Arm Holdings

IThe business

Arm sells the design of the brain of a chip. It does not make anything. Customers such as Qualcomm, Amazon, Google, Microsoft and NVIDIA license Arm’s instruction set and its ready-made processor designs, have chips built from them at a foundry, and pay Arm twice. They pay a license fee when they sign or deliver, and a royalty on every chip they ship.² The royalty is a percentage of the chip’s price or a fixed amount per chip, it falls as a customer’s volume grows, and it rises when more Arm designs go into the chip.² In the twelve months to September 2026 license and other revenue was $2.67B and royalty revenue $2.79B C (calculated).²³

The cost base is engineers. Arm had 9,584 employees at 31 March 2026, 84% of them in engineering, and it spent $2.78B on research and development in fiscal 2026, 56% of revenue.² Cost of sales is $121M, so gross margin is 97.5% on the accounting basis C (calculated). Arm files a Form 20-F each year and Forms 6-K each quarter, not a 10-K and 10-Qs, because it is a foreign private issuer based in Cambridge, England. It must retest that status each September, and if it lost it Arm would file 10-Ks and 10-Qs from 1 April 2027; the November report does not say it has.²³ Its fiscal year ends on 31 March, so fiscal 2026 ended on 31 March 2026 and the latest quarter, the second of fiscal 2027, ended on 30 September 2026.

Where the royalty comes from

Arm now reports three groups of markets. Edge AI covers smartphones and consumer and industrial devices. Physical AI covers cars and robots. Cloud AI covers servers and networking.² Phones are the base. Arm’s share of phone application processors has been above 99% for years, and mobile application processors made up 43% of royalty revenue in fiscal 2026 R (reported).² That is about $1.1B of the $2.6B C (calculated). The growth is elsewhere. Data center royalties have at least doubled year on year in each of the last four quarters, physical AI royalty rose 50% in the September quarter, and Arm says more than 10 billion Arm-based chips were reported shipped in that quarter, 1.5 billion more than a year earlier.⁴⁵⁶⁷

The data center claim is the one the price rests on. Arm says its share of server CPU compute at the top hyperscalers went from 15% in calendar 2024 to more than 50% in 2025 and will exceed 60% by the end of 2026, on current ramps of Google Axion, AWS Graviton, Microsoft Cobalt and NVIDIA’s Arm-based chips.⁷ Those are Arm’s figures. I cannot check the definition, and Arm does not publish data center royalty in dollars.

A new business: chips of its own

In March 2026 Arm announced the Arm AGI CPU, its first production chip: up to 136 Neoverse V3 cores, 300 watts, made by a foundry to Arm’s design, with Meta as lead partner and co-developer.¹¹ Arm says first commercial deliveries are on track for the March 2027 quarter, it has purchase orders from several customers into calendar 2027, and it is more confident of “up to $2 billion” of Arm AGI CPU revenue across fiscal 2027 and 2028.⁷ Revenue so far is immaterial.² Three commitments show the size of what Arm is taking on. In October 2026 it agreed to purchase $672.0M of semiconductor product from a supplier through 2027 and, at the same time, took a customer commitment from a related party to supply it; Arm acts as a reseller agent. In November it agreed to pay $200.0M of advance payments in the first half of 2027 to reserve manufacturing capacity for 2027 to 2029, with minimum purchase volumes and shortfall payments. And on 1 July it bought DreamBig Semiconductor, a networking technology company, for $265.3M.³ I come back to what this changes in Valuation.

The latest quarter

The September quarter was a record for that quarter. Revenue was $1,443M, up 27%. License and other revenue rose 50% to $773M, and royalty revenue rose 8% to $670M, which is lower than the $715M of the June quarter.⁷ GAAP operating income was $201M, a 13.9% margin. Non-GAAP operating income, which leaves out stock pay, was $638M, a 44.2% margin R (reported). GAAP earnings per share were $0.12 and non-GAAP $0.52.⁷ Arm guides the December quarter to $1.54B of revenue, plus or minus $50M.⁷ The results came out after the market closed on 4 November. The stock closed that day at $191.80 and the next day at $166.72, down 13.1% C (calculated), and closed at $186.30 on 9 November.¹

Where Arm is from

Arm began in 1990 as a joint venture of Acorn Computers, Apple and VLSI Technology. It was listed in London and on Nasdaq from 1998 until SoftBank Group bought it in September 2016.² Arm Holdings plc, the present holding company, was set up in 2018 and listed its American depositary shares on Nasdaq on 14 September 2023. The offer price was $51.00. The seller was SoftBank’s group, which sold 102.5M ADSs; Arm received none of the money.²¹² Each ADS represents one ordinary share.² At 21 May 2026 SoftBank owned 86.4% of the company, which makes Arm a controlled company, and that fact runs through the rest of this page.²

IIThe moat

Arm has one strong protection and four places where it leaks. The strong one is software.

Software built for Arm

A chip needs an instruction set, the list of operations a processor can carry out. Programs are written for one list. Arm owns the one that every key phone operating system depends on, and it says more than 350 billion Arm-based chips had shipped by 31 March 2026 and more than 30 million developers write for the platform.²⁷ A company that wants to leave must port its software and its tools, and then wait. The 20-F says it can take two to three years to design a new Arm processor and another two to three to develop a chip around it.² That time is the barrier, and it is why Arm collects a royalty on “substantially every chip” made with its designs under every license model it sells.²

The protection is a position in a language, not a law. Nobody is barred from writing another language. That is the first leak, and it is a slow one.

Where it leaks

RISC-V and x86. The 20-F names both as competitors. RISC-V is a free, open instruction set, and the filing says some of Arm’s own customers are major supporters of it. x86, from Intel and AMD, already has a large installed base in servers and PCs.² Customers who design their own cores. An architecture license lets a company build its own processor core to Arm’s instruction set, so Arm sells less design and more permission. Customers with bargaining power. The top five customers, including Arm China and SoftBank, were 57% of fiscal 2026 revenue, and some key contracts let a customer take Arm’s latest products as soon as any other customer can.² Arm itself. The 20-F warns that selling its own chips may bring Arm into competition with customers, which could make them cut back and choose x86 or RISC-V.² Bloomberg reported on 16 May 2026, citing people familiar with the matter, that the US Federal Trade Commission is examining whether Arm will refuse or degrade CPU licenses while it builds its own chips; the FTC and Arm declined to comment, and the 20-F, filed ten days later, does not name the probe.¹³² Qualcomm says something close in its latest annual report: Arm has become a chip supplier in direct competition with its licensees, and Qualcomm’s ability to license Arm’s latest technology on acceptable terms has been and may continue to be affected.⁹ I cannot tell from the outside how many customers think this way. Arm’s growth in cloud licenses says most do not, for now.

Arm China

Arm sells into China through Arm China, its exclusive distributor of licenses to Chinese customers. SoftBank controls Acetone Limited, which owns 48.2% of Arm China, and Arm holds 10% of Acetone without votes, so Arm’s indirect share is about 4.8%. The agreement runs to 2048. Arm does not control Arm China, and its revenue depends on the reports Arm China sends it.²³ Arm China was 16% of fiscal 2026 revenue, $790.6M, against about $670M in each of the two years before; the PRC in all was 18%.² In the September quarter Arm China paid $208.9M against $243.1M a year earlier, and over six months $398.3M against $444.1M, a fall of 10% C (calculated). It owed Arm $289.4M at 30 September, and Arm wrote off $16.0M of Arm China receivables after an audit.³ I treat China as a leak of unknown size: flat for two years, up in fiscal 2026, falling now, and exposed to export rules and to a distributor Arm does not control.

The Qualcomm dispute

Qualcomm was 9% of fiscal 2026 revenue, about $443M C (calculated).² It is a major customer, it is suing Arm and being sued by Arm, and Arm’s June letter notes that Qualcomm has announced plans to enter the AI data center CPU market with an Arm-based chip, the Dragonfly C1000.²⁹⁶ The table sets out where each case stands. The two sides describe the October jury trial differently, and both can be true: Arm counts verdicts by claim, and Qualcomm lists what the jury found Arm did wrong. Damages to Qualcomm were nominal. The questions that carry money are undecided: whether Qualcomm may withhold certain royalties, and the appeal on the Nuvia ruling.³⁹ In my view the lasting cost is whether customers that design their own cores believe the license will be honored as signed.

IIIThe competitive advantage

The software lock-in says Arm is hard to leave. This section is about what Arm earns for staying: how much it gets paid per chip, and whether that is rising.

The claim: more Arm in every chip

A chip with one old core pays a small royalty. A chip with eight newer cores, a graphics block and a pre-built subsystem pays more. That is Arm’s whole argument, and its letters repeat it every quarter: royalty revenue is up on higher royalty rates per chip from Armv9 and CSS.⁷ Compute subsystems are the sharpest form of it. Since Arm announced them about two and a half years ago, it had signed 21 CSS licenses across 12 companies by December 2025, five customers were shipping CSS-based chips, and the top four Android phone makers shipped devices with them.⁴ On the November call, as transcribed, the chief financial officer said it signed four more in the September quarter, two for data center chips and two for smartphones.¹⁴ In servers, Arm says higher core counts and higher prices per core are driving a rise in royalty, and Neoverse shipments reached 1.5 billion cores, 500 million of them in the last nine months.⁵⁶ AWS’s Graviton5 has 192 cores, twice Graviton4, and NVIDIA’s Vera has 88 against 72 for Grace.⁴

What I can check, and what I can’t

Arm does not publish a royalty rate for any generation. No filing gives the Armv9 rate, the gap to Armv8, or a CSS multiple. The table shows what the documents do say.

The one place I can do arithmetic is royalty per chip, and it does not give a clean answer. Before the listing, Arm reported chips shipped each year. Royalty per chip rose from 5.2 cents in fiscal 2021 to 5.5 cents in fiscal 2023, about 6% in two years.¹² The letters give a recent count: more than 10 billion chips in the September quarter and 1.5 billion fewer a year earlier. Divide each quarter’s royalty by its own count and royalty per chip fell, from about 7.3 cents to under 6.7 cents, roughly 7% to 8% C (calculated). But on the 4 November call an analyst read the same letter as showing a rise of at least mid-single digits, and the chief financial officer called it mixed and said he would not forecast from the average across 10 billion chips, as transcribed.¹⁴ The two readings point opposite ways. The chips Arm counts are those partners report in the quarter, and Arm books royalty on shipments in the quarter, so the two may sit a quarter apart. Mix matters too: a rising number of low-priced embedded chips pulls the average down even if each premium chip pays more. I do not lean on either reading, and nothing in my values uses royalty per chip. I would like to model phones, servers and physical AI separately. I can’t, because the split of royalty dollars by market is not published except for phones at 43%.²¹⁴

Do the royalty numbers add up?

Arm gives growth rates for the groups, not dollars. In the September quarter total royalty rose $50M, from $620M to $670M. Edge AI, mostly phones, was roughly flat. Physical AI rose 50%. Data center at least doubled.⁷¹⁴ Those pieces add up to a $50M rise only if data center was a small part of the year-earlier quarter. The chief financial officer said, as transcribed, that last year data center was “approaching 20%-ish” of total royalties.¹⁴ If it had been 20% of $620M, about $124M C (calculated), doubling it alone would have added $124M, more than twice the whole rise, and edge would have had to fall. Phones were 43% of royalty in fiscal 2026.² So one of three things is true: data center was a much smaller share than 20% a year ago, the doubling covers less than the group that the 20% describes, or edge fell by more than “roughly flat” suggests. I can’t tell which. It matters because the price rests on data center royalty. If the base is small, doubling it moves the total little and a soft phone quarter shows up in the total at once. If the base is large, the rest of royalty is shrinking. Publishing data center royalty in dollars would settle it.

Evidence the advantage is widening

• Data center royalty has at least doubled year on year in each of the last four quarters, and the chief financial officer expects triple-digit growth over the next few years.⁴⁵⁶⁷¹⁴
• Outside customers took $490M of licenses in the September quarter against $222M a year earlier, up 121%, so the license boom is not only a related-party story C (calculated).³
• Annualized contract value, Arm’s measure of normal license revenue, rose 12% to $1,790M.⁷
• Physical AI royalty rose 50% and Arm Total Access licenses reached 56 at March 2026, more than half of the top 30 customers.⁷⁵
• Revenue has grown more than 20% a year for five straight quarters.⁷

Evidence it is narrowing

• Total royalty growth was 27%, 11%, 22% and 8% in the last four quarters. Royalty is seasonal, so I compare each quarter with the same quarter a year earlier.⁷⁴⁵⁶
• Edge AI royalty was roughly flat year on year in the September quarter, the chief financial officer said, and Arm blames higher memory prices for cutting phone unit sales.¹⁴⁷
• Arm China revenue fell 14% in the quarter. And in fiscal 2026 all of the license growth came from related parties: license revenue from outside customers fell 9% and from related parties rose 141%.³²
• The chief financial officer guided December-quarter royalty to rise by a low-teens percentage on the September quarter and only a low single-digit percentage on a year earlier.¹⁴
• Qualcomm says Arm’s move into chips is hurting its access to Arm’s licenses.⁹

Where Arm wins, and where it loses

Arm wins where a customer wants a proven processor with software already written for it and does not want to spend three years designing one: phones, hyperscaler servers, cars, and any chip company too small to design cores. It loses to the customer big enough to build cores and unhappy enough with the price to try RISC-V, and on the fringe, in low-priced embedded chips, where the royalty is small and the free option is close. Since March it has a new way to lose: the customer who takes delivery of Arm’s finished chip and asks why it still pays Arm for the design.

IVManagement

Who runs it, and who they work for

Masayoshi Son, who founded and runs SoftBank Group, is Arm’s chairman. He also chairs the board’s pay committee and takes no pay from Arm.²¹⁵ All seven non-executive directors, Mr Son included, were named by SoftBank under a governance agreement that lets SoftBank pick seven of the eight board seats while it owns more than 70%.² Arm is a controlled company under Nasdaq rules, which lets it skip some governance requirements; it keeps a pay committee, but the controlling shareholder’s chief executive chairs it.¹⁵ SoftBank also holds the right to subscribe for new shares to keep its share, to see results before they are filed, and to approve any new share plan larger than 5% of the company.²

Rene Haas has been chief executive since February 2022. He sits on SoftBank Group’s board, advises it, and since 20 April 2026 is also chief executive of SoftBank Group International, which oversees parts of the parent’s chip and AI holdings; Arm says that role is limited and part-time.²¹⁰ The chief legal officer, Spencer Collins, was general counsel at SoftBank Investment Advisors.² This is a team that knows the owner well. It was not chosen by the outside shareholders, who own 13.6% of the company.

How they are paid, and what that pays them to do

Mr Haas’s reported pay was $60.6M in fiscal 2026, $24.5M in fiscal 2025 and $70.1M in fiscal 2024.¹⁵ Of the $60.6M, $55.6M was share awards valued on the day they vested, and about two-thirds of that, $36.6M, was the rise in the share price since the award was made C (calculated).¹⁵ His salary has been $1.35M for three years.

The plan for fiscal 2027 changes in two ways. The annual share award, 14 times salary or $18.9M, can now vest up to 200% instead of 125%, so up to $37.8M. And there is a new one-time award, 425,000 shares, that pays only if Arm’s market value reaches $1.0 trillion by March 2029, $1.5 trillion by March 2030 or $2.0 trillion by March 2031, measured on a 60-day average price.² Today Arm is worth about $201B C (calculated) on diluted shares. The top milestone is ten times that, and the 425,000 shares are worth $79M now. The pay report puts the payout at $100M if only the first milestone is reached and $799M at full vesting, which needs the share price to rise 1,208% from $143.74, its 60-day average to 30 April.²¹⁵ My table values each tranche at its own milestone on 1,070M shares and gets about $645M for all three C (calculated). Shareholders approved the policy on 9 September, with 939.3M votes for, 36.4M against and 5.7M withheld. SoftBank holds 922.7M shares. If it voted all of them in favor, which the result does not say, everyone else cast about 16.6M for and 36.4M against C (calculated).¹⁶

Follow the incentive. The annual bonus is paid half on revenue and half on non-GAAP operating income. The annual share award is paid on the same two numbers, a strategy modifier, and a quarter on share return against the S&P 500 technology index.² Non-GAAP operating income leaves out stock pay. Revenue counts the parent’s consulting fees the same as a royalty from a phone maker. The new award pays on one thing, the share price. Nothing in the plan rewards cash left after stock pay, the return on the money going into chips, or revenue from customers other than the owner. A chief executive can fairly be paid for a rising share price. This plan pays him to chase the price and does not check how it is done. The plan does not look at how the revenue is made, so it would not tell a good dollar from a poor one.

How much of their own money is in

Mr Haas owned 396,007 shares on 21 May 2026, worth $73.8M at $186.30, and Mr Child 174,232, worth $32.5M; after a sale on 21 September, Mr Child’s holding was 153,442.²¹⁷ Almost all of it, as far as I can see, came from pay. The company asks the chief executive to own ten times salary in shares, and at March’s price he held more than 27 times.¹⁵ Arm’s officers first filed Forms 3 on 18 March 2026, so there are no earlier Forms 4. Since 25 March the seven officers who file them have sold 323,715 shares for $72.2M, and none has bought a share on the market C (calculated).¹⁷ Several of the sales cite plans adopted in advance under Rule 10b5-1. A sale on a plan is not a signal in itself. But an investor at $186.30 would want to see someone with inside knowledge pay that price once.

Control is the other half of ownership. SoftBank owns 922.7M shares, 86.4%. At 21 May it had pledged 769.0M of them, 72.0% of the company, as security for a margin loan that began at $8.5B in 2023.² Bloomberg reported in September, citing people familiar with the matter, that the facility was raised by $5B to $25B, with $20B drawn as of December and the loan due in September 2027.¹⁸ SoftBank’s own June report deducts the equivalent of ¥3.22 trillion for margin loans using Arm shares, which is about $19.8B at its ¥162.39 rate C (calculated).¹⁹ At $186.30 the pledged shares are worth $143B C (calculated), so the loan is about 14% of that value, and at $25B it would be about 17% C (calculated). The pledged shares would only just cover $25B at $33 a share C (calculated). The 20-F says the lenders can call for more security, or require repayment, if the price falls below thresholds it does not disclose, and that SoftBank may have to dispose of shares.² I do not predict that. The size matters: the public float is worth $27B, and a placing of 5% of SoftBank’s stake, 46M shares, would equal 32% of the float C (calculated).

Alphabet Inc.’s 13F-HR for the quarter ended 30 September 2026, filed on 6 November, lists 1,960,784 Arm ADSs, reported under Google LLC, valued at $568.0M at the 30 September close of $289.66, which is 0.18% of the shares in issue C (calculated).²⁰¹ At the $51.00 offer price that count is $100.0M C (calculated), and the listing prospectus names a Google entity among the cornerstone investors that had indicated an interest in the offering.¹² That makes it look like an unchanged allocation from the listing and not a bet on the recent price, though neither filing says so.

What they did with capital over a full cycle

The record is short, because Arm has only been listed for three years, and it is mostly spending on the business. Arm has paid no dividend and bought no stock since listing. Its one large payment out was a $750M dividend in fiscal 2021, when SoftBank and its Vision Fund owned all of it.¹²² It spent $545M on property and equipment in fiscal 2026, up from $64M in fiscal 2023, and research and development rose from $1.1B to $2.8B over the same years.²²¹ It sold its Artisan foundation IP business to Cadence in August 2025 for a $131M gain, received $143.4M for its stake in the chip designer Ampere when SoftBank bought Ampere in November 2025, and on 1 July 2026 paid $265.3M for DreamBig.²³ It also put $16.4M into Kigen, a company a SoftBank fund controls.²³

The one thing that works like a buyback is paying employees’ tax. In fiscal 2026 Arm paid $529M to cover tax on vesting shares so it could issue fewer, and $407M more in the first half of fiscal 2027.²³ Even so, shares outstanding rose from 1,025M in March 2023 to 1,070M in September 2026, up 4.4%, though the pace slowed from 1.6% in fiscal 2025 to 0.7% in fiscal 2026 once Arm began paying the tax in cash C (calculated).²²¹³ Stock pay cost $1.3B on the income statement in the last twelve months. The tax Arm paid on vesting was $649M, and about 0.65% of the company a year was issued net, worth about $1.3B at today’s price C (calculated). Together that is about $1.9B, more than the expense, because the awards were granted when the price was lower. I can’t tell skill from luck in a record this short. Management spent on building, kept $3.6B in cash, issued shares every year and returned nothing. That does not tell me what it will do with $200B of market value.

VFinancials

Six fiscal years sit on one footing. Arm was private from 2016 to 2023, its first annual report under the present company covered fiscal 2024, and fiscal 2021 comes from the listing prospectus. Earlier years were reported by a different company, in pounds, under different rules, so I do not splice them, and the ten years shown on other pages are not available here.²¹²²²

What an owner could really take out

Over the five fiscal years from 2022 to 2026 Arm reported $3.08B of net income. Its free cash flow before stock pay, on its own definition, was $2.84B. Stock pay expense was $3.01B C (calculated). So the cash an owner could have taken, after stock pay, was −$0.17B C (calculated). In the last twelve months it was −$169M: operating cash flow of $1,734M, less $638M of capital spending, less $1,265M of stock pay C (calculated). Arm reports free cash flow of $1.10B for that period. Stock pay, which that figure leaves out, equals 73% of operating cash flow.⁷³

Two cautions run in opposite directions. The first makes the figure look worse than it is. Stock pay is a cost I am deducting at its accounting value, and operating cash flow in the first half of fiscal 2027 was helped by a $229M swing from the timing of payroll tax on vested shares.³ The second makes it look better. Receivables, contract assets and contract liabilities together took $685M of cash in the twelve months C (calculated), which is the revenue-ahead-of-cash point below, and Arm has begun paying large sums in cash to settle employees’ tax, $649M in the twelve months C (calculated).²³ I show both views and do not pick the kinder one.

Quality of earnings

• Revenue from the owner. Affiliates of SoftBank paid Arm $704M under a consulting agreement in fiscal 2026, against $146M the year before, and $402M in the first half of fiscal 2027. The agreement covers “technical consultancy and advisory services” on possible deals and partnerships for SoftBank and its affiliates, and it is the largest source of Arm’s license growth.²³ At 30 September, $617M of what Arm has booked from them had not been billed. Since fiscal 2025 Arm has booked $1.25B from these affiliates and billed about half C (calculated).³ The chief financial officer says he expects the quarterly rate for the rest of the year to stay near $209M.¹⁴ I cannot see what the services are, what they would cost from anyone else, or whether the agreement continues. The 20-F lists it as a risk: part of near-term revenue depends on arrangements with the parent.²
• License growth is related-party growth. License and other revenue from outside customers fell 9% in fiscal 2026, from $1,421M to $1,298M, while related-party license revenue rose 141%, from $418M to $1,009M.² Revenue booked over time, which is where the consulting agreement sits, went from $121M in fiscal 2024 to $1,080M in fiscal 2026, while point-in-time license revenue, the normal sale of designs, was $1,310M, $1,372M and $1,227M R (reported).² The September quarter looks different: outside license revenue was $490M against $222M. One quarter is not a trend.³
• Revenue ahead of cash. Contract assets, revenue booked but not yet invoiced, were $1,450M at 30 September against $270M at March 2023 R (reported). That is 27% of last year’s revenue, against 10% then C (calculated). In the first half of fiscal 2027 contract liabilities took $290M more cash than a year earlier, which Arm attributes to revenue recognized ahead of billing.³²² Arm stopped reporting its remaining performance obligations and license counts in the shareholder letter from the June quarter.⁷ The 6-K still shows the obligations: $2,422M, a third due within a year.³
• Growth without the owner’s fees. Leave out the SoftBank consulting fees and revenue grew 9% in fiscal 2026, to $4,216M from $3,861M, and 24% in the first half of fiscal 2027, to $2,330M from $1,884M C (calculated). My base has license and royalty revenue, the owner’s fees included, growing 14% in fiscal 2028, which sits between the two.²³
• One-offs and gains. Fiscal 2026 net income of $904M includes a $131M gain on selling Artisan, and the first half of fiscal 2027 includes $108M of unrealized investment gains. Tax was 13% of pre-tax income in the half against 22% for fiscal 2026; I use 22%.²³
• Stock pay against the headline. In the September quarter non-GAAP net income was $566M and GAAP net income $130M. Stock pay and the employer tax on it account for $425M of the $436M gap. Arm has $3.4B of stock pay still to be expensed on awards already granted, over a weighted 1.5 years, and 23M unvested awards outstanding.⁷³
• Commitments. Non-cancellable purchase obligations of $1,057M for cloud services, data centers and software run to 2036, and were raised by $305M in April 2026. On top come the $672M of chips Arm will resell to a related party, the $200M of capacity prepayments and $480M of operating leases.²³

The balance sheet

Arm has $3.6B of cash and short-term investments and no debt; the 20-F says it has none and may borrow later.² The largest assets after cash are the $1.1B of receivables and the $1.45B of contract assets, which together equal 46% of the last twelve months’ revenue C (calculated). Property and equipment has risen from $354M at March 2025 to $941M; the 20-F ties the fiscal 2026 rise in spending to data center and office expansion and computer hardware, and does not say more.²³ I count net cash of $3.6B in my values. I leave out $599M of equity investments and $480M of lease liabilities, which roughly cancel in the value per share.

VIValuation

First, how this goes badly wrong

Before the numbers I asked what would most likely hurt an owner at this price, and worked back. Three things stand out. None needs anything exotic.

The price comes down to a multiple the business can carry. At $186.30 Arm is worth $201B, 37 times last year’s revenue. NVIDIA is 16.4 times and Broadcom 17.8 times on their own pages.²³²⁴ At those multiples Arm would be worth $90B to $98B, or $83 to $90 a share C (calculated). That comparison flatters Arm: NVIDIA’s operating margin is 65.2% and Arm’s, after stock pay, is 16.7%.²³³ The stock fell 58% from June to November while quarterly revenue kept growing more than 20% a year, which is what a price resting on a story does.¹⁷ This is the likeliest way an owner loses money, and it needs no failure at Arm.

The chip business goes wrong, or costs the licenses. The chip business adds inventory, supplier commitments of $672M and $200M, a first-version gross margin of high 30s to low 40s percent, and customers that may now see Arm as a rival.³¹⁴² If the chips earn nothing and 10% of license and royalty revenue goes elsewhere, my base falls from $44 to $33 E (estimate). That moves my base by $11, a small step beside the $142 between my base and the price. What is large is the bear case, where chips lose money and licensees defect together.

The owner. At 21 May SoftBank owned 86.4% and had pledged 72.0%, against a loan of roughly $20B. It paid Arm $704M in fiscal 2026, 14% of revenue, for work I cannot price.²¹⁹ If SoftBank needs cash it can place shares, and 5% of its stake is 32% of the public float. If its consulting revenue stopped, I estimate my base falls from $44 to $42 if the cost of earning it went with it, and to $39 if the cost stayed E (estimate), a small cut either way. The harm would come through the price, not the cash flows: a falling share price raises the odds that the loan needs more security, and a large block meets a thin market.

What an owner could really take out

I value cash after stock pay and after the capital needed to stand still. Operating income is already after stock pay. From it I take 22% tax, add back depreciation, and subtract all capital spending and the cash that growth ties up in receivables and contract assets. The result for the last twelve months is −$169M C (calculated), against net income of $936M. Management’s non-GAAP figures, which leave stock pay out, show a business with a 44% operating margin; mine is 17%. The gap is the central question about this company.

My margin for FY2031 deserves one more sentence. Management says the license business will earn a 65% operating margin, but that is a non-GAAP figure, before stock pay.⁸ My base has license and royalty revenue earning 36% after stock pay in FY2031. The two agree if stock pay and its payroll tax stay near 29% of that revenue, a little above the 27% by which management’s operating income exceeds mine today C (calculated). The CFO tied the 65% to a chip business at a $15B rate, and said it might be an operating or an EBITDA margin; my base has only $6B of chips.⁸ My bull case needs stock pay and its tax to fall to about 19%. History is less kind than my base. From fiscal 2025 to the last twelve months revenue rose $1,457M and GAAP operating income $84M, an incremental margin of 6%, because stock pay grew with revenue.²³ My base needs about 36% on the extra revenue of fiscal 2028 and then keeps climbing. If the license and royalty margin is five points lower in every year my base is $40, and ten points lower, $35 E (estimate).

What the price implies

Put the other way, what does $186.30 need? At a 10.5% discount rate and 3.5% growth after FY2037, it needs revenue of about $71B in FY2037 at the margins of my bull case, which is 1.5 times that case and 80% of Broadcom’s revenue today.²⁴ In FY2031 that path is $38B of revenue, against management’s $25B target.⁸ On my base case’s shape the price needs $100B of revenue in FY2037 C (calculated). The price needs more than the plan: 50% more revenue than my bull case, which already assumes the plan is met.

What I can’t judge

• The royalty rate by generation. It is not published. I can see direction and totals, not the rate for Armv9, a CSS design or a server core, nor the dollars by market except phones.
• What the chip business earns. Management expects the first version to earn a gross margin in the high 30s to low 40s percent, and a 35% operating margin by FY2031 on $15B of chips.¹⁴⁸ No chip has shipped. I give the range, $0 to $15B of revenue and 5% to 35% margin, and let the cases show what it does to the value.
• The owner’s consulting revenue. I don’t know what Arm provides, whether others would pay for it, or whether $0.8B a year repeats.
• Stock pay. It is 23% of revenue now. A business whose revenue outgrows its payroll would normally see that fall. Arm has no record of it yet, and $3.4B is already granted.
• Hyperscaler share. Arm’s “more than 50%” is its own figure and its own definition.
• Qualcomm and China. I can’t price a court ruling or a distributor I don’t control. They are 9% and 16% of revenue.

Because I can’t know these, I show a range and the price’s position in it, and I don’t treat $44 as exact. A fair reading of my work is that the base is somewhere from $30 to $90, and the whole range sits far below the price.

The discount rate

I ask 10.5%, the rate I use for Broadcom. NVIDIA is at 10.0% and Marvell at 12.0%.²⁴²³²⁵ The ten-year Treasury yielded 4.88% on 9 November.²⁶ Arm has no plant to build, which argues for less. It has a controlling owner with a pledged stake, a lawsuit with a large customer, revenue from its owner, and a new business it has never run, which argue for more. I settle on 10.5%. At 9.5% my base is $52 and at 11.5% it is $38. My bull case is $155 at 9.5%, still below the price.

Three cases

Bear

License and royalty revenue grows 7% in FY2028 and 3% a year from FY2032, and does not accelerate. The consulting revenue from SoftBank’s affiliates fades, Arm China and Qualcomm take more than they give, and stock pay stays at about a quarter of license and royalty revenue, so that revenue earns 26% after stock pay. The Arm AGI CPU sells $0.7B to $1.0B a year at a small loss at first and then a 5% margin. Revenue reaches $8.1B in FY2031. Value $16.

Base

License and royalty revenue grows 14% in FY2028, slowing to 9% by FY2032 and 4% by FY2037, close to management’s $10B target for FY2031 but 9% short of it. The chip business sells $1.5B in FY2028, $6.0B in FY2031 and $8.5B in FY2037 at a 20% margin once it is running. License and royalty revenue earns 36% after stock pay in FY2031 and 40% from FY2034. Value $44.

Bull

Management meets its fiscal 2031 plan, $15B of chip revenue and $10.4B of license and royalty revenue, and total revenue then keeps growing about 11% a year, on average, for six more years. Chips earn 33% to 35%, the margin management gave. License and royalty revenue earns 46% after stock pay in FY2031 and 52% later, which needs stock pay and payroll tax to fall from 27% of that revenue now to about 19% in FY2031 and 13% later. Value $125.

What it takes to close the gap

The table makes one change at a time to my base, then adds the changes together. No single change closes the gap. The largest steps are the bull chip path with its 33% to 35% margin, which meets the plan and lifts the base to $89, and an 8.5% discount rate, which lifts it to $64. Added together in the order shown, the bull chip path, the bull license and royalty growth and the bull margins reach $125, my bull case, and the price needs more. My bull case at an 8.5% discount rate is $199, just above the price E (estimate). The share count moves the answer by less than a dollar: I divide by 1,079M diluted shares, and one ADS is one ordinary share, so a per-share value is also a per-ADS value.²³

Other uses of the same dollar

The best comparison on this site is Broadcom, the nearest in what it sells: designs that other companies’ chips depend on, and a large stock pay bill. At its price my base returns 7.5% a year against the 10.5% I ask, a shortfall of 3.0 points. Arm at $186.30 returns 5.3%, a shortfall of 5.2 points, and my bull case 8.7%.²⁴ NVIDIA’s base returns 9.7% on a 10.0% requirement.²³ By my numbers none of the three clears its bar, and Arm misses by the most. The ten-year Treasury yields 4.88%, so my Arm base earns 0.4 points more than that for owning a controlled company with a pledged stake.²⁶ Broadcom pays its owner in cash after stock pay. Arm so far does not.

Arm has not repurchased a share and pays no dividend. At $186.30, against my base of $44, repurchasing stock would destroy value for continuing owners, so I do not ask for it. I would ask for something else: that stock pay stop growing faster than revenue.

What the market paid before

The stock closed at $63.59 on its first day. Since then the market has paid very different prices for a dollar of Arm revenue. At the offer price it was 19.5 times, at the March 2024 close 40.2 times, and at the June 2026 high 95.2 times C (calculated). At 36.5 times it is 62% below the high, below the March 2024 close, and above the other three dates C (calculated). A price that has swung between 19 and 95 times revenue in three years has been set by changing hopes more than by changing cash.

The margin of safety

I set it from the bear case: I want a price at which my $16 bear value is at least three quarters of what I pay, which is $21. The price is 8.9 times that. The lowest close since the listing was $47.87 on 20 October 2023, and the lowest close in the last year was $104.55.¹ I am not saying the price will fall to $21. I am saying an investor there would be paying for the license business and getting the chips as a gift.

Two different questions

The first is how good the business is. By the numbers it is very good: a 97.5% gross margin, more than 99% of phone chips, a royalty on every chip, net cash of $3.6B, and the data center royalty doubling. The second is how good the price is. At $186.30 the price is 323% above my base value and 49% above my bull, and an owner has so far taken out nothing after stock pay. Those are different answers. A great business can be a poor purchase when the price already holds the next decade of its best case. This looks like that case. The market may be seeing what I am not. The strongest form of that is a data center royalty and a chip business that outgrow my bull, and I show in Section VII what it would take.

How the market could be right

Three routes lead to today’s price, and I can say how far each is from where Arm stands. The first is a plan that is met and then beaten: the price needs about 1.5 times my bull case’s revenue, $38B in FY2031 against the $25B target C (calculated). The second is a lower required return. My bull case, which is the plan met, is worth $125 at 10.5%, $155 at 9.5% and $199 at 8.5% E (estimate). A buyer who asks 3.8 points over the 4.88% Treasury yield, and who believes the plan, pays today’s price; I ask 5.6 points, and I treat the plan as the top of my range, not the middle, because Arm has not yet shipped a chip.²⁶ The third is a high multiple on the plan’s earnings. At $186.30 the price is under 20.7 times the more than $9 of earnings per share management points to for fiscal 2031. If the shares still traded at 30 times that figure in March 2031 they would be $270, a return of 8.8% a year from today; at 40 times they would be $360, 16.2% a year C (calculated). Those earnings, if they are non-GAAP as the margins given imply, leave out stock pay, which is about a quarter of revenue now, and they depend on a chip business that has not shipped, so they flatter. They are the argument as its holders would make it, and I cannot rule it out.⁸

VIIThe case for the price

What most people miss

Three facts cut in Arm’s favor and get less attention than the stock pay and the owner. Neoverse server cores took six years to reach the first billion shipped and nine months for the latest 500 million.⁶ License revenue from outside customers was $490M in the September quarter against $222M a year before, up 121%.³ And Arm Total Access, an annual license to the whole portfolio, had 56 customers at March 2026, more than half of the top 30.⁵

The bull case

I found no outside investor or analyst whose published view on Arm I could check against a dated document, so I give the case as the company makes it. In the 4 November letter, chief executive Rene Haas calls demand for the Arm AGI CPU “extremely strong” and says confidence in achieving “up to $2 billion” of chip revenue across fiscal 2027 and 2028 has risen.⁷ The 6 May letter says Arm is on track to its $15 billion chip forecast, without a date.⁵ On that day’s call, as transcribed, chief financial officer Jason Child put the date at fiscal year-end 2031, with $10B of license and royalty revenue and earnings per share above $9, and later said the license business gets to something like 65% and chips to the mid-30s at a $15B rate. The transcript does not say whether the $9 is GAAP; the margins imply non-GAAP.⁸

In plain words: AI data centers need more CPUs, the biggest cloud companies and NVIDIA already use Arm designs, each generation earns more per chip, and Arm now sells the finished chip. If that is right, revenue goes from $5.5B to $25B in four and a half years and the price is under 20.7 times the earnings management is aiming at.

Against reality

If Arm does what it says, and hits those targets, I get my bull case, $125, which earns 8.7% a year at today’s price against the 10.5% I ask. The price needs more than the plan: about 1.5 times my bull revenue, $38B in FY2031 against $25B C (calculated). To get from here to $25B by March 2031, revenue has to grow about 40% a year for four and a half years C (calculated). It grew 16% a year from fiscal 2022 to fiscal 2026 and 23% last year.²²² Three quarters of the extra revenue would come from a business with none today: $15B of chips. Then it needs EPS above $9, which looks non-GAAP, against $2.00 over the last twelve months, and stock pay, now a quarter of revenue, to stop eating the gap.⁸⁷

It helps to multiply it out. The price needs five things: license and royalty revenue of $10B by fiscal 2031 (about 55%), chips of $15B at a 35% margin by then (about 15%), stock pay that stops outgrowing revenue (about 40%), a calm owner with no forced sale or break in consulting revenue (about 80%), and Qualcomm and China unchanged (about 75%). Together that is about 2%, or about one chance in 50. They are my guesses, they are not independent, and I do not use them in the values E (estimate).

What a 10x would take. Ten times $186.30 is $1,863, or $2.01 trillion of market value on 1,079M diluted shares. At 30 times earnings after stock pay that needs about $67B of net income, which is 72 times the $0.94B of the last twelve months C (calculated). At a 30% net margin that means about $223B of revenue, 41 times the last twelve months and 45% growth a year for ten years. The chief executive’s new award sets $2.0 trillion by March 2031, four and a half years away, and a tenfold price in that time is about 69% a year.² Arm’s revenue has grown 16% a year in the past four years. A tenfold has happened to a very large company in recent years: NVIDIA’s share price is 15.9 times its low close of 14 October 2022, and its revenue went from $27.0B in fiscal 2023 to $303.0B, 11 times.³⁰²³ On this site, $223B of revenue is 2.5 times Broadcom’s $89.1B and three-quarters of NVIDIA’s $303.0B.²⁴²³ I cannot build a path that takes Arm there. NVIDIA grew into an operating margin of 65.2%; Arm’s is 17% after stock pay, and its price already pays 37 times revenue. For most companies this big a tenfold in a few years is close to impossible, and I say the same of Arm.

Where I come down

Three things moved me: data center royalty that has at least doubled four quarters running, outside license revenue up 121% in the September quarter, and $3.6B of cash with no debt. They did not move my values, because stock pay still takes more than the business keeps, a seventh of revenue comes from the owner, and royalty growth slowed to 8%. What would raise my values is royalty growth back above 15% with the owner’s revenue steady, first chip deliveries on time at the stated margin, and owner cash that turns positive after stock pay; until then the price sits above my bull case.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• A controlled company. SoftBank can pass any ordinary resolution on its own, picks the board, and consents to share plans. Minority holders owning 13.6% can’t outvote it. The 20-F says SoftBank’s interests may differ from theirs.² Its margin loan is reported to mature in September 2027, within a year, and to have been renegotiated in September.¹⁸
• Regulators. Bloomberg reported on 16 May 2026, citing people familiar with the matter, that the US Federal Trade Commission is examining whether Arm will refuse or degrade CPU licenses while it builds its own chips, and that South Korean authorities inspected Arm’s Seoul offices in 2025. Neither the FTC nor Arm confirmed it. I cannot size this; the remedy could touch the terms on which licensees buy.¹³
• Memory prices and phone units. The chief financial officer blames higher memory prices for weaker mid and low-tier phone sales. Phones are 43% of royalty.¹⁴²
• Export rules. The Commerce Department’s “Affiliates Rule,” which would widen restrictions to companies owned 50% or more by listed entities, was due to return on 10 November 2026, as Arm’s own report, written before the new rule, said it could. A Bureau of Industry and Security final rule published on 9 November extends the suspension through 10 January 2027.³²³
• Chip supply and Taiwan. The Arm AGI CPU is made by third-party foundry partners, and Arm says any escalation in Taiwan could disrupt chip supply across the world.²³
• Customer concentration. Three customers were 42% of fiscal 2026 revenue and five were 57%.²
• Key people. Engineers are 84% of the staff and many are paid in stock. A fall in the share price lowers the pay of everyone Arm needs to keep, and raises the number of shares it must issue.²
• Estimated royalty. Arm accrues royalty from estimates and trues them up a quarter later, so a miss can show as a change in a later quarter.³
• Taxes. The tax charge was 13% of pre-tax income in the first half and 22% for fiscal 2026. Pillar Two minimum taxes and the sharing agreement with SoftBank affect the rate.²³