BULL & BEAR MEMO

Arm Holdings — Bull & Bear Memo

A thinking anchor, not a valuation or a recommendation.

1. Business in one line

Arm licenses CPU architecture and processor designs to chipmakers for an upfront fee, then collects a per-chip royalty — a percentage of the chip's selling price, or a fixed fee per unit — on substantially every chip shipped, for as long as it ships. Since March 2026 it also sells a data-center CPU of its own. FY2026 revenue: $4,920m — $2,613m royalty, $2,307m license and other, at a 98% gross margin.

Type: fast grower — three consecutive years above 20% revenue growth since the IPO — sitting on a cyclical royalty base whose units track the semiconductor cycle. The classification matters because Arm has no unit share left to win in its largest market: the bull case must be a pricing story, not a volume story.

2. Bull case — Peter Lynch pitch

The simple reason this stock could work is that Arm gets paid more each year for the same chip, and has just opened two new ways to get paid for chips it previously only rented cores to.

B1. Royalty per chip rises faster than the unit base decays

What must happen — Mix shift toward Armv9 and Arm Compute Subsystems lifts average royalty per chip fast enough that royalty compounds above 20% on a roughly flat unit base.

Why Arm specifically — Royalty rates step up with how much Arm content sits in the chip, and every key mobile operating system depends on Arm cores. A customer who wants a cheaper royalty has to move a software ecosystem, not swap a core — the switching cost sits with the customer, the price increase with Arm.

Evidence — Royalty of $2,613m in FY2026, +21%, attributed in the 20-F to “an improved mix of products with higher royalty rates per chip, such as Armv9 technology”; +22% to $715m in Q1 FY2027. Mobile applications processors, where Arm has held above 99% share for years, are still 43% of royalty — that share cannot grow, so the growth is rate.

What to monitor — Royalty growth versus industry unit growth; mobile's falling share of total royalty.

B2. The data center becomes a second royalty pool of comparable size

What must happen — Hyperscaler in-house Arm CPUs keep displacing x86 sockets, each carrying a far larger royalty than a phone.

Why Arm specifically — Royalty is a percentage of chip ASP or a fee that scales with content, so a 100-core server CPU pays a multiple of a smartphone SoC — and every AI accelerator still needs a host CPU.

Evidence — Data-center royalties more than doubled year over year in Q1 FY2027. Cumulative Neoverse shipments passed 1.5 billion cores — the most recent 500 million in nine months, against six years for the first billion. Named programs: NVIDIA Vera in full production, Google Axion hosting its TPU systems, AWS Graviton5 under a multi-year Meta agreement, Azure Cobalt 200, Qualcomm Dragonfly C1000.

What to monitor — Data-center royalty growth rate; Neoverse cumulative-core milestones; any named hyperscaler program slipping.

B3. Arm AGI CPU converts dollars of royalty into thousands of dollars of silicon revenue

What must happen — The chip launched in March 2026 ships in volume from end-calendar-2026 and holds its multi-generation roadmap with Meta and the other named customers.

Why Arm specifically — Arm already owns the cores, interconnect and software stack, so the incremental cost of selling the whole chip is physical design and wafers, not architecture — chip-level revenue in a socket that previously paid only a royalty.

Evidence — Committed demand rose from a stated “$1 billion opportunity” across FY2027–28 to “exceeds $2 billion” within one quarter, capacity secured for the $1bn case, initial product delivered to multiple customers. Meta is lead partner and co-developer across multiple generations. Claimed more than 2x performance per rack versus x86.

What to monitor — Conversion of the >$2bn demand into recognized revenue; the first disclosed gross margin on the silicon line; inventory and wafer commitments appearing on the balance sheet.

B4. Today's spending is an investment phase that ends

What must happen — R&D growth falls below revenue growth and GAAP operating margin recovers from single digits.

Why Arm specifically — Arm funds one core design and amortizes it across every licensee, so a fixed R&D base spread over a rising royalty stream is the operating leverage — nothing in the licensing model requires cost to scale with revenue.

Evidence — R&D was $2,776m in FY2026 — 56% of revenue — up 34% while revenue grew 23%. GAAP operating margin was 18% for FY2026 and fell to 7.1% in Q1 FY2027 from 10.8%, even as non-GAAP operating margin rose to 41.2%.

What to monitor — R&D growth versus revenue growth; GAAP — not non-GAAP — operating margin.

Why the market might be missing it: royalty is recognized a quarter in arrears on estimated shipments, so a data-center inflection reaches Arm's reported numbers materially later than it happens in its customers' fabs.

3. Bear case — Munger invert

The most likely way I lose money is that Arm's reported growth turns out to have been bought from its own parent, while the licensees it now competes with quietly fund the alternative.

R1. Arm now competes with the customers who pay its royalties

Attacks B1, B2, B3 · Permanent

How it could fail — The AGI CPU and Arm's direct hyperscaler relationships put it in the same socket as Broadcom, Marvell, Qualcomm and other licensees. Arm's own 20-F says these customers may “seek alternative architectures”, “withhold sensitive roadmap information” or “demand more favorable commercial terms”. A licensee cannot punish Arm this quarter — it starts a RISC-V program that removes a socket in five years.

What would confirm failure — An architecture licensee announcing a non-Arm core for a segment where it currently ships Arm; licence renewals on visibly worse terms.

Damage — Permanent: the royalty base is the entire durable asset, and an architecture that leaves does not come back.

R2. Arm cannot enforce the licence terms that make architecture licensing safe

Attacks B1 · Largely permanent

How it could fail — Arm sued Qualcomm and Nuvia to force destruction of cores developed under a terminated architecture licence. The jury found the technology licensed under Qualcomm's own ALA, and in September 2025 the court granted Qualcomm judgment as a matter of law that Nuvia had not breached. Arm's appeal is pending; Qualcomm's counter-suit goes to trial in Q4 CY2026. If that stands, an architecture licence becomes a cheap option: acquire a design house, keep the cores, pay the architecture rate rather than the implementation rate.

What would confirm failure — The Third Circuit affirming; an adverse Q4 CY2026 verdict; other licensees restructuring toward architecture-only terms.

Damage — Qualcomm alone was 9% of FY2026 revenue; the larger loss is the price ceiling — implementation royalties hold only if the cheaper route stays legally closed.

R3. Most of FY2026's growth came from the controlling shareholder

Attacks B4 and the headline growth rate · Fixable, but it has already repriced the growth rate

How it could fail — Revenue under the consulting agreement with a SoftBank Group affiliate was $704.4m in FY2026 against $145.5m in FY2025 — $559m of the $913m total revenue increase, and 14% of FY2026 revenue, from an entity controlled by the shareholder that owns 86.4% of Arm. Strip it out and FY2026 revenue grew about 9%, with license and other revenue down about 5% (arithmetic from the disclosed figures, not a disclosed number). $645.8m sat unbilled at 31 March 2026 — roughly two-thirds of all contract assets.

What would confirm failure — The consulting line flattening while ACV growth stays in the low teens — ACV, Arm's own normalized licensing metric, already decelerated from +22% at March 2026 to +13% at June 2026.

Damage — Less to cash flow than to the multiple: the market has been capitalizing a 23% grower whose arm's-length business grew about 9%.

R4. Production silicon destroys the financial model the multiple is paid for

Attacks B3 · Permanent if the silicon business succeeds

How it could fail — Arm's 98% gross margin exists because it ships files, not wafers. The 20-F warns the silicon business brings “margin compression, greater quarter-to-quarter revenue volatility, increased capital intensity and working capital requirements”, plus foundry commitments that cannot flex, inventory obsolescence and a longer cash conversion cycle. At $2bn of silicon against roughly $5bn of IP revenue, blended gross margin falls mechanically however well Arm executes.

What would confirm failure — The first quarter reporting inventory, wafer purchase commitments and a gross margin below the mid-90s.

Damage — Structural, and realized by the bull case working rather than failing — which is why it is the easiest risk to underwrite away.

R5. Stock compensation already consumes the operating profit

Attacks B4 · Ongoing; fixable only by slowing hiring

How it could fail — Equity-settled share-based compensation was $1,052m in FY2026 against GAAP operating income of $900m; with employer taxes, $1,212m, or 25% of revenue. The 41.2% non-GAAP operating margin quoted in Q1 FY2027 was a 7.1% GAAP margin. Diluted shares rose only from 1,044m to 1,068m over two years — the cost lands in the income statement rather than the share count, real in either place.

What would confirm failure — GAAP operating margin staying in single digits as silicon cost of sales lands on top.

Damage — Caps the operating leverage in B4 — R&D can decelerate and GAAP margin still not recover if equity comp scales with engineering headcount.

R6. The two largest revenue concentrations are entities Arm does not control

Attacks B1 · Structural, outside Arm's control

How it could fail — Arm China — an independent distributor Arm neither owns nor manages — is Arm's single largest customer at 16% of FY2026 revenue. Arm depends on Arm China's own reported figures to determine what it is owed, and recognized $12.3m of expected credit losses against it in FY2026, taking the allowance to $28.3m. PRC revenue was 18% of total; the top five customers, including Arm China and SoftBank Group, were 57%.

What would confirm failure — A China export-control or end-use change touching the IPLA; a widening credit allowance against Arm China; PRC revenue turning negative after +17% in FY2026.

Damage — A sixth of revenue reaches Arm through a counterparty it cannot audit and a government that can close the channel.

4. Signals to monitor

5. External challenge notes

Not run. This memo is built only from Arm's own filings and shareholder letters. Ask for “external challenge” to add competitor, regulatory, litigation and short-seller findings from outside the source set — kept in this section so company evidence and outside evidence stay separable.

6. Bottom line

The stock could work because Arm is repricing a socket it already owns — royalty per chip rising above 20% a year on a flat unit base — while the data center adds a structurally higher-value royalty pool and the AGI CPU turns some sockets from a royalty into a whole chip sale.

For that, Armv9 and CSS mix must keep lifting royalty per chip, Neoverse adoption at AWS, Google, Microsoft, NVIDIA and Qualcomm must keep compounding, the >$2bn of AGI CPU demand must convert to shipped revenue from end-2026, and R&D growth must eventually fall below revenue growth.

The thesis most likely breaks not in a bad quarter but through licensees responding to Arm becoming their competitor — funding RISC-V or in-house architectures for sockets they currently pay royalties on — a decision invisible for years and irreversible once made.

I would change my mind if the SoftBank consulting line flattened while ACV growth stayed in the low teens — putting organic growth nearer 9% than 22% — if a major architecture licensee announced a non-Arm core in a segment it ships Arm in today, or if the AGI ramp reset consolidated gross margin below the mid-80s.