Arm closed fiscal 2026 with record revenue and a business model that now runs on two visibly different engines: a royalty stream that arrives per shipped chip, and a licensing stream that arrives when large deals close. This article reads the company's own shareholder letters and SEC filing rather than commentary about them.
For anyone licensing Arm IP, the useful part is not the headline. It is the mechanism raising royalty per chip, and the fact that Arm has started selling silicon of its own.
The year and the quarter
Full-year fiscal 2026, which ended 31 March 2026, produced revenue of $4.92 billion — a record, and the third consecutive year since the IPO of more than 20% revenue growth. Royalty revenue was $2.61 billion, up 21%. Licensing revenue was $2.31 billion, up 25%. Non-GAAP earnings per share was a record $1.77.
The fourth quarter was the company's highest-ever revenue quarter at $1.49 billion. Licensing revenue grew 29% year-on-year to $819 million; royalty revenue grew 11% to $671 million, and Arm states that data-centre royalty more than doubled year-on-year.
The two growth rates diverging matters. Royalty growing 11% while licensing grows 29% in the same quarter is not a contradiction — it is what happens when a chip-attached annuity and a deal-timed licence book sit in the same income statement.
Why royalty per chip keeps rising
Arm names three mechanisms, and all three are structural rather than cyclical:
- Architecture mix: newer Armv9-based designs carry a higher royalty rate than Armv8-based ones, so the royalty per chip rises as the installed base migrates
- Compute Subsystems: buying a pre-integrated, pre-verified configuration rather than discrete IP raises the value Arm captures per chip
- More Arm content per chip: GPUs, NPUs, interconnect and system IP alongside the CPU raise the royalty base independent of unit volume
The royalty rate you will not find in the filing
One caution that is worth stating plainly: Arm does not publish a blended royalty rate. The per-chip percentages that circulate in analyst notes and press coverage are third-party estimates, not company disclosure, and they disagree with each other by wide margins. Model royalty from your own negotiated terms, not from an estimate quoted as fact.
Compute Subsystems: what actually changed in what you buy
CSS is the packaging change that matters most to a design team. Instead of licensing a CPU core and integrating it yourself, you license a pre-integrated, pre-verified configuration — CPU cluster, interconnect, memory subsystem and security IP delivered as a hardened block.
Arm reported 21 CSS licences across 12 companies, with five customers already shipping CSS-based chips and two of those on a second-generation platform. The top four Android smartphone vendors are all shipping CSS-powered devices.
Arm also describes CSS as an annual refresh in smartphones, with royalty rates generally increased year on year as each new generation ships. That turns an architecture step-up into a recurring escalator. For a licensee it cuts integration risk and time-to-market — and it moves more of the per-unit cost into the royalty line, which is a modelling decision worth making explicitly rather than discovering at volume.
The data-centre shift
Arm states that Neoverse deployments have passed one billion cores and that its share of CPU compute among top hyperscalers is about 50%. Reported core counts illustrate the direction: AWS Graviton at 192 cores on Neoverse CSS V3, NVIDIA's Vera CPU at 88 Arm-based cores, Microsoft Cobalt 200 at 132 cores on Neoverse CSS V3.
Google is described as replacing x86 host processors with custom Arm-based Axion CPUs for its next-generation TPUs, with Arm citing up to 2.7 times better training performance-per-dollar and up to 80% better inference performance-per-dollar against the previous x86-based generation.
Those are vendor-stated comparative figures. They are useful for understanding where the platform is heading, and they should not be read as a benchmark you will reproduce.
The silicon pivot: Arm AGI CPU
On 24 March 2026 Arm introduced the Arm AGI CPU, described as its first production silicon product, purpose-built for agentic AI. Arm states it will deliver more than twice the performance per rack compared with x86-based platforms and that this allows AI data-centre capital expenditure to be reduced by up to $10 billion per gigawatt — both figures labelled in the filing as based on Arm internal estimates.
Meta is named as lead partner and co-developer on a multi-generation roadmap. Arm reports more than $2 billion of customer demand across fiscal 2027 and 2028, against a stated forecast of $15 billion for the business, and names Cerebras, OpenAI, Positron and Rebellions as integrating the part alongside accelerator-based systems. Systems are stated as orderable from ASRock, Lenovo, Quanta and Supermicro. More than 50 companies are listed as supporting the expansion, including AWS, Broadcom, Google Cloud, Marvell, Microsoft, Micron, NVIDIA, Oracle, Samsung, SK Hynix and TSMC.
For a team licensing Arm IP, this is the one item with a strategic edge to it: the company you license CPU IP from now also ships a CPU. That does not change an existing licence, but it is worth knowing when you negotiate field-of-use terms, and it is a fair question to ask about roadmap overlap in the segment you are designing into.
What to check before licensing Arm IP
None of the above changes the practical checklist, but it sharpens two lines on it:
- Which architecture generation and which delivery form — discrete core IP or CSS — the licence actually covers, because the royalty per chip differs
- Whether the core or subsystem is characterised and available on your specific process node and PDK release, not just on the platform name
- What the per-unit royalty is at your expected volume, written into the agreement rather than inferred from an external estimate
- Whether the licence covers the full software and toolchain support you need for your product lifetime
- Field-of-use terms, given Arm now ships silicon in the data-centre segment
