 accounts. Below 15% would materially change our view; above 22% confirms it.
• We differ from the knowledge base on price, not on the business, and the knowledge base says so itself. The single most valuation-specific claim in the lane — from business_breakdowns, dated 2025-03-02, conviction 70 — argues that "ARM's 30x-sales valuation is justified by more than AI." The ADSs trade at 57.5x sales today. We agree with the qualitative case in that claim and reject the price: the bull case our own knowledge base articulates was struck at roughly half the current sales multiple, and no claim in the lane defends the multiple that now exists.
• Positive surprise that would force a re-rating up: the Arm AGI CPU — the production-silicon product announced in March 2026 that contributed nothing to FY2026 revenue — winning a named hyperscaler design at scale. The knowledge base's freshest name-level claim, from Dylan Patel, 2026-04-07 (conviction 55), predicted exactly this: "ARM will launch a standalone CPU in a few weeks that Meta and Cloudflare will adopt, driving more standalone-ARM proliferation rather than IP licensing." The 20-F independently confirms the launch half of that prediction; the adoption half is untested and is the largest unpriced variable in the story. A silicon product carries a far larger revenue-per-socket than a royalty, and if it lands, the FY2029-FY2031 revenue path that currently looks fabricated becomes arguable.
• Negative surprise that would break the thesis: any disclosure that the Arm China relationship has deteriorated (the 20-F already records past failures to obtain timely information and to enforce audit rights); a step-up in the related-party share of revenue; a guide implying FY2028 revenue below roughly $7.0B; gross margin falling below 90% as production silicon ships; or any event touching the SoftBank Group Facility against which 72.0% of the company is pledged. Any of these hits a stock at 125.9x forward earnings with a 13.6% float and a beta of 3.767 — the worst possible combination of valuation and liquidity to be holding through bad news.

Synthos fair values

All three anchors are multiples of the FY2028 consensus non-GAAP EPS distribution (mean $3.031, low $2.813, high $3.477, 22 analysts — the best-covered forward year in the file), cross-checked against FY2027E of $2.229 and FY2029E of $4.033.

• Bear ~$155 — 55x the FY2028 consensus LOW of $2.813. Cross-check: 69.6x FY2027E; 38.4x FY2029E; still 48% above the 52-week low of $104.55 and 18.7% below the 200-day average of $190.78. The scenario: the FY2028 acceleration does not arrive, related-party revenue is disclosed as a still-larger share, the silicon pivot compresses gross margin, and the multiple de-rates toward something a related-party-dependent licensor deserves. −44.8%.
• Base ~$236 — 78x the FY2028 consensus MEAN of $3.031. Cross-check: 106.1x FY2027E; 58.6x FY2029E; corrected EV of $249.4B on FY2028E revenue of $8,180M is 30.5x sales — which is, deliberately, the same sales multiple the knowledge base's own bull claim defended. Sensitivity, stated openly: 70x gives $212, 85x gives $258, and 92.6x — today's multiple, simply held — gives $281, which is spot. The entire answer lives in a 70-93x band and no greater precision is claimed. The scenario: consensus revenue and EPS land roughly on plan, and the market applies a modest discount for the related-party composition and the non-GAAP basis. −15.7%, with zero shareholder yield to soften it.
• Bull ~$322 — today's 92.6x forward multiple, HELD, applied to the FY2028 consensus HIGH of $3.477. Cross-check: 144.5x FY2027E. This bull case requires no re-rating, only an earnings beat — the Arm AGI CPU wins volume sockets, Armv9 and Compute Subsystems lift royalty per chip faster than modelled, external licensing turns back up, and data-centre attach compounds. +14.8%.

Base is 15.7% BELOW spot; asymmetry is roughly 0.33:1 to the upside — 44.8% of downside against 14.8% of upside. That ratio is the verdict. It is not that the business is bad; it is that on the most generous defensible construction of the bull case, using consensus's own high estimate and no multiple expansion whatsoever, the reward is a third of the risk. And that offer is being made on a day the ADSs rose 17.36%.

7. Knowledge base — the deepest lane in the batch, and its own bull case says the price is wrong

Raw hits: 25. Entity matches: 22. Text-only candidates: 3. Used: 20 (18 entity + 2 text-only). Discarded: 5 (4 entity + 1 text-only). Discarded for token collision on the bare word "ARM": ZERO.

The collision check first, because it was the specific hazard here. "ARM" is a high-collision token — arm of a company, armed, arms race, arm's length. The three text-only candidates were inspected individually and not one is a false positive: all three genuinely reference the Arm architecture. That is an unusually clean result for this token and it is reported as a check that passed, not assumed.

What was discarded, and why:

The 20 used claims break down by stance as 17 bullish, 2 neutral, 1 bearish. By channel: business_breakdowns 11, jensen_huang 1 (used), dylan_patel 3, all_in 2, lex_fridman 1, bill_gurley 1, invest_like_the_best 0 (used) — seven distinct channels, hence a breadth score of 7.

Three sourcing weaknesses cap the conviction, and they must be stated before the quotes.

First: 16 of the 20 used claims carry speaker: null and are attributed only to a CHANNEL. Under the 4-lane attribution policy that is the weakest form of sourcing we accept — a channel-level attribution tells you where a claim was said, not who said it or what their record is. Only four used claims have a named speaker: Dylan Patel, "Guido", Brad Gerstner and Daniel Marcus.

Second, and more serious: 10 of the 22 entity matches — 45% of the entire lane — share a single date, 2025-03-02, on a single channel (business_breakdowns), all with no named speaker. That is not ten independent observers converging; it is one episode transcribed into ten claims. The apparent depth of this lane is roughly half illusory, and the effective independent-source count is closer to six than to twenty-two. This is exactly the concentration failure the per-speaker attribution policy exists to surface, and it is surfaced here rather than buried.

Third: no claim in the lane carries a speaker_role of management. The half-weight management discount therefore applies nowhere. One claim is about the chief executive ("CEO Rene Haas reinvigorated a 'sleepy' ARM post-IPO", conviction 70) but is not by him, so it is treated as third-party assessment, not company voice. On position reversals: the dylan_patel channel carries one bullish and one bearish claim, but from two DIFFERENT named speakers (Dylan Patel bullish 2026-04-07; "Guido" bearish 2025-09-22). Attributed by speaker rather than channel, that is disagreement between two people, not one person reversing. No reversals are visible anywhere in the lane.

The claims that carry real weight, verbatim.

> 2025-03-02 · bullish · conviction 85 · horizon: principle · entities: ARM Holdings · channel: business_breakdowns · no named speaker

> "ARM effectively holds a monopoly on chip ISA for mobile phones after x86 and legacy ISAs exited that market."

> 2025-03-02 · bullish · conviction 80 · horizon: thesis · entities: ARM Holdings, Amazon, Microsoft, Google, Alibaba · channel: business_breakdowns · no named speaker

> "Data center is ARM's biggest growth opportunity — hyperscalers designing own CPUs default to ARM since there's no real alternative."

> 2025-03-02 · bullish · conviction 75 · horizon: principle · entities: ARM Holdings · channel: business_breakdowns · no named speaker

> "ARM's mature software-compatibility layer across data-center chips is the critical moat, creating high switching costs beyond the hardware IP."

> 2025-03-02 · bullish · conviction 70 · horizon: thesis · entities: ARM Holdings, ARM · channel: business_breakdowns · no named speaker

> "ARM's 30x-sales valuation is justified by more than AI — broad expansion into data center, automotive, and IoT beyond its mobile monopoly."

> 2025-03-02 · bullish · conviction 65 · horizon: fact · entities: ARM Holdings · channel: business_breakdowns · no named speaker

> "ARM's royalty per chip is rising from ~7 cents (2023) to ~9 cents today to perhaps 12 cents in 3-4 years, expanding value capture."

> 2025-03-02 · neutral · conviction 60 · horizon: thesis · entities: RISC-V, ARM Holdings · channel: business_breakdowns · no named speaker

> "RISC-V is an emergent potential competitor with more flexibility, but very early days — not a dire threat to ARM anytime soon."

> 2025-09-30 · bullish · conviction 85 · horizon: thesis · entities: ARM · channel: all_in · no named speaker

> "ARM is the CPU link between hardware and software, uniquely positioned as AI compute moves to energy-efficient edge, wearables and robots."

> 2026-04-07 · bullish · conviction 55 · horizon: fact · entities: ARM · channel: dylan_patel · speaker: Dylan Patel (independent)

> "ARM will launch a standalone CPU in a few weeks that Meta and Cloudflare will adopt, driving more standalone-ARM proliferation rather than IP licensing."

> Test: "ARM ships its standalone CPU with Meta/Cloudflare adoption"

> 2025-09-22 · BEARISH · conviction 55 · horizon: thesis · entities: ARM · channel: dylan_patel · speaker: Guido (independent)

> "ARM is a little screwed too; its selling point was partnering with everyone who wouldn't partner with Intel, but Nvidia now has access to Intel technologies."

> 2024-10-13 · bullish · conviction 58 · horizon: thesis · entities: ARM · channel: bill_gurley · speaker: Brad Gerstner (independent)

> "Arm is well positioned as an orthogonal competitor to peel off AI compute at the edge, where Nvidia's advantage is weaker, off its ~300B installed device base."

And the two text-only claims that survived inspection, labelled as such:

> 2025-10-14 · bullish · conviction 88 · horizon: principle · entities: Nvidia (ARM named in the thesis text only) · channel: jensen_huang · no named speaker

> "New computing platforms almost never emerge—only x86, ARM, and CUDA exist; the CUDA ecosystem flywheel is a rare, durable moat."

> Used at reduced weight: the claim's subject is another company's moat, but the premise — that Arm is one of only three durable computing platforms in existence — is a direct, high-conviction statement about Arm's structural position and corroborates the ISA-monopoly claim above from an independent channel.

> 2022-03-22 · bullish (on RISC-V, therefore BEARISH on Arm) · conviction 55 · horizon: thesis · entities: none (ARM named in thesis text only) · channel: dylan_patel · speaker: Daniel Marcus (policymaker)

> "Chinese firms like Alibaba are actively adopting open-source RISC-V to de-link from proprietary ARM/Intel/AMD instruction sets and have had some successes, with military companies among the participants."

> Used at reduced weight and aged four years, but it is the only claim in the lane that speaks to the specific threat the 20-F's own risk factors name most often, and it comes from a policymaker rather than an investor.

What this lane is, stated without inflation.

The qualitative case for Arm is strongly and independently corroborated. Four separate channels, across four years, converge on the same structural claim: Arm is an architectural monopoly whose moat is the compiled software estate rather than the IP, hyperscalers designing their own CPUs default to it, and the royalty take per chip is rising. The 20-F corroborates each of those from the company's own side — 350 billion cumulative chips, 22 million developers, and "For chips where our products have provided more value, we will typically receive higher royalty revenue per chip." On the business, we do not differ from the knowledge base at all.

One knowledge-base prediction has been partially validated by the filing since it was made. Dylan Patel's 2026-04-07 claim that Arm would launch a standalone CPU is confirmed by the 20-F's disclosure that "In March 2026, the Company announced it is expanding its compute platform into production silicon products with the Arm AGI CPU." The adoption half of the claim — Meta and Cloudflare — is untested and no filing in this archive speaks to it. That is a live, gradeable test with a named speaker attached, and it is the most useful single item in this lane.

One claim we tried to verify and could not. The royalty-per-chip claim (7 cents → 9 cents → 12 cents) is checkable in principle from royalty revenue divided by annual chip units. The 20-F discloses cumulative units (over 350 billion) but the annual unit figure sits in a numeric table that the current filing extraction strips. We state that as an extraction limit rather than estimating around it. What can be said directionally: royalty revenue grew 20.5% in FY2026 and the filing attributes that specifically to "an improved mix of products with higher royalty rates per chip, such as Armv9 technology" — consistent with the claim, not proof of the numbers.

And here is the finding that matters for the verdict. The single most valuation-specific claim in the entire lane defends a 30x-sales multiple. Arm trades at 57.5x sales today. The knowledge base's own bull case was struck at roughly half the current price-to-sales multiple, and not one of the 20 used claims defends the multiple that now exists. The lane supports the business; it does not support the price. The conviction rating is Moderate on breadth and weak on sourcing, and it does not earn a valuation premium in the fair value.

8. Data integrity — what we rejected from the vendor file and why

Fourteen findings. Listed rather than silently corrected. ARM's vendor file is unusual: the segment and geographic tables are among the cleanest in the batch and tie exactly to the filing, while the derived metrics, the lease classification and the estimate block are among the worst.

1. Currency — TESTED AND PASSED, and this was the live risk on this name. profile.currency is USD, every reportedCurrency in inc_a, inc_q, bal_a, cf_a, seg_prod and seg_geo is USD, and the quote is in USD on Nasdaq. The 20-F confirms it in its own words at the front of the document: "We prepare our consolidated financial statements in accordance with generally accepted accounting principles in the U.S. ('GAAP'). We present our consolidated financial statements in U.S. dollars ('USD')." A UK-incorporated issuer reporting in US GAAP and US dollars is the exception, not the rule, and the check had to be run before any multiple was computed. The 20-F adds a second confirmation from the revenue side: "Less than 2% of our total revenue is denominated in currencies other than U.S. dollars, and the impact of changes in foreign exchange rates on our revenue for the fiscal years ended March 31, 2026 and 2025 was immaterial." No currency mixing. Every multiple in this dive is USD-on-USD. Stated because the check passed, not assumed.

2. ADS-to-ordinary-share ratio and share count — TESTED AND PASSED. This is the defect class that has broken other names in this batch, and on a foreign issuer with ADSs it had to be resolved from the filing rather than inferred.

• The ratio is 1:1. The 20-F states it directly: "American depositary shares, each of which represents the right to receive one ordinary share ('ADSs')." Not 1:4, not 1:10. One ADS is one ordinary share.
• Market cap ÷ price = $299,637,678,519 ÷ $280.56 = 1,067,998,569 shares.
• The 20-F's principal-shareholders section states the count as of 2026-05-21: "1,068,078,760 ordinary shares outstanding." Difference: 0.0075%. Agreement.
• The 20-F COVER PAGE states a different, earlier count: "The number of ordinary shares outstanding of Arm Holdings plc as of March 31, 2026 was 1,064,055,252." That is 0.37% below the implied count — the gap is seven weeks of share issuance under the equity plans, not an error. We use 1,068,078,760.
• Cross-check against the EPS denominator: the FY2026 weighted-average DILUTED share count is 1,068,000,000 — within 0.0001% of the market-cap-implied count. The Q1 FY2027 diluted count is 1,078,000,000, 0.94% higher, consistent with continued RSU vesting.
• Independent fourth check: the insider Form 4s in the payload all record the security as "Ordinary Shares", consistent with a 1:1 deposit ratio.

Four independent counts land within 1% of each other. There is no share-class error, no ADS-ratio error and no partial-unit error on this name. The vendor's market capitalisation is correctly struck. Stated at length because the price level and the ADS structure demanded the check, and the check passed.

3. The FREE FLOAT is 13.6%, not 100% — not a vendor defect, but the most important structural fact the payload does not carry. The 20-F: "As of May 21, 2026, SoftBank Group beneficially owns approximately 922,733,999 or 86.4% of our total issued and outstanding share capital and thus a majority of the total voting power of our ordinary shares. As such, our publicly traded ADSs, representing ordinary shares, is 145,344,760." The tradeable market value is approximately $40.8B, not $299.6B. Further: 769,029,000 ordinary shares — 72.0% of the company — are pledged as collateral under the SoftBank Group Facility, a margin loan, and the filing states that on default the lenders "may... exercise their rights to foreclose on and sell or cause the sale of our shares that may be pledged as collateral. The foreclosure on our shares that are initially pledged as collateral for the SoftBank Group Facility could cause a change of control of us." No vendor field carries any of this. It is reported here because a 17.36% single-day move across a 13.6% float on below-average volume cannot be interpreted without it.

4. epsdiluted is None in EVERY row of inc_a and inc_q — a systematic absence, not an FY2026 hole — and diluted EPS is derived. All six annual rows and all nine quarterly rows carry epsdiluted: None. The differently-cased parallel field epsDiluted is populated ($0.85 for FY2026), so this is a field-naming artefact rather than missing data, but it is a null in the payload and it is flagged rather than silently passed over. Our derivation, stated: FY2026 net income of $904,000,000 ÷ the FY2026 weighted-average diluted share count of 1,068,000,000 = $0.8464, which rounds to the $0.85 in the populated field. Confirming cross-check: the four FY2026 quarterly GAAP EPS figures are $0.12 + $0.22 + $0.21 + $0.29 = $0.84, consistent to a cent of rounding. We use $0.846 for FY2026 GAAP diluted EPS, and $0.975 for the trailing twelve months to 2026-06-30 ($1,044M of net income over an average diluted count of 1,071M).

5. GAAP versus non-GAAP is the largest basis gap in this batch and consensus is struck on the non-GAAP side. FY2026 GAAP diluted EPS $0.846; the four reported non-GAAP quarters sum to $1.77 — a gap of 109.1%, or $986M of net income. The FY2026 est.epsAvg of $1.75281 lands within 1.0% of the $1.77 non-GAAP actual and 107% above the GAAP figure, which is the arithmetic proof that the estimate block is non-GAAP. The adjustment is almost entirely share-based compensation of $1,052M, 21.4% of revenue. All forward multiples in this dive use consensus non-GAAP EPS and the trailing multiple is shown on both bases (150.0x non-GAAP, 287.8x GAAP). The reader is told which is which everywhere it appears.

6. inc_a FY2026 cost of revenue of $370M is contradicted three ways — the annual gross margin of 92.5% is REJECTED. The vendor's annual figure implies a 92.5% gross margin. Against it: (a) the sum of its own four FY2026 quarterly cost-of-sales figures is $264M, a 94.6% margin; (b) its own ratios_ttm.grossProfitMarginTTM is 95.35%, implying $240M of cost on TTM revenue of $5,156M; (c) the 20-F states "Cost of sales remained flat during the fiscal year ended March 31, 2026, as compared to the fiscal year ended March 31, 2025" — a year in which the vendor's own annual cost of sales was $206M. The vendor reports a 92.5% FY2026 gross margin; the quarterly series, the ratios block and the filing all point to 94.6-95.8%; we use approximately 94.6% and reject the 92.5%. Operating income of $908M is unaffected (the annual and quarterly series agree to within $7M on that line), so no valuation conclusion in this dive rests on the disputed figure.

7. netDebt omits $850M of short-term investments — REJECTED and rebuilt. The vendor's −$2,294M nets the $457M "debt" against cash and equivalents of $2,751M only. The vendor reports net cash of $2,294M; the balance sheet shows $3,144M; we use $3,144M. Consequential rejections: enterpriseValueTTM ($297.044B → $296.494B), netDebtToEBITDATTM (−1.89x → −2.72x), evToSalesTTM (57.61x → 57.50x).

8. totalDebt of $457M is entirely LEASE liabilities and predominantly OPERATING leases — REJECTED as a debt measure. longTermDebt is $0 in all three years; totalDebt equals capitalLeaseObligations exactly. The 20-F says finance-lease payment obligations are $62M and operating-lease payment obligations are $549M, so the field labelled "capital lease obligations" is carrying operating leases. Operating lease liabilities are not borrowings. The vendor's own file contradicts itself: debtToMarketCapTTM, longTermDebtToCapitalRatioTTM, interestCoverageRatioTTM and debtServiceCoverageRatioTTM are all 0, consistent with zero borrowings, three fields away from a $457M "total debt". Arm has no notes, no revolver and no aggregate principal amount disclosed anywhere in the 20-F. We treat borrowings as $0 and, conservatively, still deduct the $457M of lease liabilities in the EV build — a 0.15% difference either way. Note also the year-to-year inconsistency: shortTermDebt reads $32M (FY2024), $40M (FY2025) and $0 (FY2026), so the FY2025-to-FY2026 "total debt" comparison in the vendor file is not meaningful.

9. Trailing EBITDA is overstated in the vendor file by $217M — REJECTED and recomputed. ratios_ttm.ebitdaMarginTTM of 26.629% on TTM revenue of $5,156M implies $1,373M, and evToEBITDATTM of 216.35x resolves to exactly that figure. Building it from the four quarters to 2026-06-30 gives operating income $892M + depreciation and amortisation $264M = $1,156M (22.4% margin). We use $1,156M, and every EV/EBITDA figure in this dive is computed on it: 256.5x, not 216.3x. The correction makes the stock look more expensive, not less.

10. The est block's EBIT and EBITDA rows carry a constant-margin fabrication signature — REJECTED in full. Across all six forward years the EBITDA margin is exactly 25.9% and the EBIT margin is exactly 20.4%, while netIncomeAvg sits at 1.87x to 1.95x ebitAvg in every single year. Net income cannot be roughly double EBIT year after year on a consistent basis. What has happened is a basis mix inside a single row: netIncomeAvg/epsAvg are non-GAAP (verified against the FY2026 actual to within 1.0%) while ebitAvg/ebitdaAvg are on something nearer GAAP (FY2026 ebitAvg of $1,001.9M is 10.3% above the actual GAAP operating income of $908M). ebitAvg and ebitdaAvg are used for nothing in this dive. All forward valuation runs on epsAvg and revenueAvg, which verify cleanly.

11. The FY2030 and FY2031 estimate rows are EXCLUDED from every conclusion. They show revenue growth accelerating to +46.3% and +76.1% five and six years out, taking FY2031 revenue to $28,090M — 5.7x FY2026 — with analyst counts that move non-monotonically (10 → 12 on revenue, 10 → 19 on EPS). A forecast in which growth accelerates every year to the end of the series is not a forecast. Separately, the FY2027 row is internally inconsistent: its netIncomeAvg of $2,318.7M divided by its epsAvg of $2.22858 implies 1,040M shares, against 1,068-1,071M implied by every other row. We use the FY2027 epsAvg and ignore its netIncomeAvg.

12. Two smaller vendor inconsistencies, corrected in presentation. (a) cf_a FY2026 reports investmentsInPropertyPlantAndEquipment of −$575M and capitalExpenditure of −$545M in the same row, with free cash flow computed on the $545M. We use $545M and note the $30M gap. (b) The ratios_ttm per-share ratios do not all resolve to the same price: priceToBookRatioTTM × bookValuePerShareTTM gives $280.56 (the close), priceToSalesRatioTTM × revenuePerShareTTM gives $281.09, and priceToEarningsRatioTTM × netIncomePerShareTTM gives $283.26 — which is the day's HIGH of $283.2485. Three different prices inside one block. Every price-based ratio in this dive is recomputed at the $280.56 close.

13. quote.yearHigh/yearLow disagree with tech.hi52/lo52 — we use tech and say why. The quote block reports a 52-week high of $452.70 and low of $100.02; the computed technical block reports $439.46 and $104.55 — discrepancies of 3.0% and 4.5%, the widest in the batch. We use the tech figures throughout, because they are computed from the same six-year close series that produces the moving averages, drawdown and relative-return figures, and mixing sources within the technical section would make the percentages inconsistent. Note also tech.last of $280.52 against quote.price of $280.56 — a four-cent difference from different snapshot moments, immaterial, and not reported as a defect.

14. Fields that read as zeroed but are CORRECT, checked individually. inventory: 0 — Arm licenses designs and holds no physical stock; the 20-F confirms no inventory line. commonStockRepurchased: 0 and commonDividendsPaid: 0 in all four years — the 20-F states Arm pays no dividend and discloses no repurchase programme. lastDividend: 0, dividendYieldTTM: 0, dividendPayoutRatioTTM: 0 — all correct. daysOfPayablesOutstandingTTM: 0 and inventoryTurnoverTTM: 0 — artefacts of a business with no inventory and negligible payables ($80M), not errors. One genuine zeroed field: interestExpense is $0 in all six annual rows, while the 20-F states "Interest expense consists primarily of interest on finance leases" — so a small interest expense exists and the field is zeroed. Immaterial to every conclusion here, and reported for completeness.

Not defects, correctly reported and independently confirmed: the seg_prod royalty/licence split ties EXACTLY to reported revenue in all three years ($2,613M + $2,307M = $4,920M, matching the 20-F's stated "Total revenue increased $913 million, or 23%, to $4,920 million"); the seg_geo table ties EXACTLY to reported revenue in all three years and the 64% non-US share it implies matches the 20-F's stated "approximately 64%" to the point; the FY2026 revenue growth rates of +25% licence and +21% royalty match the 20-F's "increased $468 million, or 25%" and "increased $445 million, or 21%" to the dollar; the share count reconciles four ways; the currency reconciles three ways; and stockBasedCompensationToRevenueTTM of 22.38% is consistent with $1,052M on FY2026 revenue and the higher TTM figure. On the segment and geography lines specifically, this is one of the two or three cleanest vendor-to-filing reconciliations in the batch.

Balance-sheet freshness — tested, with one real gap. bal_a[0] is dated 2026-03-31, which is the exact date of the audited 20-F balance sheet, so it is filing-verified and not stale in the GOOGL/LLY sense. But the income statement runs one quarter further (Q1 FY2027 to 2026-06-30) with no corresponding balance sheet, and the vendor's own enterpriseValueTTM implies net cash of $2,594M, which matches no balance sheet in the file — evidence it was struck on an unexposed June-2026 sheet. On financings post-dating the statements: both 6-Ks are cover pages, Arm has no borrowings, no notes, no revolver and no shelf drawdown disclosed, so there is no financing event that could have moved the balance sheet. That half of the test passes cleanly.

Non-equity tripwire — checked and passed. ARM is common equity in the form of American Depositary Shares, each representing one ordinary share of nominal value £0.001, Nasdaq Global Select listed, CUSIP 042068205, ISIN US0420682058. The price of $280.56 is not par-like ($25 or $1,000); beta is 3.767, the highest in the batch and the opposite of the sub-0.3 typical of a fixed-income-like instrument; there is no dividend at all, fixed or variable, which rules out a preferred or a baby bond from the other direction; the 52-week band of $104.55 to $439.46 is a 320% range; and the insider Form 4s record the security as "Ordinary Shares". This is common equity, held through a depositary. The one ADS-specific wrinkle worth stating is that the depositary is entitled to charge holders annual ADS service fees, which a direct holder of the ordinary shares does not pay — a small permanent negative carry disclosed in the 20-F.

9. Technicals

• Price $280.56. −36.17% from the 52-week high of $439.46, but +168.31% above the 52-week low of $104.55. Position within the annual range: 52.6th percentile — the middle, despite the drawdown, because the low was so low.
• Moving averages, split decision: −13.4% below a 50-day average of $323.94; +47.1% above a 200-day average of $190.78. The long-term uptrend is powerfully intact; the intermediate trend is broken.
• RSI 50.97 — exactly neutral, which is unusual for a stock 36% off its high and offers no oversold cushion. MACD −19.31 — negative.
• Maximum drawdown from peak over the trailing period: −36.17%, i.e. before today's move the price was at the worst point of the drawdown.
• Beta 3.767 — the highest in the batch by a wide margin, and the single most useful number in this section. A 4-to-5% sector factor mechanically produces a 15-to-19% move in this name.
• Relative performance is extraordinary and it cuts both ways: 3-month +38.0% vs SPY +7.6% and QQQ +7.7%; 6-month +162.3% vs SPY +11.1% and QQQ +15.7%; 12-month +103.9% vs SPY +24.3% and QQQ +30.8%. ARM has outperformed the S&P by roughly 80 points over twelve months while sitting 36% below its own high. The six-month figure — +162% — is the one to hold: this stock has already had its re-rating, inside the last two quarters.
• Sentiment: 19 buy, 6 hold, 2 sell across 27 analysts; consensus target $302.14 (+7.7%), median $270 — 3.8% BELOW spot, high $500 (+78.2%), low $130 (−53.7%). The high-to-low spread is 3.8x, the widest dispersion in the batch. The consensus mean sits 12% above the median, meaning a small number of very high targets are pulling the average up. A target set in which the median is below the current price is a genuine signal and it agrees with our direction.

Today's move and what it does to the entry

ARM closed 2026-08-04 at $280.56, up 17.36% or $41.50 from a previous close of $239.06. It opened at $253.40 — a 6.00% gap — traded a $251.10 to $283.25 range, and closed within 0.95% of the day's high on 9,421,874 shares.

Three facts settle the question of what caused it.

First, there is no company-specific event. No filing in the archive is dated 2026-08-04. The most recent company event is the 6-K of 2026-07-29 — Q1 FY2027 results — which is 692 bytes of cover page with no financial statements attached. Six days elapsed between that print and today's move.

Second, 2026-08-04 was a broad semiconductor up-day and ARM's beta is 3.767. Across the complex: MRVL +12.81%, INTC +10.92%, LRCX +7.85%, MU +7.62%, AMD +7.00%, KLAC +6.95%, AMAT +5.48%, TXN +5.42%, ASML +4.22%, with PANW +5.53%, IBM +3.91% and ANET +3.06% elsewhere in the batch. ARM's +17.36% is the largest move of any name recorded on the day, and it is 1.36x the next largest. At a beta of 3.767, a market-plus-sector factor of roughly 4.6% produces exactly this move. This is amplified beta, not information.

Third — and this is the detail that should stop a buyer — volume was BELOW average. 9,421,874 shares against a 90-day average of 10,327,972, i.e. 8.8% below normal, on a day the price rose 17.36%. And the free float is 145,344,760 ADSs, so the day turned over 6.5% of the tradeable stock. A 17% advance on below-average volume in a 13.6% float is the signature of scarcity, not of accumulation. There were simply not enough shares available to absorb a sector-wide bid.

The honest read. Unlike KLA in this same batch, where a 7% gap came off a 35% drawdown into resistance and did not by itself disqualify the entry, ARM's gap does disqualify it — because ARM's valuation had no cushion before the gap and has less now. Before today the ADSs were 125.9x FY2027 consensus at $239.06; after today they are 147.7x at $280.56 on the same estimates. The gap did not resolve any uncertainty. It repriced the same uncertainty 17% higher in one session on light volume.

The chart itself is genuinely two-sided and we will not pretend otherwise. The 200-day average at $190.78 is 32.0% below spot and rising; the 50-day at $323.94 is 15.5% above and falling; RSI is dead neutral at 50.97. This is a stock in the middle of a very wide range with no technical edge in either direction. What it is NOT is a cheap entry.

What we are therefore telling the reader, plainly: we are not asking you to buy this at all today, and today's 17.36% move makes that easier rather than harder to say. The setups that would change it are the pre-registered flip conditions in Section 11, and the nearest of them is a price near $190-210 — which is 25-32% below today's close and inside the range this stock has traversed within the last six months.

10. Insiders — eight transactions, all of them OPEN-MARKET SALES, and none of them tax withholding

This table is the opposite of the KLA pattern in the same batch and must not be read the same way.

Total: 82,094 shares, $21.43M at an average price of $261.04.

Every transaction is coded S-Sale — an open-market disposition. There is not one F-InKind transaction in this file, which is the code for shares withheld by the issuer to satisfy tax on vesting RSUs. These are not tax withholdings. Four named executive officers sold shares into the market and none of them bought any.

Four observations, stated carefully.

First, two officers went to zero. Chief Legal Officer Spencer Collins sold 40,941 shares — half the total by count — and holds zero afterward. Chief People Officer Charlotte Eaton also finished at zero. A complete disposal by a named executive officer is a materially different signal from a partial trim, and it is reported as such. The payload does not carry Rule 10b5-1 plan flags, so we cannot tell whether these were pre-scheduled. We say so rather than assuming either way.

Second, the Chief Accounting Officer sold nearest the top. Bartels sold 11,306 shares at $392.70 on 2026-06-02 — 40% above today's price and within 11% of the 52-week high of $439.46. She retained 12,135 shares, so she sold roughly 48% of her position.

Third, the price spread across nineteen days is itself informative about the volatility of this name. The eight sales run from $215.00 (2026-05-19) to $392.70 (2026-06-02) — an 83% range in fourteen calendar days. That is what a 3.767-beta stock with a 13.6% float does, and it is a warning to anyone sizing a position on the assumption that they can exit near any particular price.

Fourth, and important for what this signal is NOT: the data stops on 2026-06-04, two months before this dive. Nothing in the insider file covers the July decline or the 2026-07-29 print. We do not know what officers have done since, and we do not speculate.

The reading. Zero open-market purchases and eight open-market sales by four officers, two of whom exited entirely, at an average price 7% below today's close. That is a mildly negative signal — mildly, because executive selling has many innocent explanations and because the amounts ($21.4M) are trivial against a $299.6B market capitalisation. But it is a genuinely different fact from KLA's eight F-InKind withholdings in this same batch, and conflating the two would be an error. It is reported as selling, because it is selling.

11. Verdict, kill-criteria and flip conditions

Avoid.

This is the best franchise in this batch and the worst price in it, and the verdict is entirely about the second thing.

What is genuinely excellent, and none of it is in dispute: an architectural monopoly on the instruction set that essentially all smartphones and a rising share of data-centre CPUs compile against; more than 350 billion cumulative chips shipped and more than 22 million developers, a moat made of compiled software rather than of patents; a gross margin near 95%; revenue up 22.8% to $4,920M in FY2026 with royalty revenue up 20.5% on the strength of Armv9 mix; $3,144M of net cash and not one dollar of borrowings; a customer concentration that is improving (largest customer 21% → 17% → 16%); a genuine expansion up the value stack through Compute Subsystems and, from March 2026, into production silicon with the Arm AGI CPU; and the deepest, most bullish knowledge-base lane in this batch, corroborated point-for-point by the filing.

What we are declining to pay for: 287.8x trailing GAAP earnings. 150.0x trailing non-GAAP. 125.9x the FY2027 consensus. 92.6x FY2028. 57.5x sales. 256.5x corrected trailing EBITDA. 306x free cash flow. A base fair value of $236 that is 15.7% BELOW spot and 21.7% below the street's $302.14. Zero shareholder yield — no dividend, no buyback, and roughly 1.0-1.4% of annual dilution running the other way, so 100% of the return has to come from earnings and the multiple. Free cash flow less share-based compensation that has been negative for three consecutive years. And an entry price that includes a 17.36% one-day move on below-average volume with no company news.

And what genuinely worries us, which is neither the price nor the business: 74.0% of the FY2026 revenue increase came from RELATED PARTIES — Arm China and entities under common control with SoftBank Group — while external licence revenue FELL 8.7%; related-party revenue rose from 20.6% to 30.5% of the company in one year; substantially all PRC revenue flows through an entity Arm neither controls nor consolidates and which self-reports the numbers; SoftBank Group owns 86.4% and has pledged 72.0% of the company against a margin loan whose foreclosure the filing says could cause a change of control; the free float is 13.6%; and as a foreign private issuer Arm files no quarterly financial statements at all, so between 20-Fs there is nothing to check any of it against.

The distinction that matters. We are not saying Arm is a bad business — we think it is probably the best business in this batch, and the exponential score of 8 says so. We are saying that at 125.9x forward earnings, with three-quarters of last year's growth sourced from the controlling shareholder's affiliates, a 13.6% float, no yield, and a base case 15.7% below the current price, there is no version of this position that pays you for the risk you are taking. If the multiple simply HOLDS at today's 92.6x and consensus earnings arrive in full through FY2028, the return is zero. That is not a Watch; a Watch implies we would like to own it at a modestly better price. This needs a materially different price, and it is an Avoid until it gets one.

Pre-registered KILL criteria — what would confirm Avoid and take this to a permanent exclusion:

• Related-party revenue exceeding roughly 35% of total in the FY2027 20-F, or another year in which related parties supply the majority of the revenue increment. That is the growth story being revealed as a group-internal story.
• External licence and other revenue declining for a second consecutive year. It fell 8.7% in FY2026; because the filing says a processor takes two to three years to design and a chip another two to three to build, a second decline is a direct hit to royalty revenue four-to-six years out.
• Any credit event, refinancing difficulty or foreclosure notice touching the SoftBank Group Facility, against which 769,029,000 shares — 72.0% of the company — are pledged.
• Any disclosure that the Arm China relationship has deteriorated, that timely or accurate information is again unobtainable, or that payments are again late. The 20-F records all three as having happened before.
• Gross margin falling below 90% as production silicon ships, which would confirm the filing's own warning that the AGI CPU carries "materially different margin profiles."
• A guide implying FY2028 revenue below roughly $7.0B — 14% under the current consensus of $8,180M — which removes the acceleration the entire multiple is paid for.

Pre-registered FLIP conditions — what would take this to Watch, and then to Buy — Tactical:

• Price near $190-210 with no fundamental deterioration. This is the single most likely path and it is only 25-32% below today's close, inside the range this stock has traversed within six months. At $200 our base of $236 is an 18% premium, the bear case is a 22% drawdown rather than 45%, and the asymmetry inverts to better than 1:1. This would be a Watch at $210 and a Buy — Tactical at $190.
• A 20-F or 6-K showing EXTERNAL revenue growth above roughly 20% with the related-party share falling toward 20%. That single disclosure would remove the largest analytical objection in this dive and would justify a materially higher multiple than we currently apply. This is the fastest possible route to being wrong, and we would want to be told about it.
• The Arm AGI CPU winning a disclosed, named, volume design at a hyperscaler, validating the untested half of Dylan Patel's 2026-04-07 knowledge-base claim. Production silicon carries far more revenue per socket than a royalty, and if it lands at scale the FY2029-onward revenue path becomes arguable rather than fabricated.
• Share-based compensation falling below roughly 15% of revenue, which would close most of the gap between the GAAP and non-GAAP earnings the multiple is struck on, and would make free cash flow less stock compensation positive for the first time since FY2023.
• A meaningful reduction in SoftBank's stake through an orderly secondary that widens the float — painful in the short run because of the supply, and structurally positive because it replaces a controlled-company discount and a 13.6% float with a normal shareholder register.

Where ARM fits in the Synthos Framework Portfolio. The semiconductor architecture and intellectual-property sleeve, at 0% today with a 1.5-2.5% target on a fill near $190-210 or on a disclosed improvement in external revenue growth. On the batch overlap question: ARM is not the same trade as KLAC, LRCX, AMAT or ASML — those are capital-equipment names levered to fab spending budgets, while Arm is levered to chip units shipped and to royalty rate per unit, which is a different and structurally better exposure. A reader who owns the equipment complex is not already long Arm. That is precisely why this is worth tracking rather than dismissing: the exposure is genuinely differentiated and genuinely exponential, and the only thing wrong with it is the price. Logged as a tracked Synthos call (Avoid) as of 2026-08-04 at $280.56.

Single biggest risk: that the reported growth is not an arms-length growth rate. $676M of the $913M FY2026 revenue increase — 74.0% — came from related parties inside a group SoftBank Group controls at 86.4%, and external licence revenue fell 8.7%. The entire valuation architecture — the 125.9x forward multiple, the 35.1% FY2028 revenue acceleration, the $302.14 street target — is built on a growth rate that is one-quarter arms-length. There is no valuation cushion to absorb the re-rating that a second year of the same composition would deserve. At 125.9x forward earnings and 57.5x sales, a disappointment does not produce a proportional decline; it produces a de-rating on top of it, which is what the $155 bear case is.

Most fragile assumption in the price: that non-GAAP EPS reaches $3.031 by FY2028, a 71.2% increase from FY2026's $1.77, while the company invests hardest. Every other assumption in 6a is a volume assumption and volume is observable quarter by quarter. This one is a margin assumption made against the direction of the company's own most recent disclosed quarter — GAAP operating margin of 7.6% in June 2026 against 29.5% in March, on research and development that jumped $140M sequentially to $838M, annualising 21% above the whole of FY2026 — and against the filing's own warning that production silicon "may have materially different margin profiles, revenue recognition characteristics, and sales cycles compared to our IP licensing business." Arm is being asked to expand margins in precisely the years it has told you it is building its first chip. If FY2028 non-GAAP EPS lands at $2.60 rather than $3.031, then at 78x the ADSs are worth $203 — 28% below today's close.

Provenance & disclosures

• Traceability: 25 tagged knowledge-base claims name Arm (entity matches 22, text-only candidates 3, used 20, discarded 5; breadth 7, net conviction positive-moderate). Sources by channel: business_breakdowns (11 used, of which 10 share a single date, 2025-03-02 — a single episode, and 45% of the whole lane), dylan_patel (3 used: Dylan Patel bullish 2026-04-07, "Guido" bearish 2025-09-22, Daniel Marcus text-only 2022-03-22), all_in (2, both 2025-09-30, no named speaker), jensen_huang (1 used text-only, 2025-10-14; three others discarded), lex_fridman (1, 2021-02-18), bill_gurley (1, Brad Gerstner 2024-10-13). Sixteen of the twenty used claims carry speaker: null and are attributed only to a channel — the weakest sourcing the 4-lane attribution policy admits, and it is 80% of this lane. No claim carries a management speaker_role, so the half-weight management discount applies nowhere. No position reversals are visible; the one bullish/bearish pair on the dylan_patel channel comes from two different named speakers, which is disagreement rather than reversal. Discards, with reasons: two jensen_huang claims from 2013-01-16 (Windows RT on ARM; Project Denver) discarded for a thirteen-year staleness that removes all evidentiary value; one invest_like_the_best claim (2025-09-16) on private-equity returns in which ARM is an incidental example; one jensen_huang claim (2025-11-12) whose subject is the United Kingdom's research base; and one text-only jensen_huang claim (2025-03-13) whose subject is another company's Grace Hopper product. Collision check on the bare token "ARM" — arm of a company, armed, arms race, arm's length — returned ZERO false positives across all 25 hits, an unusually clean result for that token, and it is reported as a check that passed rather than assumed. The most valuation-specific claim in the lane defends a 30x-SALES multiple; the ADSs trade at 57.5x sales, so the knowledge base's own bull case does not support the current price, and this is stated as a finding. All quotes are verbatim from the stored claim text.
• Data as-of: fundamentals — income statement, balance sheet and cash flow for fiscal 2026 ended 2026-03-31, filing-verified against the 20-F filed 2026-05-26; the June-2026 quarter (Q1 FY2027) is VENDOR-SOURCED ONLY and cross-checked against the earnings-calendar entry for 2026-07-29 (revenue actual $1,289M against $1,262.683M estimated; non-GAAP EPS $0.45 against $0.40) · estimates 2026-08-04 · prices 2026-08-04, quote timestamp 1785873601 = 2026-08-04T20:00:01Z ($280.56, +17.36%; 50-DMA $323.94; 200-DMA $190.78; RSI 50.97; MACD −19.31; beta 3.767) · knowledge-base claims 2026-08-04. Fiscal/calendar note: Arm's fiscal year ends 31 MARCH. "FY2026" in this dive means the twelve months ended 2026-03-31; "Q4 FY2026" means the March 2026 quarter; "Q1 FY2027" means the June 2026 quarter. All figures come from the Synthos vendor data file for ARM or from the SEC filings in the ARM archive; no figure comes from memory, recall or external retrieval.
• Filing archive contents and its gaps — this is a FOREIGN PRIVATE ISSUER and the archive is unlike every other name in the batch: Form 20-F filed 2026-05-26 covering fiscal 2026 ended 2026-03-31 (813KB); Form 6-K filed 2026-05-06 (691 bytes) tagged to the period 2026-01-01 to 2026-03-31; Form 6-K filed 2026-07-29 (692 bytes) tagged to 2026-04-01 to 2026-06-30. THERE IS NO 10-K AND NO 10-Q — Arm does not file them and never has. BOTH 6-Ks were opened in full and BOTH are cover pages only: registrant name, commission file number 001-41800, the Form 20-F/40-F checkbox, an "EXHIBIT INDEX" heading and a "SIGNATURES" heading. Neither contains a single financial figure; the earnings releases they furnished sit in exhibits that were not extracted. The 2026-07-29 6-K is the Q1 FY2027 earnings furnish, filed six days before this dive, and it tells us nothing. As a foreign private issuer Arm is exempt from quarterly reporting — the 20-F states "we are not required to file periodic reports and financial statements with the SEC as frequently or as promptly as U.S. public companies" — so no filing covers the June 2026 quarter and none will. The most recent filing-verified financials in this dive are as of 2026-03-31. Separately, numeric tables are stripped from the current filing text extraction, so the related-party revenue split in Section 2 is derived from the 20-F's own prose percentages and reconciled to the audited $4,920M and $4,007M totals, and the annual chip-unit figure needed to test the royalty-per-chip claim is unavailable. Both limits are stated rather than estimated around.
• Where the filings contradicted or corrected the vendor (detailed in Section 8): totalDebt of $457M rejected as a debt measure — the field is entirely capitalLeaseObligations and the 20-F says finance-lease obligations are only $62M against $549M of operating-lease obligations, so operating leases have been folded into "total debt"; Arm has ZERO borrowings, which the vendor's own debtToMarketCapTTM, interestCoverageRatioTTM and longTermDebtToCapitalRatioTTM of 0 independently confirm three fields away. Net cash rebuilt from $2,294M to $3,144M (the vendor omits $850M of short-term investments), enterprise value rebuilt from $297.044B to $296.494B, and every EV multiple recomputed. Trailing EBITDA recomputed from the vendor's $1,373M to $1,156M (operating income $892M plus D&A $264M over the four quarters to 2026-06-30), taking EV/EBITDA from 216.3x to 256.5x — the correction makes the stock more expensive, not less. inc_a FY2026 cost of revenue of $370M and the 92.5% gross margin it implies REJECTED against the vendor's own quarterly sum ($264M, 94.6%), its own ratios_ttm (95.35%) and the 20-F's statement that "Cost of sales remained flat" on a FY2025 base of $206M. est.ebitAvg and est.ebitdaAvg rejected in full for a constant-margin signature (exactly 25.9% and 20.4% in every one of six years) and for netIncomeAvg running at 1.87-1.95x ebitAvg in every year — a basis mix inside a single row. FY2030 and FY2031 estimate rows excluded for revenue growth accelerating to +46.3% and +76.1% with non-monotonic analyst counts; the FY2027 row's netIncomeAvg discarded for implying 1,040M shares against 1,068M elsewhere. quote.yearHigh/yearLow ($452.70/$100.02) rejected in favour of tech.hi52/lo52 ($439.46/$104.55), a 3.0% and 4.5% discrepancy. ratios_ttm per-share ratios found to resolve to three different prices ($280.56, $281.09 and $283.26, the last being the day's high); all price-based ratios recomputed at the $280.56 close. epsdiluted is None in all fifteen income-statement rows and diluted EPS was derived ($904M ÷ 1,068M = $0.846, cross-checked against the four quarterly figures summing to $0.84). And the entire related-party revenue analysis in Section 2 — the most important finding in this dive — exists ONLY in the 20-F's prose and appears nowhere in the vendor payload in any form. Where vendor and filing AGREED — worth recording, because the agreements here are unusually strong: the currency (USD presentation and US GAAP, confirmed verbatim in the 20-F and matching profile.currency and every reportedCurrency field); the ADS ratio of 1:1 ("each of which represents the right to receive one ordinary share") and the share count, which reconciles four independent ways to within 1% (market cap ÷ price 1,067,998,569; 20-F principal-shareholders section 1,068,078,760 at 2026-05-21; FY2026 diluted denominator 1,068,000,000; 20-F cover 1,064,055,252 at 2026-03-31); the seg_prod royalty/licence split, which ties exactly to $4,920M / $4,007M / $3,233M in all three years and whose +25% and +21% growth rates match the 20-F's prose to the dollar; the seg_geo table, which also ties exactly in all three years and whose implied 64.2% non-US share matches the 20-F's stated "approximately 64%"; the absence of any dividend or buyback, confirmed by the 20-F's "we... do not currently intend to pay dividends"; and the zero-inventory balance sheet, correct for an IP licensor.
• Basis note — the most important disclosure in this dive after the related-party finding: analyst estimates and every consensus figure quoted here are NON-GAAP; reported diluted EPS is GAAP. The gap is 109.1%. FY2026 GAAP diluted EPS was $0.846 against a reported non-GAAP total of $1.77, a difference of $986M of net income, of which essentially all is $1,052M of share-based compensation equal to 21.4% of revenue. Forward multiples in this dive use consensus non-GAAP EPS (125.9x FY2027E, 92.6x FY2028E) because that is the basis consensus is struck on, and the trailing multiple is shown on both bases (150.0x non-GAAP, 287.8x GAAP). The reader should hold one fact steady: free cash flow LESS share-based compensation was negative in FY2024 (−$90M), FY2025 (−$642M) and FY2026 (−$73M).
• Estimate coverage: 19 analysts on FY2027 revenue and 16 on FY2027 EPS; 21 and 22 on FY2028, the best-covered forward year and the one all three fair-value anchors are struck on; 10 and 16 on FY2029. FY2030 (10 revenue / 10 EPS) and FY2031 (12 / 19) are excluded from every conclusion for the reasons in Section 8. epsAvg and revenueAvg were verified against FY2026 actuals to within 0.97% and 0.34% respectively; ebitAvg and ebitdaAvg were verified and failed, and are used for nothing.
• Peer note: the vendor peer set contains no clean comparable. Applied Materials ($434.0B), Lam Research ($397.4B), KLA ($255.3B) and Intel ($509.1B) are capital-equipment and manufacturing businesses with entirely different economics; Amphenol, Intuit, ServiceNow, Sony and Texas Instruments are not comparables at all. The only structurally relevant name in the set is Qualcomm ($170.8B), which is simultaneously an architecture licensee and Arm's litigation counterparty in two active Delaware actions. No peer multiples are supplied in the file, so no peer-multiple comparison is drawn. The 20-F names RISC-V as the competitive vector and the PRC's semiconductor self-sufficiency policy as the vehicle, and that is used instead.
• Fair-value caveat: the $155 / $236 / $322 anchors are multiples of the FY2028 consensus non-GAAP EPS distribution — 55x the low of $2.813, 78x the mean of $3.031, and 92.6x (today's forward multiple, held) the high of $3.477. Stated arithmetic, not a discounted cash flow. The base is deliberately sensitivity-disclosed: 70x gives $212, 85x gives $258, and 92.6x — today's multiple simply held — gives $281, which is spot. The whole answer lives in a 70-93x band and no greater precision is claimed. The bear case assumes a de-rating; the bull case assumes no re-rating at all, only an earnings beat. Note the implication of the base case's own arithmetic: if the multiple holds at today's level and consensus earnings arrive in full through FY2028, the total return is zero — there is no dividend and no buyback to add to it.
• Timing: Q1 FY2027 results were furnished on 2026-07-29, six days before this dive, on a 6-K containing no financial statements; the vendor's earnings calendar records a beat on both lines ($1,289M revenue against $1,262.683M; $0.45 non-GAAP EPS against $0.40). The next print is 2026-11-04, 92 days away (consensus revenue $1,376.798M, non-GAAP EPS $0.44) — no near-term company-specific catalyst, which is an explicit input to the verdict rather than an oversight. 2026-08-04 was a broad semiconductor up-day (MRVL +12.81%, INTC +10.92%, LRCX +7.85%, MU +7.62%, AMD +7.00%, KLAC +6.95%, AMAT +5.48%, TXN +5.42%, ASML +4.22%); ARM's +17.36% was the largest move in the batch, carried no company-specific filing, occurred on volume 8.8% BELOW average across a free float of only 145,344,760 ADSs, and is treated as sector beta amplified by a 3.767 beta and by float scarcity.
• Governance disclosure: Arm is a "controlled company" under Nasdaq rules and is eligible to rely on exemptions from corporate-governance requirements that protect holders of non-controlled companies. SoftBank Group beneficially owns 922,733,999 ordinary shares — 86.4% — as of 2026-05-21, and 769,029,000 shares, 72.0% of the company, are pledged as security under the SoftBank Group Facility, a margin loan whose foreclosure the 20-F says "could cause a change of control of us." Masayoshi Son is Chairman of Arm's Board and Chairman and CEO of SoftBank Group. A Shareholder Governance Agreement gives SoftBank pre-emptive rights, board-designation rights (up to eight of nine directors while it holds more than 70%), registration rights and consent rights. Arm is also party to a consulting agreement under which it advises SoftBank on SoftBank's own proposed investments and acquisitions. As a foreign private issuer Arm's audit committee is not subject to the independence determination applicable to Nasdaq-listed US companies. A material weakness in information-technology general controls was identified for fiscal 2022 and 2021 and the 20-F states remediation was completed during fiscal 2025. Arm is a plaintiff and defendant in two active Delaware actions with Qualcomm and Nuvia; on 2025-09-30 the court granted Qualcomm judgment as a matter of law and Arm has appealed to the Third Circuit. All of the above is treated as a governance risk input to the score of 8, not as an allegation of wrongdoing.
• Accessibility note: no information in this dive is conveyed by colour. All emphasis is carried by bold text, table structure and explicit labelling.
• Not investment advice. Independent research, educational and informational only, never personalised. No recommendation to buy, sell or hold any security is made to any person.
• Version: 2026-08-04-full.