Author: Godot
After the market closed yesterday, Arm released its financial report for the fourth quarter of fiscal year 2026. With the official announcement of AGI CUP, Arm has transformed from not manufacturing chips itself and only providing instruction set architecture to becoming an AI CPU provider. Its business model has changed from IP licensing to becoming a supplier of customized silicon and computing subsystems.
However, Wall Street is divided on Arm's future. One side bets that AI agents will completely reshape the CPU landscape.
On the other hand, there are concerns about issues such as memory costs and TSMC's capacity constraints, and concerns that ARM's current extremely high valuation multiple has severely overdrawn future expectations of perfect performance.
1) Firstly, at the data level
Total revenue for the fourth quarter of fiscal year 2026 was $1.49 billion, a 20% year-over-year increase, exceeding Wall Street's consensus estimate of $1.47 billion.
Total revenue for the year reached a record $4.92 billion, up 23% year-over-year.
There is a clear discrepancy between licensing revenue and royalties.
Licensing and other revenue rose 29% year-on-year to $819 million as chip design companies accelerate their acquisition of next-generation architecture licenses from Arm in order to seize the AI opportunity.
A leading indicator of the long-term health of the licensing business, annualized contract value (ACV), grew by 22% year-on-year, exceeding management's mid-to-high single-digit long-term target.
Royalty revenue was $671 million, up 11% year-over-year, but below the market's previous expectation of $700 million.
The booming front-end licensing market coupled with sluggish back-end royalties suggests a pessimistic outlook for the recovery of downstream consumption.
In terms of cost and expense control, Arm continues to increase its R&D investment in AGI CPUs, computing subsystems and full-chip solutions.
Non-GAAP operating expenses reached $734 million in the fourth quarter, a 30% increase year-over-year. Full-year operating expenses surged 33% year-over-year to $2.717 billion.
Fortunately, the IP licensing model comes with high gross margins, with the non-GAAP operating profit margin at 49.1% in the fourth quarter and maintained at 43.0% for the whole year.
In terms of cash flow and balance sheet, net cash flow from operating activities was $1.524 billion, and non-GAAP free cash flow was $882 million.
The company had US$3.601 billion in cash, cash equivalents, and short-term investments at the end of the period, providing ample resources to secure advanced wafer and packaging capacity from TSMC.
2) Multi-narrative: Agentic AI pushes the CPU back to the center
Starting this year, AI is entering the Agentic AI stage, represented by Openclaw, which is also known as intelligent agent AI.
Traditionally, large models are static, waiting for user input prompts. Intelligent agents (AI) are automatic, with the CPU handling control and orchestration.
Arm's internal data shows that in complex agent workflows, up to 50% to 90% of system latency is actually caused by insufficient CPU scheduling capabilities.
Hardware configurations were thus forced to be adjusted. In the past, the ratio of CPU to GPU in AI training clusters was often 1:8 or even lower.
But this ratio is changing to 1:4, and may even approach 1:1 in the future.
Arm's conservative estimate is that the future demand for CPU core density in data centers will surge from the current 30 million cores per gigawatt of power to 120 million cores, a fourfold physical expansion that is almost certain.
At the same time, each concurrent agent needs to maintain a large key-value cache and complex context states, and the CPU platform also has to undertake massive memory management tasks.
UBS predicts that the total potential market size for server CPUs will surge from a baseline of approximately $30 billion in 2025 to $170 billion in 2030, nearly a fivefold increase.
Furthermore, since supercomputing centers will place extreme emphasis on power efficiency and high-density scalability in the era of Agentic AI, Arm's reduced instruction set architecture, with its low power consumption advantage, is expected to capture most of the new market share.
Arm's market share in the server CPU market is projected to jump from about 15% in 2025 to 40% to 45% by 2030.
This is precisely the core basis for Evercore ISI and other institutions' bullish view on Arm's long-term explosive potential.
At the product level, just six weeks after the release of Arm AGI CPUs, orders from customers spanning fiscal years 2027 and 2028 have already more than doubled from the initial estimate of $1 billion to over $2 billion.
In addition to Meta, Verda, a European AI cloud provider, has clearly stated that it will deploy AGI CPUs on a large scale in its next-generation infrastructure, in conjunction with NVIDIA's GB300 and Vera Rubin systems for intelligent agent computing power orchestration.
Cerebras, OpenAI, and ODM manufacturers such as Lenovo, Quanta, AMD, and ASRock have also launched server solutions based on this chip.
3) The hidden concern is the production capacity of TSMC's packaging and HBM.
However, Arm expects that substantial sales revenue from the first batch of silicon wafers produced will not be recognized until the fourth quarter of fiscal year 2027, with large-scale production not expected until fiscal year 2028. This represents a mismatch of more than a year between demand and financial realization.
Arm AGI CPU architecture integrates a large amount of HBM, and TSMC's exclusive CoWoS packaging is currently the only mature and reliable interconnect solution.
Compared to NVIDIA and AMD, Arm, as a newcomer, is at a disadvantage in TSMC's priority list. HBM supply is also tight.
In the next 12 to 18 months, Arm's core competitiveness will temporarily shift from its architectural design capabilities to its ability to seize supply chain capacity.
4) Bearish perspective: Valuation has already been overextended
Therefore, the cautious camp, represented by Vivek Arya of BofA Securities, downgraded Arm's rating to neutral and lowered its target price range to $120 to $135.
Its core short-selling logic includes two points.
First, smartphones are facing a severe headwind from rising memory costs, and Arm's traditional royalty revenue growth is nearing its peak. The decline in royalty growth to 11% this quarter confirms this pressure.
Secondly, the company's recent rapid growth in licensing revenue is heavily reliant on internal cash flow within its parent company, SoftBank. These related-party transactions account for nearly 30% of total licensing revenue, raising concerns in the market about revolving financing and revenue quality.
BofA stated that Arm's current valuation multiples significantly overdraw on future expectations of perfect execution.
The $2 billion AGI CPU order, constrained by a global shortage of packaging capacity, cannot be converted into substantial cash flow to support the financial statements before fiscal year 2028.
This is the root cause of the divisions on Wall Street.
In conclusion,
Arm stands at a crossroads, fraught with both architectural advantages and production capacity bottlenecks. Agentic AI offers the most imaginative ticket to the next five years, but to cash in on this ticket, it needs to navigate TSMC's packaging facilities, HBM's supply chain, and so on.
The bulls see a $170 billion market in 2030, while the bears see a cash flow vacuum before 2027. Both seem correct, just at different paces.
Attached is the link to the financial report
