ACQ2: How ARM Became The World's Default Chip Architecture (with ARM CEO Rene Haas)

2025-04-03 - 74 min - source - Read full transcript

Ben Gilbert (host)David Rosenthal (host)Rene Haas

Key insights

RISC was arguably the more efficient architecture from the start, but x86/CISC won the PC era purely through software lock-in, not technical merit.

IBM's 1981 decision to build an open PC around Intel's 8086 and Microsoft's DOS (rather than in-house) let Compaq, Dell, and other clone makers build compatible machines; once Lotus 1-2-3 and other software was optimized for x86, switching architectures required a paradigm shift Rene estimates needs to be far more than a 10-15% improvement to be worth the software rewrite cost.

risc-vs-cisc-and-software-lock-in

Nearly every 1980s-90s CPU architecture besides x86 and ARM died off once venture capital abandoned semiconductor startups for software.

Rene lists a graveyard of well-engineered but now-extinct architectures (68000, Power PC, AMD 29000, DEC Alpha, SPARC); as SaaS and subscription software attracted the bulk of dot-com-era investment, new CPU-architecture funding dried up, leaving only the two platforms with enough accumulated software investment to survive.

risc-vs-cisc-and-software-lock-in

ARM's shared-success licensing model let it survive decades without needing a single dominant flagship win, unlike vertically integrated CISC competitors.

Rather than building and selling its own chips, ARM charges a modest upfront licensing fee (a proxy for R&D the customer skips) plus a royalty on shipped volume, aligning incentives with licensees; this let ARM accumulate design wins across many small, low-margin markets while vertically integrated rivals needed to win outright to survive.

arm-licensing-business-model

ARM strictly forbids custom, non-compliant instructions in architecture licenses to preserve a single unified software target.

Even customers building their own ARM-compliant chips (like Apple) cannot add proprietary instructions, because software developers would then have to guess which chips support which extensions and default to a 'lowest common denominator' approach anyway - undermining the whole point of a shared architecture.

arm-licensing-business-model

ARM's newer 'compute subsystems' business packages pre-verified CPU, interconnect, and memory-controller bundles with guaranteed performance targets on a specific fab process.

This goes beyond selling individual IP blocks (CPU cores, mesh networks like CMNs) toward something close to a virtual chipset: ARM will work directly with TSMC, Samsung, or Intel to guarantee a customer gets a specific clock frequency, saving customers three to nine months of integration engineering.

arm-licensing-business-model

ARM's first real design win came from a pure power/cost fit with no software ecosystem behind it at all.

Texas Instruments needed a small, low-power co-processor to run baseband functions in mid-1990s GSM phones and chose ARM purely because nothing else was as power-efficient; the resulting TI chip inside early Nokia phones gave ARM its initial foothold years before any real software or developer ecosystem existed on the architecture.

arm-origin-and-design-wins

The ARM-based iPod, not the iPhone directly, was the pivotal design lineage that made ARM the smartphone-era default.

When building the iPhone, Apple debated stripping down its Intel-based Mac OS (running on Intel's underpowered Atom chip) versus building up from the ARM-based iPod into a new operating system called iOS; the iPod team's clean-sheet, ARM-based approach won internally, and once Android followed, ARM became the smartphone standard by 2007-2008.

arm-origin-and-design-wins

ARM's data center rise rests on power efficiency, hyperscaler custom-silicon economics, and vanishing architectural alternatives compounding together.

Rene cites customers like Microsoft, Google, and AWS reporting roughly 60% performance-per-watt gains from custom ARM-based silicon; because ARM licenses to many vendors, hyperscalers with enough volume can build fully custom chips (memory, storage, interconnect) tailored to their own data centers - an ROI that isn't available in the two-vendor x86 world of Intel and AMD.

arm-ai-and-data-center-strategy

Haas argues the GPU-centric AI boom expands rather than threatens CPU demand, because every accelerator still needs a host CPU and training converts into far larger volumes of inference.

He rejects the investor framing that GPUs make CPUs obsolete, comparing it to removing tires and a steering wheel from a car with a bigger engine; because 'training is the teacher, inference is the student,' the resulting inference workloads will run across data centers, cars, headsets, and wearables - environments where a full GPU won't fit but a CPU always will.

arm-ai-and-data-center-strategy

NVIDIA's blocked 2020 acquisition of ARM is now viewed as underpaying, not overpaying, given how much ARM has grown since.

NVIDIA offered $40B for ARM in 2020 when NVIDIA's own market cap was $350-400B and ARM had roughly $2B in revenue versus NVIDIA's $25B; critics called the price excessive at the time, but ARM's revenue has since roughly quadrupled and its market cap has grown to about $150B, reversing that verdict in hindsight.

nvidia-arm-acquisition-attempt

The ARM/NVIDIA deal collapsed more from customer and ecosystem opposition than from a clean-cut antitrust violation.

Rene notes it was legally a vertical merger (NVIDIA and ARM don't compete directly), which regulators typically only block if it forecloses a market; because NVIDIA had no smartphone business at the time, the near-term case for blocking was weak, but broad opposition from ARM's other chip-industry customers and regulators' concern about long-term foreclosure ultimately killed the deal, which both Rene and Ben read as evidence of how structurally important a neutral ARM had become.

nvidia-arm-acquisition-attempt

NVIDIA's own multi-decade ARM relationship, including Rene running its Windows-on-ARM business around 2009-2011, previewed today's architecture transitions.

NVIDIA doubled down on ARM through its Tegra mobile chips, then through Microsoft's push to get Windows running natively on ARM (which Rene personally managed at NVIDIA), then through automotive (NVIDIA Drive) and ultimately Grace/Grace Blackwell data center CPUs - showing a single company progressively expanding an ARM bet across four distinct computing eras.

arm-origin-and-design-wins

Media referenced

ARM & SoftBank (Acquired, 2019) - podcast - Ben references Acquired's own 2019 episode on ARM's early history out of Cambridge University as the prequel to this conversation.
Rene Haas and Jensen Huang's podcast - podcast - Rene mentions recording a separate podcast with NVIDIA CEO Jensen Huang, where Jensen made the comment that 'software never dies,' a line Rene reuses to explain architecture lock-in.

Companies

ARM Holdings - The company Rene Haas runs; designs the CPU instruction set architecture and chip designs licensed into roughly 29 billion chips shipped in FY2024, now a ~$150B market cap company after IPO.
SoftBank - Took ARM private in 2016 for $32B, held it through the failed NVIDIA acquisition, and took it public again; Rene frames the $32B purchase and eventual re-listing as vindicated by ARM's subsequent growth.
NVIDIA - Attempted to acquire ARM for $40B in 2020 (blocked by regulators in 2022); Rene was a VP there before joining ARM and discusses NVIDIA's long ARM history via Tegra, Windows on ARM, and the Grace/Grace Blackwell data center chips.
Apple - Licenses the ARM instruction set to build its own CPUs (A-series, M-series); the ARM-based iPod, not a stripped-down Intel Mac OS, was the design lineage Apple chose to build iOS from, making the iPhone ARM's breakout design win.
Texas Instruments - ARM's first major design win in the mid-1990s: TI needed a small, low-power co-processor to run baseband functions in 2G/GSM phones and chose ARM because nothing else was as power-efficient.
Nokia - The end customer behind TI's baseband chips; the first Nokia GSM phones using TI's ARM-based chip gave ARM its initial foothold before any ARM software ecosystem existed.
IBM - Choosing Intel's 8086 and an external DOS from Microsoft for the 1981 IBM PC (rather than building in-house) created the open clone market that locked x86/CISC into the PC era, according to Rene.
Intel - Won the CISC/x86 design win inside the IBM PC and has remained one of only two vendors (with AMD) building x86 chips; Rene contrasts its closed, vertically integrated model with ARM's licensed, multi-vendor model.
Microsoft - Chosen by IBM to supply DOS over Gary Kildall's company; later drove Windows on ARM starting around 2009-2011, a project Rene personally ran the NVIDIA side of.
TSMC - Referenced as the leading-edge fab whose massive capex requirements Rene and David compare to the software investment moat that keeps ARM and x86 as the only two surviving CPU architectures.
Samsung - Built the chip inside the first iPhone under an ARM license and is one of several vendors (with Qualcomm, MediaTek, Tesla, Amazon) that license ARM designs or the ARM ISA directly.
Qualcomm - One of several ARM-licensed chip vendors named repeatedly as evidence of ARM's multi-sourced ecosystem, versus x86's effective duopoly of Intel and AMD.
Tensilica - A late-1990s semiconductor CPU startup (founded by ex-Synopsys and ex-MIPS engineers, later acquired by Cadence) where Rene worked from 1997-2004; cited as one of the last funded processor-architecture startups before VC money shifted entirely to software.
Cadence - Acquired Tensilica around 2012; also cited generally as one of the backend chip-design tool vendors (alongside Synopsys) whose maturing tools in the early 1990s made ARM's licensing model viable.

Techniques and frameworks

RISC vs. CISC - The foundational architecture distinction: reduced instruction set computers (RISC, ARM's lineage) trade fewer, simpler instructions for lower power and smaller memory footprint, versus complex instruction set computers (CISC, x86) that carry legacy instruction baggage.
Shared-success licensing model - ARM's founding business model: a modest upfront license fee (a proxy for R&D the customer no longer has to fund in-house) plus a per-unit royalty once the customer ships in volume.
Architecture license vs. implementation license - Customers can either license ARM's pre-built CPU 'implementation' (the blueprint) or an 'architecture license' that lets them design their own ARM-compliant core in-house, as Apple does; ARM never lets licensees add non-compliant custom instructions.
Compute subsystems - ARM's newer packaged offering: pre-integrated, pre-verified bundles of CPU cores, coherent mesh network interconnect, and memory controllers, sometimes with a guaranteed frequency target on a named fab process - described as 'Lego blocks' versus a finished statue.
Training is the teacher, inference is the student - Rene's framing for why AI compute demand favors ARM: training happens in concentrated data centers, but inference runs everywhere (phones, cars, wearables), so inference workloads will vastly outnumber training workloads.

Summary

Rene Haas, CEO of ARM Holdings, joins Ben Gilbert and David Rosenthal to trace how a company founded in 1990 Cambridge to power the Apple Newton became the architecture inside roughly 29 billion chips shipped last year - about four for every person on earth. The conversation opens by cataloguing the sheer ubiquity of ARM (cars, doorbell cameras, refrigerators, TVs, game consoles, phones) before rewinding to the RISC-versus-CISC debate of the 1970s-80s: RISC's simpler instruction set promised lower power and smaller memory footprints, but IBM's 1981 decision to build an open PC around Intel's 8086 and Microsoft's DOS locked x86/CISC into the PC era through pure software lock-in, regardless of which architecture was technically superior.

ARM's own path ran through the margins of that world. Rather than competing head-on for PC or mainframe sockets, ARM built a shared-success licensing model - a modest upfront fee plus a volume royalty - that let it accumulate design wins in low-power niches nobody else wanted: first Texas Instruments' baseband co-processor for Nokia's early GSM phones, then Symbian-era smartphones, and critically the ARM-based iPod, whose architecture Apple chose to build iOS from rather than stripping down its Intel Mac OS. Once Android followed suit around 2007-2008, ARM became the smartphone-era's default architecture, effectively playing Intel's historical role but licensed to many chipmakers (Samsung, Qualcomm, MediaTek, Apple itself) instead of just one.

A structural thread runs through the whole conversation: as venture capital abandoned semiconductor architecture startups for software after the dot-com boom, nearly every CPU architecture besides x86 and ARM (68000, Power PC, AMD 29000, DEC Alpha, SPARC) quietly died off, leaving two survivors by attrition as much as merit. That dynamic now plays out again in the data center, where Haas argues ARM's advantages - power efficiency, an open multi-vendor model that lets hyperscalers like Microsoft, Google, and AWS build fully custom silicon (getting roughly 60% performance-per-watt gains), and the near-total absence of remaining alternatives - explain why AI-era demand keeps compounding in ARM's favor. He pushes back hard on the idea that GPU-centric accelerated computing threatens CPUs, arguing instead that every accelerator still needs a host CPU and that inference workloads (spread across every edge device, unlike concentrated training runs) will vastly outnumber training workloads.

The episode closes on two forward-looking threads: ARM's newest "compute subsystems" business, which packages pre-verified CPU, interconnect, and memory bundles with guaranteed clock-frequency targets on named fabs - a step closer to a finished chip design without ARM becoming a manufacturer - and NVIDIA's blocked 2020 attempt to buy ARM for $40B. Haas, a former NVIDIA VP himself, walks through why that deal died (a legally vertical merger that nonetheless drew fierce customer and regulator opposition) and notes the wry irony that the $40B price once called an overpay now looks like a bargain given ARM's subsequent growth to a $150B market cap.

Notable Quotes

"A CPU is only as good as the software that's written on it and how long that software survives." - Rene Haas

"You need the 10x advantage to make the switch. I'm not sure it's 10x, but it's not 15%. It's got to be something that's quite material." - Rene Haas

"If 1981 is 2007, ARM is Intel, except the benefit that ARM has is that instead of Intel being Intel... ARM is licensing the architecture." - Rene Haas

"Training is the teacher, inference is the student. There are far more students than teachers in the universe." - Rene Haas

"There was a lot of opposition that we got from regulators, customers, ecosystem partners... maybe the company is more important than folks had originally gave us credit for." - Rene Haas