Question: The ARM deep-dive (2026-05-25-arm-deep-dive, section (b)/(d)) used order-of-magnitude per-chip royalty ranges — $0.30–0.50 smartphone, $1–5 PC, $20–50 datacenter — explicitly flagged as "industry channel-check estimates, not from ARM disclosure." Can those be bounded against actually-disclosed data: ARM's F-1 royalty economics, the Apple-payment report, and the ARM v Qualcomm trial figures? (Open follow-up filed 2026-05-25; closes the last of five ARM deep-dive follow-ups.)

Verdict: The deep-dive's ranges are broadly defensible — two refinements. (1) The empirical invariant ARM actually discloses is a rate, not a dollar: royalty ≈ 1.7% of chip TAM value, with management's "typical" rate 1–2% of chip ASP. Per-chip dollars are just that rate × the chip's ASP. (2) Architecture-license (ALA) customers pay far below the 1–2% headline because they pre-pay via upfront fees — Apple's <$0.30 on a ~$130 chip is an ~85% discount to the list rate (0.23% effective). That makes the $20–50 datacenter range the least-grounded number in the deep-dive: it assumes near-list rates, but AWS/NVIDIA/Microsoft hold ALAs and could carry Apple-like discounts, which would pull effective datacenter $/chip lower (offset by $2–5K server ASPs keeping the absolute dollars large). Smartphone should widen slightly to $0.30–0.80. Blended-across-all-chips is **$0.05/chip** (FY2023) rising to ~$0.08 (FY2026) — confirming ARM's whole thesis is mix-shift to high-ASP chips, not rate hikes.

What we're asking

The deep-dive's royalty math (section (b)) and the friend-not-foe captive-CPU argument (section (d)) both rest on the per-chip dollar ranges. If those ranges are wrong by an order of magnitude, the datacenter mix-shift arithmetic ("1 server CPU = 50–100 smartphone SoCs in royalty") breaks. This follow-up pins the ranges to disclosed evidence and isolates which segment is most/least reliable.

What we found

Anchor 1 — ARM's only hard self-disclosure is a rate, not a $/chip (ARM F-1, FY2023)

ARM does not publish a per-chip dollar figure. What it does disclose:

• "Royalties are generally either set as a percentage of the licensee's average selling price per chip or as a fixed amount per chip." Rates decline as volume rises, subject to a contractual minimum royalty per chip. (F-1, Royalty Revenue section.)
• "We estimate that our royalty revenue as of December 31, 2022 represented approximately 1.7% of the industry TAM containing Arm-based chips." — the single most useful aggregate invariant. ARM captures ~1.7% of the value of all chips containing its IP.
• Higher value per chip = higher rate: "a chip with an Arm CPU based on the latest Armv9 architecture and with eight cores, plus an Arm GPU, will typically have a much higher royalty rate than a chip with a dual-core Arm CPU core based on an older architecture."
• Mix concentration: "Royalties from smartphones and consumer electronics comprised greater than 50% of our royalty revenue for the fiscal year ended March 31, 2023."

The "typical 1–2% of ASP" framing is corroborated by secondary reporting (Gigazine/MacRumors, Nov 2023, citing ARM's licensing structure).

Anchor 2 — the blended $/chip, computed from disclosed totals (ARM F-1)

Cumulative Arm-based chips shipped: >250 billion (since inception, per F-1). FY2023 total revenue $2,679M, gross margin 96%, op margin 25%.

The blended number (~$0.05–0.08) is the floor anchor and the most reliable single figure. It is dominated by the tens of billions of low-royalty embedded/IoT/microcontroller chips. The rise from $0.055 → $0.077 over three years is the entire ARM bull thesis in one number: not a rate hike, but mix-shift toward higher-ASP chips (v9, more cores, datacenter). Cross-check: $1,675M ÷ 0.017 ÷ 30.6B ⇒ implied aggregate chip ASP ≈ $3.22 — consistent with an embedded-dominated mix.

Anchor 3 — Apple is the disclosed low-rate outlier (report Nov 2023)

• Apple pays ARM less than $0.30 per chip — ARM's lowest rate, per a widely-cited Nov-2023 report (MacRumors / Tom's Hardware / Gigazine, sourcing an analyst note).
• On a ~$130 A17 Pro SoC, the 1–2% list rate would imply $1.30–$2.60; Apple pays ~$0.30 ⇒ ~0.23% effective — roughly an 85% discount to the headline rate.
• Apple and ARM signed a long-term architecture agreement extending beyond 2040 (Sept 2023), far past the usual ~5-year term. (The "perpetual royalty-free license" claim is a myth — Apple does pay, just very little; The Chipletter debunk.)

This is the key refinement to the deep-dive's section (d). The deep-dive correctly said ALA customers "pay larger upfront fees in exchange for lower per-chip rates" — but the magnitude of the discount (≈85% below list, for ARM's largest unit customer) is larger than the body implied. Architecture licensees can pay a small fraction of the 1–2% list rate.

Anchor 4 — Qualcomm gives a total-payment anchor, not a clean per-chip (ARM v Qualcomm trial, Dec 2024)

• Qualcomm pays ARM ≈ $300 million per year in total fees+royalties (trial reporting, Tom's Hardware).
• Qualcomm's per-chip ALA royalty was far below Nuvia's ("many multiples" lower per trial testimony); ARM CEO Rene Haas testified the lower rate cost ARM ≈ $50M/year in revenue. Nuvia was acquired by Qualcomm for $1.4B (2021).
• Caveat — do not conflate: the often-cited Qualcomm "2.275% / 3.25% of handset ASP" figures are Qualcomm's own cellular-patent royalties charged to handset OEMs — not what Qualcomm pays ARM. The ARM-relevant figure is the ~$300M/yr total.

Rough cross-check: ~$300M ÷ Qualcomm's order-of-500M chips/yr ≈ **$0.60/chip blended for Qualcomm**, weighted toward premium Snapdragon — consistent with a smartphone-AP range around $0.40–0.80.

Bounded per-chip estimates (refined)

Verdict + reasoning

The deep-dive's ranges survive empirical bounding, with two corrections worth folding back.

1. What ARM actually sells is a rate (~1–2% of ASP; 1.7% blended), not a fixed toll. Every per-chip dollar figure is that rate applied to a chip's ASP. This strengthens the deep-dive's core mix-shift thesis: ARM's revenue grows as chips get more expensive and more core-dense, even with flat unit counts. The blended $/chip rising $0.055 → $0.077 in three years is direct evidence of the mix-shift already in the numbers.

2. Smartphone: widen to $0.40–0.80 (Apple anchors a <$0.30 floor on its favorable legacy ALA; Android/Qualcomm sits ~$0.40–0.80). The deep-dive's $0.30–0.50 was right for Apple, slightly low for the Android/Qualcomm side.

3. Datacenter $20–50 is the single least-grounded number in the deep-dive and should be flagged as such. Two opposing forces: (a) server-CPU ASPs of $2–5K make even a small rate a large absolute dollar; but (b) AWS/NVIDIA/Microsoft are architecture licensees, and the Apple precedent shows ALA customers can negotiate ~85% below the list rate. If hyperscalers hold Apple-like terms, effective datacenter royalty could be $5–25/chip, not $20–50. ARM's confirmed non-disclosure of datacenter royalty in dollar terms (see deep-dive addendum, 2026-05-28) means this stays an estimate — the bear-case point "captive-CPU per-chip ASP is opaque" is the correct posture. The "1 server CPU = 50–100 smartphone SoCs" framing still holds directionally (even $15/server ÷ $0.55/phone ≈ 27×; $50 ÷ $0.40 = 125×), but the multiple has a wide confidence band.

Net: No thesis change. The datacenter mix-shift arithmetic is intact at order-of-magnitude. The refinement is epistemic — we now know which number (datacenter $/chip) is soft and why (undisclosed + ALA-discountable), which sharpens what to watch: any ARM disclosure of datacenter royalty in dollars, or any hyperscaler ALA-terms leak, would be the highest-value future datapoint.

Routed back to the deep-dive

Section (b) per-chip block and section (d) captive-CPU argument refined inline in the May 25 ARM deep dive (addendum extended). No watchlist/perspective mutation (ARM coverage already correct; thesis unchanged).

Sources

• ARM F-1 (FY2023, filed 2023-08-21): SEC EDGAR Archives/edgar/data/0001973239/000119312523216983/d393891df1.htm — royalty-revenue figures ($1,675M FY23 / $1,562M FY22 / $1,313M FY21), >30.6B chips FY23, >250B cumulative, "1.7% of the industry TAM", "% of ASP or fixed per chip / declining with volume / minimum per chip", v9+cores = higher rate, "smartphones + CE >50% of royalty FY23", total revenue $2,679M / 96% GM / 25% OM. Downloaded locally to /tmp/arm-f1.htm for extraction.
• Apple <$0.30/chip: MacRumors (2023-11-29) "Apple Pays Less Than 30 Cents in Royalties to Arm Per Chip"; Tom's Hardware / Gigazine (2023-11-30) corroborating; The Chipletter "Apple's Perpetual, Royalty-Free License and Other Arm Myths" (debunk of the royalty-free claim + beyond-2040 agreement).
• ARM v Qualcomm trial (Dec 2024): Tom's Hardware "Arm vs. Qualcomm trial begins" — Qualcomm ~$300M/yr to ARM, Nuvia $1.4B acquisition, ~$50M/yr revenue impact, Nuvia rate "many multiples" of Qualcomm ALA; Computerworld / The Register on the mistrial outcome.
• Qualcomm own-patent royalty rates (2.275% / 3.25%, do-not-conflate caveat): Qualcomm 5G Handset Licensing Program PDF — these are Qualcomm-to-OEM cellular royalties, NOT Qualcomm-to-ARM.
• Deep-dive being refined: 2026-05-25-arm-deep-dive (section (b)/(d) + Disclosure addendum 2026-05-28).