An Arm in Every Pie

Except the Most Delicious

From a quiet, predictable industry, the semiconductor market has turned into a quagmire of hourly news that either predicts the arrival of the singularity’s endless prosperity or the end of the universe. The delivery vehicles are AI-generated catastrophizing infographics on LinkedIn that are sending AI stocks on a zigzag course. I don’t even want to think about what the dystopia formerly known as Twitter looks like these days. I am too young and innocent to ever open that app again.

While the industry is under bombardment from the open blender of the Trump administration, it does not crack easily. Time after time, the industry has proven highly resilient, and businesses have found a way.

With the US excursing in the Gulf, the industry's helium supply is now in jeopardy, just as the neon supply was during the Russian bombardment of Mariupol in Ukraine.

Tariffs and embargos have had an impact, but not the intended one. Working as effective shields against foreign competition, these have allowed Chinese start-ups to achieve sustainability and continue building the AI and Semiconductor Tool industries without economic aid from the Chinese government.

You know that I follow all the news, but I do not respond to every hourly catastrophic event. While I do not have a crystal ball, I select the events and timing of my posts to align with the arrival of relevant data that can contribute to the conversation about what is happening in the most important industry ever.

My mental model for announcements can be seen below:

Announcement does not mean Action

Action does not mean Scale

Scale does not mean Completion

Completion does not mean Outcome.

While I do write about announcements as I will today, it will not be treated as a binary that can send stocks up or down like the AI-generated short-selling slop.

I will let people discuss if announcements are “true” or “false”, but the reality is that time will tell.

My approach is to keep the analysis and the opinions separate. They might be in the same paragraph, but my aim is to separate them clearly. I am not trying to sell you anything but a subscription to my blog, and I believe that is best done through a thorough analysis with my opinion on top.

The most important development in the Semiconductor industry right now, apart from the idiotic war in the Gulf, is the memory shortage, or, more correctly framed, the down cycle in investments by memory companies.

Two recent announcements could have affected the memory market, but only one did.

The announcement of Google’s TurboQuant model, which allegedly reduces memory usage by at least 6x for LLM inference. The practical effect is that the same hardware can support much longer context windows or larger batch sizes without running out of memory. That directly affects the memory intensity of inference workloads, especially the DRAM/HBM needed.

While memory stocks had already responded to the Micron result on the 15th of March, they were all sent lower by the TurboQuant announcement on the 25th of March.

The market quickly formed an opinion about the announcement and sent the shares down. This might be the proper response, but it was not formed based on analysis. An announcement was sufficient to create a response.

This was very similar to the response to the launch of DeepSeek in late January of 2025.

Before you think that I am turning into a stock advisor, I want to remind you that I believe announcements are overrated as a source of information.

Both responses are based on the theory that lower (memory) cost models lead to lower memory sales, but there is no evidence that this is the case.

So far, the AI buildout has followed Jevons Paradox, which argues that increased efficiency makes a resource relatively cheaper, encouraging greater demand that outweighs the initial savings. AI companies are not trying to achieve a certain token cost, but to achieve a better token cost than other AI companies. Cheaper tokens mean more use, not lower investments. This can obviously change, but that does not change my opinion that announcements are overrated compared to longer-term analysis. As always, feel free to form your own opinion.

By now, you know this post is going to be about another announcement, and that is a competing one from the ARM Everywhere event. For me, this is an excuse to finally dive into the financials of the most important IP company supporting the AI revolution.

This is not an argument for or against the growth of the memory market. That is already decided for the foreseeable future, unless the war in the Gulf continues or expands. Then all bets are off.

An Arm in every pie

ARM began in 1990 as a very British sort of triumph. Clever engineers, Cambridge roots, and a business model built on brains rather than blast furnaces. It grew from a joint venture involving Acorn, Apple, and VLSI into one of the few genuinely world-class technology assets Britain could point to without clearing its throat first.

Then came the familiar national ritual: celebrate the heritage, sell the asset. Arm was taken private by SoftBank in 2016, and while it still trades on British intellectual cachet, control has long since moved elsewhere.

In that sense, it now sits comfortably beside the great British automotive tradition — beloved at home, owned abroad, and still wheeled out as proof that Britain remains exceptional at producing things other people eventually buy.

ARM Holdings was spun off from the Acorn RISC Machine in 1990, with expertise in low-power computing and set to conquer the world.

In 2016, ARM was acquired by SoftBank, which still holds a majority stake in the company today. The ownership made it possible to invest and expand into other verticals.

The value of ARM was broadly recognised, and Nvidia tried to acquire the company in the early twenties, but the $40B deal failed due to regulatory hurdles, leading to a public offering on Nasdaq in 2023.

By then, ARM controlled most of the IP in the semiconductor industry, leading to more than 350B devices shipped with ARM IP.

Up until now, ARM's business has been based on License revenue and royalties, but the recent ARM Everywhere event will change that, as ARM unveiled its first merchant processor. While I will let brighter heads deal with the technical aspects of the ARM AGI CPU, it comes in the shape of a CPU board and various server configurations, from a 36KW air-cooled open rack with 8160 CPU cores and 180TB of memory to a 200KW liquid-cooled rack with 45,696 CPU cores with over 1 petabyte of low-latency memory. This is very similar to Nvidia's business model, but for CPU-only, whereas Nvidia’s servers are GPU/CPU AI servers.

While CPU-based servers sound boringly early-2020s and like something Intel and AMD are fiddling with as they try to get to AI leadership, the ARM Everywhere event revealed insights that are already showing up in Intel and AMD’s earnings calls. The increased demand for server CPU’s.

Rene Haas of ARM expects CPU consumption in data centres to increase from 30M CPU cores to 120M per GW of capacity, a 4x increase due to the explosion of the AI agentic workload. The agentic workloads are 15x more compute-intensive, and while the accelerators generate tokens, the CPU handles scheduling.

Since CPUs are much cheaper than Nvidia GPUs, this could move the needle on memory, as overall memory consumption will increase. As the GPU-to-CPU ratio slides, an expensive GPU will be replaced by several cheaper CPUs, each with a similar amount of memory.

Once again, I will reiterate my mental model for announcements and let others speculate whether the claim has substance.

It is time to dive into the financials of ARM

The first thing to understand about ARM is its two revenue streams: Royalties and Licenses.

The revenue architecture functions as a two-stage financial engine where licensing serves as the leading indicator of future market dominance and royalties act as the ultimate harvest of that success. The time span is typically 2 to 3 years for a License to turn into royalties.

The characteristics of the two revenue streams can be seen below:

In 2021, the ARM business model transitioned from single, distinct IP licenses to a subscription-based "all-you-can-eat" model. This grants customers access to a broad portfolio of IP for experimentation and tape-out, making their R&D costs more predictable. For ARM, this drastically reduced sales churn and shifted financial tracking toward a stable Annualised Contract Value (ACV).

ARM’s revenue by type is shown below.

After a steady decline from a mid-sixties per cent share, royalty revenue fell below 50% until 2025, then began to surge. This is mainly due to newer business models kicking in.

In the newer, more sophisticated Compute Subsystems (CSS) model, the licensing phase is no longer just about providing access to basic IP blocks but rather about delivering a pre-integrated, verified platform that significantly reduces a customer's time-to-market. This high-value licensing is directly linked to a royalty multiplier effect, in which Arm can command per-unit fees roughly double those of standard implementations.

While ARM has been working on a custom solution for the stargate project that also involves Softbank, the introduction of the ARM AGI CPU will expand the business model into competing with its customers in traditional chip sales.

This move is obviously risky but also carefully calculated, as the main competitors are x86-based CPUs from Intel and AMD. Much riskier is it to poke the 500-pound gorilla in the server jungle, but ARM has likely made a deal with Nvidia that gave its blessing in the event, along with other praiseomercials from Broadcom, Marvell, Google, AWS, Microsoft, Oracle, Micron, SK Hynix & Samsung.

One thing that should not be underestimated is ARM's relationship with key companies in the semiconductor industry, and ARM's IP is instrumental to non-merchant semiconductor companies, from Apple to the large hyperscalers.

The company is punching above its weight and getting away with more than a normal partner would. Over the last two years of earnings calls, ARM has name-dropped a parade of important semiconductor, cloud, and other companies.

The different business models have not emerged as a random evolution but are part of a carefully crafted strategy. As ARM is unavoidable in low-power processing and most other semiconductor IP, the strategy is less about attracting customers and more about shaping the customer journey in a meaningful way.

From a transactional commodity IP block that could be interchanged between suppliers, ARM has evolved its business model towards stickiness and increased royalty content. This has all been driven by the company’s relentless focus on Annualised contract value. Turning customers into long-term partners.

The entry points were set wide apart. The ARM Flexible Access model allows new customers and startups to experiment with different IPs at relatively low cost, while the all-you-can-eat ARM Total Access model gives customers predictable Royalty costs in return for a predictable Annualised Contract value.

The later additions of CCS and now Silicon as a service drive both the stickiness and the Royalty value

At the same time, the architectural models also evolved away from a transactional engagement. Both the V7 and V8 architectures gave customers a volume track with relatively low per-unit royalty/license value, while the V9 architecture is much stickier and more valuable.

While most companies have a distribution of value and stickiness that goes through the middle of the chart, ARM has an Ikean strategy of directing the customer journey around the middle of the chart. It is like an ABC customer model with no B customers.

The outcome of the strategy is shown below.

As can be seen, the total number of licensing customers has plateaued, while the Total Access customers continue to grow. The licensing revenue from AFA is quite low, as the entry fee is around $85K, while the revenue contribution from these customers are approximatly half of ARMs’ revenue.

The royalties by market is show below.

This highlights the growing reliance on mobile applications over the last couple of years. This is not surprising given ARM’s legacy of low-power compute, but not a great growth story in a market that is entirely driven by the AI cloud applications.

The relative weakness of ARM in client and server applications stems from the dominance of the x86 architecture, driven by Intel and AMD.

While the share of ARM architecture has been increasing, the share of x86 architecture has remained resilient over the last few years. The x86 applications have not traditionally been ultra-power-sensitive, except in the laptop segment, but this is changing rapidly in the server market, and gives ARM a chance to attack the x86 fortress. This is one of the major drivers of ARMs strategy.