Arm: The Silicon Blueprint - [Business Breakdowns, EP.200]

Summary

• Goldberg’s core frame: Arm is the “blueprint” licensor of chip design—the standard bathroom plumbing every architect can copy and paste so customers differentiate elsewhere. Qualcomm, NVIDIA, and Broadcom license Arm’s instruction-set IP, pay an upfront license fee plus a per-chip royalty, and build their differentiation—modems, graphics, and AI—on top. Zack Fuss’s setup put the stakes at nearly $5B of run-rate revenue and a roughly $150B market cap, or about 30 times sales.
• The AI story is real but indirect: large GPU systems also need CPUs, and NVIDIA’s Grace CPU is Arm-based. Goldberg’s caution: “Arm’s valuation is justified by more than just AI—it’s this broad expansion into so many other things,” above all data center, where the “super 7” hyperscalers—Amazon, Microsoft, Google, Facebook, Alibaba, Baidu, and Tencent—are designing their own Arm-based CPUs as alternatives to expensive Intel/AMD parts.
• The value-capture thesis is central: Arm earned about 7 cents per chip in 2023, about 9 cents today, and Goldberg’s rough math gets to 12 cents over the next 3–4 years. On a $100,000 NVIDIA system, “Arm probably makes a buck or two”—so “maybe there’s a little room for a little bit more to go Arm’s way,” multiplied across end markets, with each nominal royalty increase flowing through to a bottom line supported by 90%+ gross margins and 40–50% operating margins.
• Rene Haas’s turnaround is the catalyst: after SoftBank paid $32B for Arm, “Arm went to sleep” for most of a decade; the failed $54B NVIDIA deal three years ago “woke SoftBank up.” Haas “fixed pricing, he fixed product, he fixed marketing,” including freemium-like tiers—lower upfront license payments and higher ongoing royalties—that get designers “in the door” and then almost lock them in, since switching architectures would take a major customer such as Qualcomm roughly a decade.
• The RISC-V threat is blunted, not dead. RISC-V is an open standard, not owned by one company and free to use. Goldberg says it “is not ready for data center workloads today”—years away from smartphones, let alone data centers—and Haas’s repricing “greatly reduced, maybe eliminated” its appeal for many use cases. Still, about 1,000 RISC-V-centered chip companies took root in China’s embedded/IoT market while Arm “took its eye off the ball.” Even Qualcomm, currently being sued by Arm, “can’t just go out tomorrow” and switch.
• The bear case isn’t competition—it’s the attach rate. Depending on whom you ask, the ratio of AI accelerators to Arm-based CPUs is about 2:1–8:1 today. If it goes toward 100:1, “the TAM, or total addressable market, is nowhere near as big as we would have expected and growth plateaus.” The bull case: Arm is “not quite a monopoly legally but functionally they’re pretty close” on ISAs and could “really flex their muscles” on royalty rates.
• Growth vectors beyond mobile: automotive is “the big opportunity”—a few hundred dollars of semiconductor content per car, growing double digits annually, before any autonomy—though 5–10-year product cycles mean patience. Goldberg’s closing lessons: R&D at 20–30% of revenue is non-negotiable; Arm underinvested in AI, tensor, and graphics cores over the last decade. Its ecosystem of hundreds of licensees is a powerful force multiplier.

Deep dive

1. The blueprint model: license the plumbing, let customers differentiate

• Goldberg’s opening analogy, worth keeping in full: an architect differentiates on curb appeal and light flow, not on “the best-designed bathrooms”—so you “take the generic blueprints for a standard bathroom and copy and paste those around your design.” Arm supplies that low-level plumbing—the basic math functions every chip needs—freeing Qualcomm to differentiate on modems and NVIDIA on graphics and AI processing.
• The mechanics: Arm makes no chips. It licenses IP to designers, who send designs to foundries like TSMC, and charges an upfront license payment plus a royalty per chip shipped. Zack Fuss’s setup numbers frame the stakes: roughly $150B market cap, sales approaching $5B, and “a rather robust 30 times revenue multiple.”

2. CPUs in an AI world: attach rate, not center stage

• The engineering trade-off spelled out: a CPU is the “general-purpose jack-of-all-trades” running the OS and low-level functions; a GPU does graphics, and now AI math, better. Arm sits at the heart of the general-purpose side: “in most compute systems you’re always going to need some control node,” even in NVIDIA systems linking 72 GPU cards, and NVIDIA probably prefers to sell its own attached Grace CPU, which is Arm-based.
• Goldberg resists making Arm a pure AI trade: there’s real content in CPUs and networking chips “adjacent to the core NVIDIA AI accelerators,” but “I personally think that Arm’s valuation is justified by more than just AI—it’s this broad expansion into so many other things.” Data center is “first and foremost their biggest growth opportunity”; the hyperscaler group includes Amazon, Microsoft, Google, Facebook, Alibaba, Baidu, and Tencent. IoT is “a messy market” where “nobody does stellar”; automotive looks “pretty good right now” but runs on 5–10-year design cycles versus one or two elsewhere.

3. From Acorn to the Matrix phone to the iPhone

• The origin story as told: Acorn Computers in the UK spun off its chip team into a three-way JV with VLSI and Apple—Apple wanted the IP for the Newton, which “didn’t do particularly well,” but “Arm actually did pretty well, made a lot of money from Newton.” Early customers included, “of all things,” fax machines.
• The inflection: Nokia and Texas Instruments adopted Arm, and “the first Arm-powered mobile phone was actually the 8110—the Matrix phone.” Arm’s power-efficiency reputation versus x86 made it appealing in feature phones; then 2007’s iPhone launch “really kicked it into high gear,” because suddenly phones needed very advanced processors and “the best way to do that was to start with an Arm IP at the core.”

4. RISC vs CISC—and why flexibility, not just watts, was the real prize

• Goldberg demystifies the old religious war: CISC (x86) bakes complex instructions like a square-root function into silicon; RISC strings simpler steps together. Cheap on-chip memory made RISC much more attractive, and its reduced instruction set contributed to its power-efficiency reputation, though Goldberg called that an oversimplification: “there are always trade-offs.”
• His deeper claim: what RISC really meant was licensee flexibility—“it wasn’t just the power savings… it was much more about the flexibility”—to optimize per application. Structurally, x86 is owned by two companies while “anybody with enough money can get an Arm license.” Hundreds of licensees “innovating and exploring the market and finding every segment and every niche” gave Arm a broader ecosystem than Intel and AMD could explore, and that volume “really powered the growth of TSMC.”
• RISC-V takes that flexibility further: it is an open standard, not owned by one company and free to use, though downstream software complexity is a drawback.
• Historically, not manufacturing was also a financial necessity: chip production involves foundry payments, mask sets, inventory, and working capital. Licensing let Arm avoid those burdens while building a broad ecosystem. Goldberg thinks Arm may now move up the stack—helping customers “harden” IP into manufacturable designs and getting “very, very close to doing a full design of their own chips”—though probably not selling a merchant-labeled part. The key compatibility issue is software: code written for Microsoft’s Arm CPU can be ported “fairly easily” to Amazon’s, after “a lot of blood, sweat and tears.”

5. SoftBank’s sleep, NVIDIA’s bet, and the Haas reawakening

• The corporate saga: SoftBank paid $32B about 10 years ago amid “a lot of head-scratching,” then “Arm went to sleep”—a mobile ISA monopoly without public-market pressure. NVIDIA’s $54B offer was, in Goldberg’s read, classic Jensen Huang: “his superpower in my mind is he’s willing to make big bets… he’s not afraid of failure.” Part of NVIDIA’s interest was its own CPU effort and desire to accelerate Arm’s data-center work, though Goldberg also thought the acquisition was “not necessarily fully thought out.”
• The deal’s collapse three years ago “woke SoftBank up”; new CEO Rene Haas “fixed pricing, he fixed product, he fixed marketing.”
• The value-capture arithmetic Goldberg keeps returning to: about 7 cents average royalty per chip at the 2023 IPO, about 9 cents today, and “rough math” of about 12 cents over the next 3–4 years. Against a $100,000 NVIDIA system yielding Arm “probably… a dollar or two,” the room for capture is obvious—and the pervasiveness is the point: “pretty much every electronic device in your house” has Arm content, from AirPods to thermostats to TVs. Even x86 PCs may contain Arm cores, perhaps for Bluetooth or Wi-Fi.

6. Software-like economics, a blunted RISC-V, and the attach-rate bear case

• The financial engine: 90%+ gross margins, 40–50% operating margins, and marginal cost of a royalty sale “essentially zero”—but Goldberg’s caveat matters: “this is not software… you can’t patch this. This gets baked in the chips.” The new freemium-style tiers—small upfront payment, higher royalties as usage expands—exploit the fact that “once you have that, they’re almost locked in.” Even Qualcomm, sued by Arm, has no easy exit: switching to RISC-V “would take them a decade.”
• On RISC-V, Goldberg says it “is not ready for data center workloads today.” Raw performance “maybe” is comparable, but the design ecosystem, hardening, and software mean “years and years” before RISC-V is ready for smartphones, let alone data centers. Its main foothold so far was China’s late-2010s chip boom, where about 1,000 RISC-V-centered companies emerged in embedded IoT while Arm “wasn’t quite ready for it and wasn’t totally paying attention.”
• The scenario tree: base case, steady incremental value capture; upside, Arm “really flexes its muscles” as a functional near-monopoly on ISAs and drives significant royalty increases; bear case, “a radical shift in how we do compute”—accelerator-to-CPU attach going from 2:1–8:1 today toward 100:1, shrinking the TAM so “growth plateaus.” That’s where Goldberg would focus bear-case work.
• Closing lessons, plus credit where due: invest 20–30% of revenue in R&D. Arm’s underinvestment in AI, tensor, and graphics cores is the counterfactual—“there’s definitely an alternate universe” where it mattered even more today. Goldberg credits three people with Arm’s success: Sophie Wilson, an often-unsung founder of the original design team who did not move to Arm when it spun off; founding CEO Simon Saxby, who established the business model; and Haas, who repositioned it for the next wave. The licensee ecosystem remains a major competitive force multiplier.