ARM Holdings PLC (ARM) Report Interpretation

Overview

The report summarizes Bernstein's meeting with Arm management at its London SDC. It argues that Arm's first in-house AGI CPU remains on schedule, with customer demand, especially from Meta and Agentic AI workloads, exceeding supply; execution and advanced-node foundry capacity are the principal constraints.

Core views

Arm's AGI CPU program remains in final qualification, with approximately 4,000 CPUs in the final testing round with Meta. Assuming no material qualification issues, management expects first shipments in CQ1 2027, or FQ4, generating about US$100m of revenue. Meta supported the R&D and qualification process and receives preferential pricing; it is expected to contribute roughly 90% of the planned US$1bn FY28 revenue opportunity. Its planned use cases include MTIA head-node functions, general-purpose computing and Agentic AI workloads, across products ranging from up to 136-core versions to higher-clock-speed head-node SKUs. The near-term opportunity is constrained by supply, not demand. Management said demand already exceeds US$2bn, while Arm initially secured around US$1bn of TSMC supply. Arm has since obtained additional wafer commitments and now believes revenue can exceed its original US$1bn target, although total supply remains below demand. Visibility on FY28 production should improve by next March because of an approximately nine-month lead time. TSMC's 3nm capacity is effectively sold out, while 2nm capacity has not yet been broadly allocated; incremental supply could come from TSMC expansions or from reallocations when customers release reserved capacity. Wafer pricing remains dynamic: 2027 pricing is contracted, whereas 2028 pricing remains under negotiation. Management views TSMC as the only realistic foundry partner for Arm's first internally developed CPU because a first-generation advanced chip requires intensive engineering support and advanced-node execution beyond the needs of IP licensing. Intel Foundry was viewed as lacking sufficient customer support and advanced-node experience for this program. Samsung is considered the only credible alternative, but Arm currently expects even its second-generation 2nm CPU to remain at TSMC. The report links the larger CPU opportunity to Agentic AI. Management views published CPU TAM estimates as demand measures rather than feasible supply forecasts: Arm's public estimate is US$100bn, competitors have discussed US$200bn or more, and internal demand scenarios could approach US$300bn. These figures remain highly uncertain and depend on Agentic AI adoption, customer deployment plans and availability of logic, memory and advanced-packaging supply. Arm argues that agents are CPU-intensive because they rely on sequential task execution; this could drive higher CPU requirements and future products with potentially 500 cores. Arm sees datacentre CPUs as an expansion opportunity for customers that want advanced processors but lack hyperscalers' internal chip-design resources. Management attributes Meta's adoption of Arm-based CPUs to superior performance per watt versus x86, arguing that Arm can often provide similar performance at roughly half the power consumption. This matters as power availability becomes an AI-infrastructure bottleneck. AMD and Intel are considered the primary datacentre CPU competitors, Qualcomm a longer-term threat, and Nvidia less of a standalone CPU competitor because of its integrated rack-level strategy. Arm assumes no CPU revenue from Amazon, Google or Microsoft because of their in-house processor efforts. The company also highlighted increasing adoption of Compute Subsystems (CSS), as customers choose pre-integrated subsystems to reduce development time and engineering complexity. It is shifting customers from traditional licences toward subscriptions: about 30 of its top 50 customers used subscriptions last year, with management expecting 40-45 over time. Subscriptions are intended to improve retention and royalty streams, although revenue can remain lumpy because around 50% is recognized at signing and agreements usually run for five years. Management expects AI workloads to migrate gradually toward edge devices as customers seek lower inference costs and hyperscalers shift some computing burden to endpoints. It expects competition increasingly to center on cost per token, power use and efficiency rather than peak compute performance. It also expects CPUs to become more important in inference: training historically required roughly one CPU for every two GPUs, but inference uses more CPU resources, a trend Agentic AI could reinforce. In China, Arm said it can still ship the vast majority of its products and retains strong ecosystem participation. Management stated that its CPU market share is roughly 95% in China versus approximately 50% in the rest of the world, although domestic semiconductor-capacity expansion is slower under continuing restrictions. Internally, Arm has access to OpenAI models and is developing a CPU-design-focused model, but management believes meaningful AI-assisted CPU design remains years away because CPU-design data and open-source ecosystem support are much more limited than in software development. Over the next several years, management's focus is execution as Arm moves from IP licensing toward designing and selling complete CPUs. It targets IP-business operating margins of roughly 65% by 2030, up from about 55% currently, while expecting the CPU business to reach approximately 30% operating margins as it scales. Capital allocation is expected to prioritize engineering recruitment, selective team acquisitions and product development rather than near-term buybacks. Longer term, management identified robotics, autonomous systems and humanoid robots as potential additional sources of CPU demand, while reiterating that it has no current plan to enter GPUs. Bernstein values Arm by applying a 42x multiple to its FY31 EPS estimate of US$11.37, resulting in a US$480 price target. It rates the stock Outperform.

Analysis framework

Bernstein synthesizes management comments from its SDC meeting, then assesses the AGI CPU launch timetable, Meta adoption, supply availability, foundry choices, competitive positioning, licensing-model evolution and long-term AI-driven CPU demand. The target price is based on a price-to-earnings multiple applied to Bernstein's FY31 EPS estimate.

Methodology notes

• Valuation methodsP/E and PEG Valuation

  Price-to-earnings valuation

  Bernstein applies a 42x multiple to its FY31 EPS estimate of US$11.37 to derive the US$480 target price.

• Industry AnalysisSupply-demand framework

  Advanced-node CPU demand versus TSMC wafer supply

  The report frames AGI CPU revenue potential through demand exceeding US$2bn against constrained TSMC capacity, with expansion and reallocation as possible supply releases.

• Industry AnalysisUpstream-Midstream-Downstream Transmission

  Agentic AI workload growth, CPU requirements and semiconductor supply constraints

  The analysis connects Agentic AI adoption to CPU-intensive sequential workloads, then to processor demand, foundry capacity, memory and advanced-packaging availability.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• ARM Holdings PLC (ARM)

  Primary covered company; the report links its growth outlook to AGI CPU commercialization, AI-driven datacentre CPU demand and licensing-model expansion.

  Strengths

  AGI CPU program is in final qualification; Meta adoption; performance-per-watt positioning; CSS momentum; strong China market share; targeted IP margin expansion.

  Weaknesses

  Arm is entering the materially different business of designing and selling complete CPUs rather than licensing IP.

  Comparison

  Management views AMD and Intel as primary datacentre CPU competitors, Qualcomm as a longer-term threat, and TSMC as the only realistic initial foundry partner.

  Risks

  Slower AGI CPU growth, weaker server-CPU share gains, slower royalty-rate increases, slower v9 and CSS penetration, higher ramp costs and higher operating expenses.

Key data

• AGI CPU demandMore than US$2bnManagement said demand exceeds current available supply.
• Initial FY28 AGI CPU opportunityApproximately US$1bnArm initially secured around this level of TSMC supply and now believes revenue can exceed it.
• Meta share of initial FY28 opportunityApproximately 90%Meta supported R&D and qualification and receives preferential pricing.
• Initial shipment timing and revenueCQ1 2027; approximately US$100mExpected if final qualification concludes successfully.
• China CPU market shareRoughly 95%Compared with approximately 50% in the rest of the world, according to management.
• IP-business operating margin targetApproximately 65% by 2030Up from roughly 55% currently.
• CPU-business operating marginApproximately 30%Expected as the CPU business scales.
• FY31 EPS estimate and valuation multipleUS$11.37 and 42xInputs used by Bernstein to derive its US$480 price target.

Impact & implications

Bernstein's central implication is that Arm's AGI CPU upside is increasingly determined by qualification execution and advanced-node supply rather than end demand. The report also sees power efficiency, expanding CPU intensity in Agentic AI inference, CSS adoption and subscription migration as support for Arm's longer-term growth and margin opportunity.

Risks

• Slower growth in the new AGI CPU business could pressure Bernstein's price-target case.
• A lack of market-share gains in server CPUs is a downside risk.
• Royalty-rate increases across end markets could be slower than expected.
• v9 and CSS penetration could progress more slowly than expected.
• AGI CPU ramp-up costs and the operating expenses needed to support projected growth could exceed expectations.