Arm AGI CPU

The Announcement

On March 24, 2026, Arm Holdings unveiled the Arm AGI CPU at its "Arm Everywhere" keynote in San Francisco. It is the first production silicon chip in the company's 35-year history.¹

Arm has spent over three decades as the most influential chip company that does not make chips. Its architecture powers virtually every smartphone on earth, most of the world's embedded devices, and an expanding share of cloud servers. More than 280 billion Arm-based chips have shipped since 1990.² Arm itself has never sold a finished processor until the AGI CPU.

The AGI CPU is a 136-core server processor built on TSMC's 3nm process, designed for AI data center workloads. Meta is the lead customer and co-development partner. Eight other launch partners, including OpenAI, Cloudflare, Cerebras, SAP, SK Telecom, F5, Positron, and Rebellions, have committed to deploying the chip. Commercial systems are available from ASRockRack, Lenovo, and Supermicro.^1,3

Arm claims the chip delivers more than 2x the performance per rack compared to incumbent x86 servers, with potential capex savings of up to $10 billion per gigawatt of AI data center capacity.³ By entering the silicon business directly, Arm is shifting from a pure IP licensing company toward a hybrid model that captures more value from the AI infrastructure buildout.

A Brief History of Arm

Arm traces back to Cambridge, England, where in 1985 two Acorn Computers engineers, Sophie Wilson and Steve Furber, designed the ARM1: the world's first commercial RISC processor. It worked on its first silicon run and consumed so little power that it kept running after the team accidentally disconnected the power supply, surviving on residual current through the I/O pins.⁵

In 1990, Arm was spun out as a joint venture between Acorn, Apple, and VLSI Technology. Apple invested $3 million; it needed a low-power chip for the Newton PDA. From day one, Arm chose licensing over manufacturing: design the architecture, let others build the chips.^5,6

The bet worked because mobile happened. Nokia, Qualcomm, and Texas Instruments all needed power-efficient processors for phones, and Arm's RISC design was the natural fit. By the iPhone launch in 2007, Arm was ubiquitous. Today, Arm-based chips power roughly 99% of smartphones and more than 280 billion chips have shipped in total.^2,6

SoftBank acquired Arm for $32 billion in 2016 and still owns roughly 90% of shares. After NVIDIA's $40 billion acquisition attempt was blocked by regulators in 2022, SoftBank re-IPO'd Arm in September 2023 at a $54.5 billion valuation.⁷

Throughout all of this, the business model never changed: design the blueprints, license them out, collect royalties. The AGI CPU is the first time in 35 years that Arm has decided to build the house itself.

How Arm Makes Money

Arm's business model is a bit unusual in semiconductors. The company does not manufacture or sell chips. It designs processor architectures and instruction sets, then licenses that intellectual property to chipmakers who build their own processors using Arm's designs. There are two core revenue streams.⁸

Licensing Revenue

Companies pay upfront fees to access Arm's IP. There are several licensing tiers. At the simplest level, a customer licenses a pre-designed Arm processor core (like the Cortex-A78 or Neoverse V3) and integrates it into their system-on-chip. At the highest level, architecture licensees like Apple and Qualcomm license the instruction set itself and design entirely custom cores that implement Arm's architecture but share none of its microarchitecture.⁸

In Q3 FY2026 (ended December 2025), license and other revenue was $505 million, up 25% year-over-year.⁹

Royalty Revenue

For every chip shipped using Arm's IP, the chip manufacturer pays a per-unit royalty. The royalty rate varies by chip type, complexity, and the licensing agreement, but it typically ranges from low single-digit percentages of the chip's selling price to fixed per-unit fees. More advanced architectures command higher royalties. The transition from Armv8 to Armv9 is particularly important: Armv9 chips carry roughly double the royalty rate of Armv8, and Armv9 represented approximately 25% of royalty revenue as of early fiscal 2025 and is growing rapidly as adoption accelerates.^8,9

Beyond Armv9, Arm's Compute Subsystems (CSS) program has further transformed the economics. CSS provides chipmakers with more complete, pre-integrated designs rather than bare IP blocks. The effect is dramatic: CSS licenses command royalty rates north of 10% per chip, compared to the 1-3% typical of traditional licensing. Twenty-one CSS licenses have been signed, with five already shipping.²²

In Q3 FY2026, royalty revenue hit a record $737 million, up 27% year-over-year. Data center royalty revenue grew more than 100% year-over-year.⁹

The Revenue Gap

Here is the tension at the core of Arm's valuation story. Arm's technology is in 99% of smartphones and a growing share of data center servers. It has roughly 260 customers including every major technology company on earth. Yet its annual revenue run rate is approximately $5 billion. Apple alone generates over $380 billion in annual revenue from products built on Arm's architecture. NVIDIA's data center GPUs, many of which use Arm-based host processors, generated $115 billion in annual revenue in the most recent fiscal year.²

Arm captures pennies on the dollar for every device its technology enables. The licensing model is highly profitable (non-GAAP gross margins around 98%, operating margins around 40%) but fundamentally limited in how much value it can extract from the AI infrastructure buildout.

The AGI CPU changes this equation. Instead of collecting a royalty of a few percent on a $10,000 server chip, Arm would sell the entire chip. Management has framed the opportunity as expanding from a roughly $3 billion royalty TAM toward a potential $100 billion market by selling finished silicon.¹⁰

The Dual Revenue Model

What makes Arm's position structurally unique: the AGI CPU creates a dual-monetization model that no other chip company can replicate. Arm continues to collect royalties on every chip its licensees sell (AWS Graviton, Google Axion, Microsoft Cobalt, NVIDIA Grace) while simultaneously selling its own competing silicon. It gets paid whether it wins the socket or its customers do.^18,21

Arm CFO Jason Child told CNBC the company is selling the AGI CPU at roughly 50% gross profit. That is significantly lower than the ~98% gross margins on pure IP licensing. But the dollars per unit are in a different league. On a $10,000 server chip where Arm might collect a few hundred dollars in royalties, selling the chip directly at 50% gross margin could yield several thousand dollars in gross profit per unit.¹⁹

"It expands our market to include customers that were not interested in an IP model, gives our current customers choice, and for Arm it creates a much larger profit opportunity," Child said. "We're not going to force any of our existing customers to migrate to this new model."¹⁹

There is a real market need here. Until now, getting an Arm server chip meant designing one yourself. That requires roughly 1,000 engineers and a $500 million budget, according to chip analyst Patrick Moorhead. AWS and Google can do that. Most data center operators cannot. The AGI CPU is an off-the-shelf Arm server processor they can just buy.¹⁸

Inside the AGI CPU

The AGI CPU is built around 136 Arm Neoverse V3 cores on TSMC's 3nm process node. It uses a dual-chiplet design, meaning two silicon dies are packaged together in a single module, each containing half the cores. This is a common approach in modern server processors (AMD's EPYC uses a similar multi-die architecture) that allows higher core counts while keeping individual die sizes manageable for manufacturing yield.^1,4

Source: Arm product page, The Register^1,4

Design Philosophy

The AGI CPU was purpose-built for AI data center workloads. No simultaneous multithreading, no turbo boost, no legacy application accelerators. Mohamed Awad told The Register the team "specifically didn't want to add things that weren't going to be 100 percent utilized in the mission of this device." Where Intel and AMD maintain backward compatibility with decades of x86 software, the AGI CPU dedicates its transistor budget entirely to AI-relevant performance.^1,4

The tradeoff: it will not run legacy enterprise workloads. It is built for environments where the software stack already runs on Arm, which describes most modern cloud-native and AI infrastructure.

Unlike AMD and Intel, which separate I/O and memory controllers onto dedicated chiplets, the AGI CPU integrates them onto the same dies as the compute cores. At 300 watts for 136 cores, it runs at 2.2 watts per core, compared to 3.9 watts for Intel's 128-core Xeon 6 P-cores (500W) and 2.3 watts for Intel's efficiency E-cores at less than half the clock speed. That efficiency gap is the foundation of Arm's 2x performance-per-rack claim.^4,23

Rack Density

At a standard 36kW air-cooled power envelope, a single rack holds up to 8,160 cores. In a liquid-cooled 200kW configuration, Arm has validated 45,696 cores per rack, more than twice the density of NVIDIA's Vera racks. Beyond the flagship 136-core part, 128-core and 64-core variants are planned. Production begins H2 2026, with initial customer shipments in Q4.^3,4,21

Meta and the Agentic Thesis

The Co-Development Partnership

Meta is not just the AGI CPU's first customer, it is also the co-developer. The partnership started three years ago when Meta asked Arm to build a finished chip, not just license IP. Paul Saab, a Meta software engineer who has been on the project since 2023, told CNBC: "In today's world, you really only have a couple of players. This adds yet another player to the ecosystem for us."¹⁸

Meta's Head of Infrastructure, Santosh Janardhan, framed the chip as part of a "robust and adaptable portfolio of custom silicon solutions, purpose-built to accelerate AI workloads." The AGI CPU is designed to work alongside Meta's own MTIA accelerators (a family of RISC-V inference chips developed with Broadcom, now in its fourth generation) as well as GPUs from NVIDIA and AMD. Meta is spending $115-135 billion on capital expenditure this year. Even a low-single-digit share going to Arm would be transformative for Arm's revenue.^20,18

Critically, Meta insisted from the start that this not be an exclusive chip. "The first conversations we had with Arm were, 'Hey, if we build this, we don't want to keep this only within the company,'" Saab told CNBC. Meta will release its board and rack designs for the AGI CPU through the Open Compute Project. The chip is available to anyone.^18,20

This matters because it explains who the AGI CPU is really for. Not hyperscalers with their own chip design teams, but the next tier: companies that need Arm server performance but cannot afford a $500 million development budget and a 1,000-person engineering team.¹⁸

Why CPUs Are Coming Back

Meta validates that there is real customer demand for Arm server silicon. The question is whether the underlying market is large enough. Arm's bet: agentic AI is restructuring how data centers consume compute. The first phase of the AI buildout was GPU-dominated. The next phase looks different. Agentic systems run continuously, breaking tasks into sub-steps, calling tools, querying databases, and coordinating across models. That is sustained, multi-step compute that runs primarily on CPUs.¹¹

The supporting evidence is increasingly credible. A November 2025 arXiv paper found that CPU-side tool processing accounts for up to 90.6% of total latency in agentic workloads.¹² Futurum Group projects the data center CPU market will reach $76.6 billion by 2029, with growth accelerating to 34.9%, outpacing GPUs.²¹ At GTC, NVIDIA CEO Jensen Huang told CNBC that CPUs are "becoming the bottleneck," and unveiled an entire rack of standalone Vera CPUs.¹⁸

Arm estimates agentic workloads require a fourfold increase in CPU cores compared to traditional cloud deployments, roughly 120 million cores per gigawatt of data center capacity. Bank of America estimates the server CPU market could more than double from approximately $27 billion in 2025 to $60 billion by 2030.^11,13

That said, the thesis still depends on a few things that are early: that agentic workloads scale as projected, that Arm's power efficiency advantage persists at data center scale, and that the software ecosystem evolves quickly enough to support widespread Arm adoption. Each is plausible but not yet confirmed at the scale Arm is targeting.

Competitive Landscape

The AGI CPU enters a server CPU market with several entrenched incumbents and a growing set of Arm-based alternatives. Here is where each stands.

Intel Xeon

Intel remains the largest server CPU vendor by installed base, though share has eroded. The current Xeon 6 offers up to 288 E-cores; next-generation Clearwater Forest on Intel's 18A process arrives later in 2026. Intel's strength is the installed base of x86 software. Its weakness is power efficiency, a metric that increasingly determines data center economics as power becomes the binding constraint.^14,23

AMD EPYC

AMD has been the biggest x86 share gainer. Current EPYC "Turin" offers up to 192 cores; next-generation "Venice" (256 Zen 6 cores, TSMC 2nm, late 2026) claims a 70% generational leap. AMD's strongest argument against Arm is software compatibility: moving enterprise workloads from x86 to Arm requires recompilation and sometimes significant refactoring. Meta, notably, is a major AMD customer alongside its Arm partnership.^14,21

AWS Graviton

Amazon Web Services has been the most aggressive Arm adopter in the data center. Its Graviton processors (now in their fourth generation at 96 Neoverse V2 cores, with a 192-core Graviton5 in development) serve nearly 100,000 cloud customers and drive over half of AWS's CPU compute demand. Graviton proved that Arm can compete in servers at scale, but it is exclusive to AWS. The AGI CPU opens a similar proposition to every cloud provider and enterprise that cannot afford a custom chip team.^21,22

NVIDIA Grace and Vera

NVIDIA's Grace CPU, based on Arm's Neoverse cores, was designed as a companion processor for GPUs in the Grace Hopper and Grace Blackwell superchips. But at GTC last week, NVIDIA unveiled Vera, a new standalone CPU with 88 custom "Olympus" cores targeting exactly the agentic AI orchestration market that Arm is after. Vera's NVLink Fusion integration gives it a differentiated position for workloads that need tight GPU-CPU coupling. NVIDIA is now both Arm's biggest GPU customer and a direct CPU competitor.^4,21

Ampere (SoftBank)

Ampere Computing, founded by former Intel president Renee James in 2018, was the first significant Arm-based server chip startup. Its AmpereOne processor offered up to 192 cores, and its Aurora chip (in development) targets up to 512 cores with AI acceleration. In March 2025, SoftBank acquired Ampere for $6.5 billion, giving Arm's parent company a second chip design team and production experience in server silicon.¹⁵

The relationship between Arm and Ampere is still being defined. SoftBank owns both, and the Ampere acquisition likely informed the AGI CPU program. Ampere's no-SMT, no-turbo design philosophy is visible throughout the AGI CPU's architecture. Next Platform reports that Ampere went "kind of quiet" after the acquisition, and "strongly suspect[s] that Ampere Computing was acquired to give Arm a second chip design team so it can drive its homegrown CPU roadmap." The two companies currently operate independently, with the AGI CPU using Arm's own Neoverse V3 cores rather than Ampere's custom designs.^15,23

Sources: Company product pages, The Register, Futurum Group^3,4,14,21

Futurum Group notes that the server CPU market "has not been this broadly contested in two decades." Six viable architectures from major vendors are all competing for AI data center sockets, with significant launches from every player scheduled for 2026-2027.²¹

The Financial Picture

Arm has delivered four consecutive billion-dollar revenue quarters, a milestone that underscores the growing monetization of its IP as AI proliferates across device categories.

Source: Arm quarterly results, investor relations^9,16

Data Center: The Growth Engine

The most striking number in Arm's recent results is data center royalty growth: more than 100% year-over-year in Q3 FY2026. Management has stated that data center is expected to become Arm's largest revenue segment within a few years, surpassing mobile. This trajectory reflects both the organic adoption of Arm-based server chips by cloud providers and the early stages of the broader market transition toward Arm in the data center.⁹

Valuation and the Revenue Bet

Prior to the March 24 AGI CPU announcement, Arm traded around $135 per share. The stock popped 15% on the news, pushing the market cap to roughly $164 billion. CEO Haas projected $25 billion in annual revenue by 2031, with $15 billion from AGI CPU sales and $9 in earnings per share. "We may be under-calling that number," Haas said. "I think the demand is higher than we think it is."^19,23

To put that in context: Arm generated just over $4 billion in revenue in 2025. Haas is projecting more than 6x growth in five years, with the majority coming from a product line that does not yet exist in production. At approximately 180x trailing earnings, the stock prices in substantial execution on both fronts.

The R&D investment required to become a chip company is already visible. R&D spending surged 46% to $512 million in Q3 FY2026, compressing operating margins by 430 basis points to 40.7%. This is not a temporary spike. It reflects the ongoing cost of chip design teams, the $71 million Austin lab operation, and the multi-generation AGI CPU roadmap.²²

Analyst Ben Bajarin of Creative Strategies told CNBC: "Arm has typically been modeled purely on their licensing and royalty business and now they have given investors a new market opportunity and business to wrap their head around and model, so it isn't a surprise it will take some time for folks to wrap their head around the valuation and new revenue targets."¹⁹

The AGI CPU Revenue Opportunity

Futurum Group projects the data center CPU market will reach $76.6 billion by 2029, with growth accelerating to 34.9% annually, outpacing even GPU growth. Arm's management frames the opportunity as expanding from a roughly $3 billion royalty TAM toward a $100 billion addressable market by selling finished silicon.^10,21

The AGI CPU has not generated meaningful revenue yet. Production begins in H2 2026, with initial customer shipments in Q4 and volume ramp in 2027. The roadmap shows an annual cadence: a second-generation chip (likely on TSMC's 2nm High NA process with significantly higher core counts) is already in development, with Ampere Computing's engineering team, acquired by SoftBank for $6.5 billion in March 2025, widely believed to be contributing to future generations.^15,23

Whether the $15 billion target is achievable depends on wafer supply (TSMC 3nm allocation is fiercely contested), customer adoption beyond Meta, and the competitive response from AMD, Intel, and NVIDIA. It also depends on whether Arm can sell silicon at scale while maintaining the licensing relationships that fund the existing business.

What to Watch

The bull case for the AGI CPU is strong, but there are a few dynamics worth tracking as this plays out.

Arm has spent 35 years building the most widely licensed processor architecture in the world, and the AGI CPU is its bet that the next phase of AI infrastructure favors the company that designed the architecture, not just the companies that licensed it. The core question going forward is whether Arm can scale a silicon business without undermining the licensing relationships that still fund everything else, and whether the agentic AI workload shift materializes fast enough to justify a valuation that already assumes it will. We will be watching the production ramp in H2 2026, Meta's deployment timeline, and any signals from AWS, Google, or Microsoft on how they view Arm's new role as both supplier and competitor.