Arm's share price rests on 'stories'

Revenue growth in FY 26

Arm reported that during FY26 it recognised $704.4 million of revenue from licensing and service arrangements with a SoftBank Group affiliate, versus only $145.5 million in FY25. SoftBank owns around 86% of Arm Holdings and has ambitions of becoming an AI infrastructure company. Consequently, it has funded many startups in robotics, autonomous driving, AI hardware etc. using Arm’s ecosystem, creating a peak in licensing deals. Licensing revenue from external customers declined from $1.4bn to $1.3bn.

Expenses and margins nowhere near financial model

ARM targets non-GAAP operating margin for its IP/CSS business in FY31 to be greater than 65%. However, in the last three years non-GAAP operating margins have been 43%, 46.7% and 43.6% respectively, despite the revenue CAGR of 23% for those years. One of the major things that the market does not focus on for ARM’s expenses is its share-based compensation, which has been massive, bringing in further dilution by increasing the number of shares outstanding. If we take them into consideration, GAAP operating margins for the last three years have been 18.3%, 20.7%, and 3.4% respectively.

Share based compensation is not free money to improve margins. By that logic, any company can improve its margins by paying the majority of expenses in shares. Big Tech companies spend billions in buying back shares which have been vested by their employee share based compensation. ARM, however, has not done any major share buyback programme since going public, because it has not generated any extra cash flow. Consequently, the number of shares outstanding have been gradually growing, and will continue to grow as more shares are vested over time. This dilutes current shareholders.

Growth in R&D and S,G&A expenses both outgrew the revenue growth achieved in FY26, with or without share-based compensation considered.

Snapshot of expenses and increasing number of shares outstanding

Free cashflow generation has been low compared to revenue

The dream - AGI CPU strategy

Before launching its AGI CPU strategy, Arm's business was built around being a neutral IP supplier to the entire semiconductor industry. The move into selling its own server CPUs fundamentally changes that position by putting Arm in direct competition with many of its own licensees, including Qualcomm, NVIDIA, Marvell and others. While Meta is reportedly a key customer, Qualcomm has also announced Meta as a customer for its server CPU roadmap, suggesting Meta intends to multi-source rather than rely on a single vendor. More broadly, the addressable market is concentrated among a handful of hyperscalers, many of which are increasingly developing their own custom CPUs, limiting the size of the merchant CPU opportunity.

The strategy also introduces execution risks that Arm has never faced as an IP licensing company. Manufacturing and selling chips require managing foundry capacity, packaging, inventory, customer support and supply chain operations, all of which are capital- and execution-intensive. In addition, merchant silicon carries significantly lower margins than Arm's traditional licensing business, reflected in its long-term target of over 30% operating margin for the AGI CPU business versus over 65% for the IP business. Finally, ARM risks eroding the ecosystem neutrality that has been its greatest competitive advantage, potentially straining relationships with partners and inviting greater regulatory scrutiny over whether it can remain an impartial steward of the ARM architecture while competing directly with its own customers.

ARM estimates that CPU market for AI in the agentic era will grow from $25bn to $100bn within a span of five years, as agents will be running all the time requiring more CPUs. Hardware manufacturers are excited, but there has been no successful software yet using the agentic systems. The issues regarding consent, privacy, security all seem too big for now on the software layer for a fully autonomous system. Costs of running LLMs have been a major issue with companies as frontier model providers have started token-based subscription. The unreliability of ‘how many tokens’ will it take exactly makes it impossible for companies to calculate their own ROI on deploying AI. Furthermore, the unit economics for LLM providers like OpenAI and Anthropic have not been working, is now widely known. All in all, there are big doubts about whether this kind of capex investment can continue till FY 31. 

Valuation

In its new strategy, by FY 31, ARM expects $10bn revenue from its IP business with > 65% operating margin and $15bn revenue from its chip business with > 30% operating margin. If we take 65% and 30% as our targets (although very unlikely), we end up with $11bn operating profit on $25bn sales. Taking a very optimistic only 20% tax rate, earnings will be $8.8bn and the number of shares outstanding are already 1.068bn in FY 26. We expect this number to rise further by FY 31. Our back of the envelope EPS comes out at $8.24. The company reports expected >$9 in FY 31. 

The shares are trading at $330 so around 40x EPS, expected five years from now, which is also very optimistic by any estimation and involves major customer, execution, and market risks. Current P/E stands at 230X. All thanks to artificial intelligence's artificial demand generation through a free float of only 12% and selling big dreams. It’s a pass for us, despite the noise.