The Barn and the $150 Billion Question
On September 14, 2023, at an initial public offering price of $51.00 per share, a company that does not manufacture a single chip — that has never fabricated a transistor, never operated a foundry, never sold a processor to an end consumer — debuted on the Nasdaq with a valuation north of $54 billion. Within five months, that number had more than doubled. By early February 2024, after a quarterly earnings report that showed licensing revenue up more than 100% year-over-year and a guidance upgrade that blindsided even the bulls, Arm Holdings plc was worth $116.8 billion — adding $37.6 billion in a single trading session, a figure roughly equivalent to the entire market capitalization of Experian. The stock traded at nearly 90 times forward earnings, more than double Nvidia's multiple and almost twice AMD's. For a company whose annual revenue at the time of its IPO was approximately $2.68 billion, these numbers seemed to violate the laws of financial physics.
But consider what Arm actually is: the company estimates that 70% of the world's population uses its technology. More than 325 billion chips based on its architecture have shipped since the early 1990s. Every iPhone, every Android device, every Nvidia Grace server CPU, every AWS Graviton processor, every chip in your car's infotainment system and your refrigerator's controller and your thermostat's brain — all of them speak the same fundamental language, an instruction set architecture born in a converted barn outside Cambridge, England, where twelve engineers set out to design a computer that could run on a battery. Arm does not sell the shovels in the gold rush. It licenses the metallurgy.
By the Numbers
The Architecture of Everything
325B+Arm-based chips shipped to date
$3.93BFY2025 total revenue
~$150B+Market capitalization (late 2025)
70%Global population using Arm-based products
22M+Software developers building on Arm
7.1BArm-based chips shipped per quarter (Q2 FY24)
~90%SoftBank's ownership stake post-IPO
The paradox at the center of Arm's existence is this: it may be the single most consequential technology company that almost nobody outside the semiconductor industry could name. Its architecture is more ubiquitous than Windows, more pervasive than Google Search, more embedded in daily life than any product from any company on earth. And yet its revenue, even after the AI-driven surge, remains a rounding error compared to the trillions of dollars of economic activity its designs enable. The company's entire business model is a bet — maintained for over three decades — that it is better to be the standard than the manufacturer, better to be the lingua franca than the speaker. Whether that bet can sustain a valuation that implies Arm is among the most valuable semiconductor companies in the world, despite having revenue roughly one-fifteenth of Nvidia's, is the question the market is now pricing in real time.
The Acorn That Became the Forest
To understand Arm you must first understand a machine almost nobody in America has heard of: the BBC Micro. In 1981, the British Broadcasting Corporation launched a national computer literacy initiative and needed a home computer to anchor it. Acorn Computers, a small Cambridge outfit founded by Hermann Hauser and Chris Curry, won the contract. The BBC Micro became an emblem of British computing — sold in schools, featured on television, beloved by a generation of programmers who would go on to build much of the UK's technology industry. But Acorn had a problem. By the mid-1980s, the company needed a next-generation processor for its machines. The existing options — Intel's x86, Motorola's 68000 — were too expensive, too power-hungry, or too encumbered by licensing complexity. So Acorn's engineers decided to build their own.
The critical figure here is Sophie Wilson, one of the most quietly important computer scientists of the twentieth century. Wilson, who had already designed the instruction set for the BBC Micro's 6502-based system, led the design of what would become the Acorn RISC Machine — ARM. The philosophy was radical in its simplicity: a reduced instruction set computing (RISC) architecture that prioritized efficiency over raw power, doing less per clock cycle but doing it faster and with far less energy. The first ARM1 processor, completed in 1985, contained roughly 25,000 transistors — compared to the hundreds of thousands in Intel's contemporary 386 — and drew so little power that the first prototype reportedly worked on the first attempt. Wilson and her colleague Steve Furber (whose book
ARM System-on-Chip Architecture remains the definitive technical reference) had built something that didn't just work. They had stumbled onto a principle that would take two decades to reveal its full significance: in a world increasingly powered by batteries, the most valuable architecture isn't the most powerful one. It's the most efficient one.
Acorn, however, was not destined to be a great company. The BBC Micro's commercial moment passed, the PC clone market swallowed the home computing category, and by 1990, Acorn needed to find a way to keep the ARM project alive without the overhead of a failing computer manufacturer. The solution was a joint venture: Acorn, Apple Computer, and VLSI Technology each invested roughly $3 million to create Advanced RISC Machines Ltd. — twelve people in a barn in Swaffham Bulbeck, a village outside Cambridge. Apple needed a low-power processor for the Newton PDA. VLSI wanted a licensable core to sell. Acorn wanted to spin off its most promising technology. The new company's first CEO was Robin Saxby, a semiconductor industry veteran who would prove to have one of the most consequential strategic intuitions in the history of technology.
The Decision Not to Make Things
Saxby's insight was deceptively simple, and it went against every instinct in the semiconductor industry of the early 1990s. Intel made chips. AMD made chips. Texas Instruments made chips. The entire industry was organized around the assumption that if you designed a processor, you manufactured and sold it. Saxby decided that ARM would do neither.
We were going to be a global semiconductor intellectual property company. We would license our technology to anyone who wanted to use it, and they would pay us a royalty for every chip they shipped.
This was, in 1990, a business model that essentially did not exist. There was no "IP licensing" category in the semiconductor industry. The few companies that licensed processor designs did so as a sideline, not as their entire reason for being. ARM would design processor cores, license those designs to chipmakers — who would then integrate the ARM core into their own system-on-chip products — and collect two revenue streams: an upfront license fee for access to the design, and a per-unit royalty on every chip shipped. The royalty would be small — typically one to two percent of the chip's selling price, sometimes even less — but it would be collected on volume. Every chip, forever.
The genius of this model — and the reason it took so long for the market to appreciate it — is that it aligned ARM's interests with the broadest possible adoption of its architecture. Intel's business model rewarded exclusivity: the higher Intel's margins, the more money Intel made, and the fewer competitors existed, the better. ARM's model rewarded ubiquity. Every new licensee, every new market, every new device category expanded the royalty base. ARM wanted its architecture inside everything — phones, printers, hard drives, cars, industrial controllers, set-top boxes — and the way to achieve that was to make the licensing terms so attractive, and the power consumption so low, that choosing ARM became the path of least resistance for any chipmaker building a product that needed to run on a battery.
The early years were lean. The Apple Newton, which had been one of the catalysts for ARM's creation, was a commercial failure. The company's revenue in its first years was measured in single-digit millions of pounds. But Saxby and his team were patient accumulators. They licensed the ARM7TDMI core to Texas Instruments, which put it into the Nokia GSM handsets that were just beginning to proliferate across Europe. They licensed to Cirrus Logic, Samsung, Sharp, and dozens of other companies. Each deal was individually small. Collectively, they were building something unprecedented: a platform standard that no single company controlled but that ARM defined.
The Mobile Bet Pays Off
The mobile phone — specifically, the smartphone — is the event that turned ARM from a clever licensing business into the most pervasive computing architecture in history.
Key inflection points in Arm's architecture adoption
1990ARM Ltd. founded as a joint venture of Acorn, Apple, and VLSI Technology. Twelve employees in a Cambridge barn.
1993ARM7TDMI core licensed to Texas Instruments for Nokia GSM phones.
1998IPO on the London Stock Exchange and Nasdaq. Market cap approximately £1.8 billion.
2004ARM11 processor powers the first mass-market smartphones.
2007Apple selects ARM-based Samsung chip for the original iPhone.
2008Google launches Android, built on ARM architecture.
2011ARM-based chips reach 7.9 billion units shipped annually.
When
Steve Jobs stood on stage in January 2007 and introduced the iPhone, the device that would reorganize the technology industry around a glass rectangle, the processor inside was a Samsung-designed system-on-chip built on ARM's architecture. It was an ARM core not because Apple loved ARM, but because nothing else could deliver acceptable performance inside a battery-constrained form factor. Walter Isaacson's
Steve Jobs captures Jobs's obsessive focus on thinness and battery life — requirements that, computationally, led inexorably to ARM. When Google launched Android the following year, it too was built for ARM. The entire smartphone revolution — two billion devices sold annually at its peak — ran on ARM's instruction set.
The scale of this is difficult to overstate. By the time Arm filed its F-1 in August 2023, it reported that more than 250 billion Arm-based chips had been shipped since the company's founding. Roughly 70% of the world's population used Arm-based products daily. In smartphones, Arm's market share was functionally 100%. Not "dominant." Not "leading." Total. Every meaningful mobile processor — Apple's A-series and M-series, Qualcomm's Snapdragon, Samsung's Exynos, MediaTek's Dimensity — was built on Arm's instruction set architecture.
And yet the financial paradox persisted. On each of those billions of chips, Arm collected royalties measured in pennies. The company's total revenue for fiscal year 2023 (ending March) was approximately $2.68 billion. Apple alone generated more revenue in a single week. Arm was the most important company in the mobile ecosystem, and it was also, by revenue, one of the smallest. The license-and-royalty model that guaranteed ubiquity also capped monetization. Every chip that shipped was validation of the architecture's dominance. Every quarterly revenue report was a reminder of how cheaply that dominance had been sold.
Son's Gambit
Masayoshi Son saw something in Arm that the public markets had undervalued for years. The SoftBank founder — a man whose career is a series of enormous, conviction-driven bets, from Yahoo Japan to Alibaba to Sprint — believed that the Internet of Things would put Arm's architecture inside a trillion devices. Not billion. Trillion. In July 2016, Son offered £24.3 billion (approximately $32 billion) for Arm Holdings, a 43% premium to the pre-announcement share price. It was the largest-ever acquisition of a European technology company.
The reaction in the UK was somewhere between horror and resignation. Arm was the crown jewel of British technology — the rare post-industrial success story, a company that had proven you could build a globally essential semiconductor business from Cambridge without fabrication facilities or American-scale capital. Son promised to keep Arm's headquarters in Cambridge, to double the UK headcount within five years, and to maintain the company's neutrality among licensees. He kept most of those promises. But the acquisition also removed Arm from public market discipline and placed it inside SoftBank's portfolio, where it sat alongside WeWork, OYO Rooms, and a menagerie of Vision Fund bets whose connection to Arm's careful, engineering-driven culture was approximately zero.
Son's IoT thesis, in its original articulation, did not play out as predicted. The trillion-device internet of things has arrived more slowly and at lower average selling prices than the most optimistic projections. But Son — in one of those reversals that make his career so difficult to evaluate from the outside — landed on the right macro thesis anyway. The demand driver for Arm's architecture turned out not to be IoT refrigerators but something far more consequential: artificial intelligence.
NVIDIA's $40 Billion That Never Was
In September 2020, Nvidia announced it would acquire Arm from SoftBank for $40 billion, in a deal structured as a mix of cash and Nvidia stock.
Jensen Huang, Nvidia's CEO, described the combination as "the company of the age of AI." The logic was compelling: Nvidia's GPU dominance in AI training combined with Arm's CPU ubiquity across inference and edge devices would create an unmatched computing platform. Arm's designs were already inside Nvidia's own Grace CPU, its automotive chips, and its data center networking products.
The regulatory response was ferocious. Arm's entire value proposition depended on being a neutral platform — Switzerland in the semiconductor industry. If Nvidia owned Arm, every Nvidia competitor that depended on Arm's architecture (which was essentially every semiconductor company on earth) would face a potential conflict of interest. The UK's
Competition and Markets Authority, the European Commission, the U.S. Federal Trade Commission, and China's antitrust regulators all signaled opposition. Qualcomm, Google, Microsoft, and others lobbied aggressively against the deal. In February 2022, after sixteen months of regulatory review, the acquisition was formally abandoned. Nvidia paid SoftBank a $1.25 billion breakup fee.
The failed deal revealed something important about Arm's structural position. The reason regulators blocked the acquisition was precisely the reason the acquisition was valuable: Arm is too central to the semiconductor ecosystem to be owned by any single participant. Its neutrality is not a nice-to-have feature of the business model. It is the business model. The moment Arm favors one licensee over another — the moment its roadmap tilts toward one customer's needs at the expense of another's — the architecture's credibility as a universal standard collapses. This constraint is simultaneously Arm's greatest strength (no one can replicate the trust) and its greatest strategic limitation (Arm can never fully capture the value it creates).
The Re-IPO and the AI Inflection
With the Nvidia deal dead, SoftBank pivoted to Plan B: take Arm public again. In September 2023, Arm listed on the Nasdaq in the biggest IPO of the year, raising $4.87 billion at $51 per share. The cornerstone investor list read like a who's who of the semiconductor industry: Apple, Google, Nvidia, Samsung, Intel, AMD, TSMC Partners, Qualcomm (through a subsidiary), Cadence, Synopsys, and MediaTek collectively indicated interest in purchasing up to $735 million of shares. That these companies — many of them fierce competitors — all wanted equity in Arm told you everything about the architecture's centrality. SoftBank retained approximately 90.6% of the outstanding shares, making Arm a controlled company under Nasdaq rules.
The IPO arrived at the precise moment when AI was rewriting the semiconductor industry's economics. Arm's first quarterly shareholder letter as a re-listed public company, for Q2 FY2024 (ending September 30, 2023), reported revenue of $806 million — up 28% year-over-year and topping $800 million for the first time. Licensing revenue had surged 106% as AI-driven demand kicked off increased investment across all end markets. Non-GAAP operating margin was 47.3%.
Licensing revenue was up over 100% year-over-year as the demand for AI has kicked off increased investment across all end markets. Our royalty revenue benefited from market share gains in automotive and cloud compute as our latest technologies, such as Armv9, increased penetration across all markets where AI is driving the need for our unique combination of performance and power efficiency.
The key mechanism was Armv9, the latest generation of the architecture. Arm had structured Armv9 royalty rates at roughly double those of the prior Armv8 generation — a pricing lever the company had never before been able to pull so aggressively. The adoption was rapid: by Q3 FY2024, Armv9 accounted for 15% of total royalties and was accelerating. Nvidia's Grace server CPU used Armv9. Apple's iPhone 15 and subsequent models used Armv9. Samsung's Galaxy S24 used Armv9. The entire AI-driven hardware cycle — from cloud to edge, from data center to smartphone — was migrating to a more expensive version of Arm's architecture, and Arm was collecting a larger royalty on every chip.
Rene Haas and the [Pivot](/mental-models/pivot) to Premium
Rene Haas became Arm's CEO in February 2022, succeeding Simon Segars, who had led the company since 2013. Haas was a semiconductor industry lifer — he'd spent two decades at Nvidia before joining Arm in 2013 as president of the IP Products Group. He was, in the most literal sense, an Nvidia-trained operator placed at the helm of the company Nvidia had just failed to acquire. Where Segars was an Arm institutional insider, a Cambridge engineer who'd joined the company in 1991 as employee number sixteen, Haas brought a Silicon Valley commercial sensibility — an instinct for monetization, for premium positioning, for extracting more value from each unit of technology shipped.
70% of the world's population uses Arm, so I think by definition you can't really run AI without Arm.
Haas's strategic vision centered on two shifts. First, move Arm upstack — from licensing bare processor cores to licensing Compute Subsystems (CSS), essentially pre-integrated, pre-verified processor complexes that customers could drop into their chip designs with less customization. CSS reduced customers' engineering time and accelerated time-to-market, but it also meant Arm was delivering more value per license and could charge more for it. The Neoverse CSS for cloud compute, announced in late 2023, was the flagship product of this strategy. Second, aggressively penetrate data center and automotive markets where Arm had historically been a marginal player. In cloud, AWS's Graviton processors had proven that Arm-based server chips could deliver superior price-performance to x86 incumbents. Microsoft's Cobalt, Google's Axion, and Nvidia's Grace followed. In automotive, the shift toward software-defined vehicles and advanced driver-assistance systems (ADAS) was driving demand for high-performance, energy-efficient compute — Arm's native territory.
The financial results validated the strategy. For fiscal year 2025 (ending March 2025), Arm reported total revenue of approximately $3.93 billion, up significantly from $2.68 billion in FY2023 and $3.23 billion in FY2024. Royalty revenue was growing in the double digits, driven by the Armv9 mix shift and market share gains in cloud and automotive. Licensing revenue was growing even faster, fueled by AI-driven demand for new chip designs across every end market. Gross margins remained stratospheric — consistently above 95% — reflecting the near-zero marginal cost of licensing intellectual property.
The Qualcomm War and the Price of Neutrality
Arm's relationship with Qualcomm, long one of its most important licensees, deteriorated into open litigation in a dispute that illuminated the tensions inherent in the licensing model. The conflict centered on Qualcomm's 2021 acquisition of Nuvia, a chip design startup founded by former Apple engineers. Nuvia had its own Arm architecture license, which it used to develop custom CPU cores. When Qualcomm acquired Nuvia and began integrating Nuvia's custom cores into its Snapdragon processors — most notably the Snapdragon X Elite for AI-capable laptops — Arm argued that Nuvia's license did not transfer to Qualcomm and that Qualcomm needed to negotiate a new, presumably more expensive, license agreement.
Arm eventually cancelled Qualcomm's chip design license, an extraordinarily aggressive move given that Qualcomm was one of the largest Arm licensees in the world. The lawsuit went to trial in late 2024, and while the details of the resolution remain complex, the case sent a signal through the industry: Arm under Haas was willing to enforce its intellectual property rights more aggressively, even at the cost of antagonizing major customers. The traditional Arm — the neutral, accommodating, ecosystem-first licensor — was being replaced by a company that intended to capture more of the value its technology created.
This shift carried real risk. Arm's leverage depends on the absence of viable alternatives. As long as the switching costs of moving off the Arm architecture are prohibitively high — as long as the 22 million developers, the billions of lines of Arm-optimized software, the entire mobile ecosystem create a gravitational field that no competitor can escape — Arm can raise royalty rates and tighten licensing terms. But if Arm pushes too hard, the incentive for large customers to invest in RISC-V, the open-source instruction set architecture, intensifies. RISC-V is still years away from matching Arm's ecosystem maturity in most markets, but companies like Qualcomm, Google, and several Chinese semiconductor firms have invested meaningfully in RISC-V development. The line between "pricing to value" and "pricing yourself into disruption" is drawn in invisible ink.
The Chip That Arm Might Make
In late 2024, reports surfaced that Arm was exploring designing its own chip — not just a licensable IP core, but a complete semiconductor product manufactured by a foundry partner like TSMC. If true, this would represent the most fundamental strategic pivot in the company's history, a move from being the Switzerland of semiconductors to becoming a competitor to its own customers.
The logic, from Arm's perspective, is not insane. The company's Compute Subsystems already represent a significant move up the value chain, providing more complete, production-ready designs. A finished chip is the logical endpoint of that trajectory. And the AI data center market, where hyperscalers are spending tens of billions of dollars annually on custom silicon, represents an enormous revenue opportunity that Arm currently captures only through modest per-chip royalties. Designing its own chip could unlock revenue measured in the billions rather than the hundreds of millions.
But the risks are existential. Arm's ecosystem — the 500-plus licensees, the $30+ billion semiconductor companies that depend on the architecture — exists precisely because Arm is not a competitor. The moment Arm ships its own chip for, say, the cloud server market, every other Arm licensee designing cloud chips (Nvidia, Qualcomm, Ampere, AWS, Google, Microsoft) faces the question of whether they are developing on a platform controlled by a rival. The company that was too important for Nvidia to own might become too conflicted for anyone to trust.
Physical AI and the Next Trillion
Haas has increasingly positioned Arm as an AI company — not because Arm designs the GPUs or accelerators that train large language models, but because Arm's architecture underpins nearly every device where AI inference happens. The smartphone in your pocket, the autonomous vehicle navigating a city street, the robot on a factory floor, the smart speaker interpreting your voice — all of them run AI workloads on Arm-based processors.
I think in the next five years, you're going to see large sections of factory work replaced by robots—and part of the reason for that is that these physical AI robots can be reprogrammed into different tasks.
The data center opportunity is real but contested. AWS's Graviton processors, which are Arm-based, now handle a significant and growing share of Amazon's cloud computing workloads, and AWS has publicly stated that Graviton instances offer up to 40% better price-performance than comparable x86 instances. Microsoft's Cobalt and Google's Axion followed the same playbook. Nvidia's Grace CPU, designed for AI-adjacent workloads in data centers, is Arm-based. The x86 duopoly that Intel and AMD maintained for decades in the server market is fracturing, and Arm is the beneficiary.
But the secular opportunity that Haas talks about most — the one that could justify Arm's premium valuation over a decade-plus time horizon — is what he calls "physical AI": the convergence of artificial intelligence with the physical world through autonomous vehicles, humanoid robots, industrial automation, and smart infrastructure. Each autonomous vehicle contains dozens of high-performance processors. Each humanoid robot requires real-time AI inference at the edge. Each smart factory deploys thousands of connected controllers. If this market materializes at scale, Arm's royalty base could expand by orders of magnitude — not because each chip generates more royalties, but because the number of high-value chips per device multiplies dramatically.
Haas has also been candid about the semiconductor supply chain's fragility. "The semiconductor supply chain has many single points of failure," he told a Fortune audience in late 2025. "There's TSMC, which is in a very obviously interesting part of the world geopolitically. There is also a very sophisticated device that has to go into these fabs that comes from one company on the planet… called ASML." Arm's own centrality to this chain — its architecture present in virtually every connected device — makes it both a beneficiary and a vulnerability in an era of semiconductor nationalism.
The Instruction Set as Empire
What Arm has built, over three decades, is something that resists easy categorization. It is not a technology company in the way most people understand the term — it makes no products that consumers purchase. It is not a semiconductor company — it operates no fabs. It is not a software company — though its developer ecosystem of 22 million is larger than most. It is, perhaps most accurately, a standard — a set of rules for how silicon should compute, embedded so deeply in the global technology stack that extracting it would require rebuilding the entire edifice of modern electronics.
The financial expression of this position is unusual. Gross margins above 95%. Operating margins approaching 50% on a non-GAAP basis. Revenue that grows not through sales effort but through the autonomous expansion of devices in the world that speak its language. The business has the economic profile of a software company — high margins, near-zero marginal cost, scalable revenue — without the churn or competitive dynamics of software. No one cancels their Arm license the way they cancel a SaaS subscription. The architecture is load-bearing.
As detailed in Daniel Nenni and Don Dingee's
Mobile Unleashed: The Origin and Evolution of ARM Processors in Our Devices, the company's history is a case study in the compounding power of patience — of choosing ubiquity over margin at every juncture, of building an ecosystem so vast that it becomes self-reinforcing, of designing an architecture so efficient that when the world's computational demands finally caught up with its constraints, Arm was already everywhere.
For fiscal year 2025, Arm reported that 7.1 billion Arm-based chips were shipped in a single quarter — Q2 FY2024 alone — taking the cumulative total past 272 billion. By late 2025, the company claimed more than 325 billion total. The numbers have a quality that defeats human intuition. There are more Arm processors on earth than there are insects.
The barn in Swaffham Bulbeck is long gone. The twelve engineers became seven thousand. The architecture designed for a battery-powered handheld became the substrate on which the entire AI revolution now runs. In Cambridge, in a purpose-built headquarters that would have seemed unfathomable to Sophie Wilson sketching instruction sets in 1985, Arm's engineers are designing version 10 of the architecture — the one that will power the autonomous vehicles and humanoid robots and AI data centers of the next decade. The royalty rate, as always, will be measured in pennies per chip. The chips, as always, will be measured in billions.
