Arm Holdings

Arm is moving from chip blueprint to the platform—and now the silicon—of AI computing

Arm designs the instruction-set architecture, processor cores, system intellectual property and software foundations used by semiconductor companies to build chips. Its historic model is unusually attractive: customers fund manufacturing and inventory, while Arm receives licence fees during design and royalties when chips ship. The architecture became dominant in mobile and embedded devices because it combined power efficiency with a broad ecosystem. AI creates a route from that base into cloud, edge, vehicles and physical systems.

The strategy is also changing. Compute Subsystems package more complete, validated designs and capture more value than standalone cores. In 2026 Arm introduced the AGI CPU, its first production silicon product, for agentic AI data centres. This can accelerate adoption and expand revenue per system, but it crosses a boundary: the neutral supplier to chipmakers is becoming a potential competitor and assuming product, supply and customer-concentration risk. The investment debate is whether greater integration compounds the platform or weakens the ecosystem that made it ubiquitous.

The business in one map

The instruction set is valuable because software already expects it

An instruction set defines how software communicates with a processor. Once operating systems, compilers, tools and applications support an architecture, every new compatible chip inherits a large software base. Developers can target the platform with less work, customers can change chip suppliers without rewriting everything and semiconductor companies can focus engineering on differentiation.

Arm’s architecture spans tiny controllers, phones, laptops, base stations, vehicles and servers. That breadth produces a feedback loop: more shipped devices attract software investment; more software makes the architecture safer for new markets; new licensees extend volume and use cases. Compatibility must be preserved while performance, security and features evolve.

The architecture is not perfectly uniform. Custom cores, system designs and operating environments vary, and optimisation for one device may not carry everywhere. Arm invests in reference platforms, developer tools and open-source support to reduce fragmentation. The ecosystem moat is strongest when software portability remains high without forcing customers into an identical chip.

Licensing creates the design pipeline; royalties monetise success

Licence revenue is recognised as customers obtain technology and support through multi-year agreements. It can be large and lumpy because contract scope, timing and accounting differ. The more important strategic function is to seed future chips. A signed licence only creates royalty value if the customer completes the design, wins sockets and ships at volume.

Royalty revenue is generated when licensees report shipments. It reflects unit volumes, chip prices, technology generation and contractual rates. Newer architecture and more complete IP can command greater economics, while the massive base of older low-cost chips produces smaller royalty per unit. Mix can therefore matter more than total devices.

The delay from licence to royalty may span years. This gives visibility when design pipelines are broad but makes near-term inference difficult. Customers may delay, cancel or lose market share. Underwriting should connect licence cohorts to disclosed product launches and royalty generations rather than treating current licence growth as immediate recurring revenue.

Compute Subsystems move Arm from components toward a repeatable platform

Designing a modern system-on-chip requires CPU clusters, interconnect, memory, security, power management, verification and physical implementation. As complexity rises, customers may not create value by rebuilding common infrastructure. Compute Subsystems provide a more complete and validated starting point that can be configured for cloud, client, automotive and edge use.

CSS can shorten time to market, reduce engineering risk and deliver predictable performance on a selected manufacturing process. Arm supplies more intellectual property and performs more integration, justifying higher licence and royalty economics. Customers retain room to add accelerators, interfaces and product-specific features.

The tension is differentiation. Leading chip companies may prefer custom cores and system design because silicon is strategic. Smaller companies may welcome a nearly complete platform but have less volume. Arm must offer a ladder from core to subsystem without making its best customers feel that architecture access is being restricted to favour higher-priced products.

The AGI CPU is a strategic boundary shift

Arm’s AGI CPU moves beyond intellectual property and subsystems into an Arm-designed production data-centre chip. It targets CPU work around AI accelerators: retrieval, orchestration, databases, networking, model serving and persistent agent workflows. Meta is the lead development partner, while server, cloud and software companies are supporting the platform.

The move can accelerate time to deployment for customers that want Arm efficiency without designing a custom processor. It gives software vendors a common reference target and lets Arm optimise memory, I/O and rack performance. Silicon revenue can be far larger per unit than a royalty, and direct product feedback may improve the underlying platform.

It also changes risk. Arm must manage foundry capacity, packaging, inventory, quality, sales and product road maps. Licensees that sell server processors may question whether their supplier now has preferential insight or competes for customers. A successful strategy expands total Arm adoption while preserving equal architecture access. If it reallocates existing demand from partners to Arm, ecosystem economics worsen.

Agentic AI increases CPU work around the accelerator

Training large models concentrated attention on accelerators. Production agents create a broader system workload. They retrieve data, call tools, maintain memory, coordinate services, enforce policy and run business logic between model invocations. These tasks require general-purpose CPU compute, memory bandwidth and input-output capacity even when the model itself runs on an accelerator.

Data-centre power is a hard constraint. A CPU that performs orchestration efficiently leaves more of the rack’s power budget for valuable accelerator work and permits greater compute density. Arm’s power-efficient heritage and customisation model fit heterogeneous systems where CPUs, accelerators, networking and memory are designed together.

The opportunity should not be reduced to a claimed performance multiple. Workload mix, compiler maturity, memory, software and total system cost determine customer choice. x86 has deep enterprise compatibility, while custom Arm CPUs are already optimised by hyperscalers. AGI must show broad production performance and reliable supply beyond a lead customer.

Neoverse allows cloud customers to customise the CPU economics

Large cloud operators design Arm-based processors to control cost, power and workload optimisation rather than relying only on merchant x86 CPUs. Neoverse provides server-class cores and system technology, while an architecture licence can support deeper customisation. The cloud controls the software environment and can make migration easier through managed services.

Arm benefits from each successful custom chip without funding the data centre or wafer inventory. The hyperscaler gains negotiating leverage and a processor tuned to its fleet. Growth depends on workloads moving to Arm instances and on successive designs using newer, higher-value technology. Announced chips are not the same as utilised instances.

This model can pressure merchant CPU vendors and expand Arm’s share of cloud compute. It also gives the largest customers considerable leverage because they own system design, software distribution and purchasing scale. CSS and AGI increase Arm’s value capture but must coexist with customers whose strategic reason for using Arm is control.

Mobile is the royalty base and the discipline

Smartphones established Arm’s scale. Application processors use Arm CPUs and often Arm graphics or system IP, while an enormous developer base targets the architecture. Premium devices adopt newer cores and Armv9 features, lifting royalty value even when total handset units are mature. More compute for on-device AI can expand silicon area and performance requirements.

Arm’s challenge is to capture more value without disrupting customer economics. Major mobile chip designers differentiate through custom cores, modems, accelerators and system integration. They can use architecture licences or Arm cores depending on strategy. CSS appeals when time and engineering efficiency matter, but the most capable customers resist commoditisation.

Unit maturity does not mean irrelevance. Mobile funds the software ecosystem and provides a proving ground for performance per watt, security and AI libraries. The risk is customer concentration and bargaining power, especially where a few chip suppliers account for premium volume. Royalty growth should be separated into units, price, architecture generation and additional IP.

Edge and physical AI turn power efficiency into reach

AI inference is spreading into cameras, factories, appliances, network devices and sensors. Local processing reduces latency, bandwidth cost and dependence on connectivity while keeping some data on the device. These systems operate under strict power, thermal and cost budgets—the environment where Arm historically excels.

Arm can combine CPU, Ethos neural processing, Mali or Immortalis graphics, security and interconnect. Kleidi software libraries help applications use CPU and accelerator features without each developer hand-optimising the hardware. A common architecture from cloud to edge can simplify deployment and model portability.

Edge volumes can be enormous but royalty per chip may be small, and alternatives include proprietary microcontrollers and open instruction sets. AI must produce real device value, not simply add a feature label. The attractive mix is higher compute content, newer Arm technology and long product lives in industrial or infrastructure applications.

Automotive is a long-duration design-win market

Vehicles are consolidating dozens of electronic control units into central and zonal computers. Driver assistance, cockpit, connectivity and autonomy require more compute, while functional safety and lifecycle support are demanding. Arm’s existing embedded footprint and efficient processors provide a natural path into higher-value vehicle compute.

Automotive CSS can package validated compute and safety foundations, reducing chip-development time. Royalty duration can be attractive because a design remains in production for many years. However, qualification and launch cycles are long, volumes ramp slowly and vehicle delays push revenue far beyond the initial licence.

Competition comes from x86 in selected cockpit workloads, RISC-V, custom accelerators and incumbent automotive semiconductor platforms. Arm should be measured by production vehicles, content per system and adoption across central compute—not by the number of conceptual partnerships. Safety certification and software continuity deepen switching cost once deployed.

Software is the invisible investment behind royalty economics

Arm supplies no value if software cannot use the chip. The company contributes to operating systems, compilers, cloud tooling, security standards, libraries and developer frameworks. This investment is expensed today while its return appears later through customer adoption and royalties. It is therefore part of the economic moat, not simply support.

AI makes software enablement more important because models and kernels must be optimised across diverse CPUs, GPUs and neural processors. Kleidi provides common performance libraries, while broad frameworks and operating systems reduce porting cost. The aim is for developers to target Arm once while hardware partners differentiate underneath.

The risk is fragmentation or dependence on other ecosystems. CUDA dominates accelerator development, and cloud software often abstracts the CPU. RISC-V communities can reproduce open tooling over time. Arm must keep libraries open enough to attract developers while using performance and compatibility to pull demand for licensed IP.

RISC-V is a strategic price umbrella and a real long-term competitor

RISC-V offers an open instruction set that allows companies to build processors without paying for architecture access. It is attractive for microcontrollers, accelerators and regions seeking technology independence. An open standard can gather innovation from universities, start-ups and semiconductor companies and place pressure on Arm licence economics.

An instruction set alone is not a complete platform. Customers need high-performance cores, verification, security, tools, operating systems, libraries and support. Arm’s decades of investment and shipped volume reduce execution risk. The gap is widest in complex mobile and data-centre systems and narrower in simple embedded applications.

Arm should respond through faster innovation and superior total development economics rather than relying on contractual control. CSS makes the make-versus-buy decision more favourable when customers face rising design cost. If RISC-V reaches adequate performance and software maturity in high-volume markets, it can limit price and reduce Arm’s share even without being technically superior.

The SoftBank relationship adds capital and governance complexity

SoftBank remains Arm’s controlling shareholder. A long-term owner can support ambitious research and strategic moves that public markets might resist. It also creates a limited public float and potential related-party questions when SoftBank-backed companies participate in the AI ecosystem or when the owner sells shares.

Minority investors need confidence that intellectual property, commercial terms and capital allocation are set for Arm rather than the wider portfolio. Independent governance, transparent transactions and equal customer access are especially important as Arm enters silicon. Perceived favouritism could damage licensee trust even without immediate financial harm.

Arm China is another structural complexity. The Chinese business provides access to an important semiconductor market through a distinct local arrangement. Collection, reporting, technology restrictions and strategic alignment introduce risk. China can contribute licensing and royalty growth, but investors should apply a higher uncertainty to control and cash conversion.

AI reaches Arm through five reinforcing layers

Competitive landscape

A scale checkpoint, not a valuation shortcut

The asymmetry between platform reach and current revenue explains both the opportunity and the valuation risk. Arm can capture more per device through newer architecture, CSS and silicon, but customers will resist a tax unrelated to delivered value. Growth must come from better technology, higher compute content and new markets rather than simply raising rates on a captive base.

The investment debate

What would disconfirm the thesis

How to underwrite Arm

Model licences and royalties separately. Build royalty by end market, units, chip value, architecture generation and IP content. Apply a realistic delay from licence signature to production and allow for failed designs. Analyse newer technology as a mix shift rather than assuming all installed devices move immediately.

Separate capital-light IP economics from AGI silicon. Direct chips require cost of goods, supply commitments, inventory and product support that royalties do not. Evaluate gross profit, working capital and concentration after foundry and packaging cost. Do not award the IP margin to silicon revenue before the product proves it.

Finally, monitor ecosystem health. Arm’s value comes from enabling many customers to innovate on one compatible base. Higher value capture is sustainable when it reduces their design cost or improves their products. If customers perceive that Arm is narrowing access, competing unfairly or pricing beyond delivered value, they have a reason to invest in custom or open alternatives.

Bottom line