The stock price soared by 30% against the trend. Has ARM, which has started making AI chips, achieved a great victory?

When AI Agents no longer solely rely on GPUs, ARM has finally entered its own era.

In September 2023, ARM went public on the NASDAQ at a price of $51 per share. SoftBank sold less than 10% of its shares, cashing out approximately $4.9 billion. It was the largest technology IPO of that year, but many people had a question in mind - How can a company that doesn't manufacture chips but only sells design blueprints justify a valuation of $54.5 billion?

In less than three years, ARM provided an answer that no one had anticipated.

On May 21, 2026, ARM's stock price skyrocketed by over 15% in a single day, reaching a record high of $259. Starting from $201 at the end of April, it increased by 27% in just three weeks. During the intraday trading on May 22, the stock price reached a peak of $285.

Since the IPO price of $51, in less than three years, the stock price has more than quintupled.

This is not just a simple market rally driven by earnings reports or analyst recommendations. ARM is undergoing the most profound transformation in its 35 - year history. Can it completely turn into a different company?

AGI CPU: The First Time in 35 Years to Manufacture Its Own Chips

To understand ARM's recent market performance, we need to first figure out how it used to make money.

ARM's business model is unique in the semiconductor industry. It doesn't produce any chips but licenses the underlying chip designs to other companies - Apple, Qualcomm, MediaTek, Samsung, and even NVIDIA. Customers pay an upfront licensing fee to obtain the design blueprints and then pay ARM a royalty for each chip produced based on the ARM architecture.

According to ARM's disclosure during the IPO, the average royalty per chip is about 5 cents.

Thanks to this model, ARM's architecture has entered 99% of smartphones globally. As of before the listing, the cumulative shipments of chips based on ARM designs exceeded 160 billion. However, the problem is obvious - The global mobile phone market has stopped growing significantly.

ARM's revenue in the fiscal year 2023 was $2.68 billion, showing a slight year - on - year decline. A company that dominated the entire mobile computing era has seen its revenue stagnate. The 5 - cent - per - chip royalty model is destined to bring only linear growth in an era when the mobile phone market has reached its peak.

To break through the ceiling, ARM must find a new growth engine.

ARM's AGI CPU | Image source: Livemint

On March 24, 2026, ARM held a press conference in San Francisco and announced something it had never done in its corporate history - Releasing and selling its own designed finished chips.

This chip is named "AGI CPU". Based on ARM's latest Neoverse V3 architecture, it has up to 136 cores, is manufactured using TSMC's 3 - nanometer process, and has a TDP of 300 watts. It is not for training large models (that's the job of GPUs) but is specifically designed for the "CPU side" work in AI data centers - data pre - processing, model inference scheduling, network management, and most importantly, The workflow orchestration of AI Agents.

CEO Rene Haas said at the press conference: "This is a critical turning point for the company."

This statement carries much more weight than it seems. ARM has adhered to a principle for the past 35 years - it only sells designs and doesn't compete with its customers. Apple uses ARM's architecture to make its M - series chips, Qualcomm uses it for Snapdragon, and NVIDIA uses it for Grace. Everyone gets what they need and there is no conflict. Now that ARM is entering the chip - manufacturing business itself, in essence, it is competing with its own customers for the same market.

However, ARM's confidence lies in the fact that it has found a new battlefield that its customers haven't had time to occupy.

The Underlying Logic of Data Centers Is Being Rewritten

In the past few years, everyone has been talking about GPUs. NVIDIA's H100 and B200 are the hard currency in the AI era, and the construction logic of data centers revolves around GPUs. What about CPUs? They are just supporting roles, and are usually configured in a ratio of 1:4 to 1:8 alongside GPUs.

But Agentic AI is changing this equation.

Different from traditional chatbots, AI Agents need to autonomously call tools, manage multi - step tasks, and coordinate multiple subsystems. These tasks are essentially serial and logic - intensive, which are exactly what CPUs, rather than GPUs, are best at. When an Agent is busy calling APIs, waiting for external responses, and conducting security checks, the expensive GPU beside it is actually idling.

ARM's judgment is that In the era of AI Agents, the demand for CPU cores per gigawatt of data center power will soar from 30 million to 120 million - a four - fold increase. TrendForce's research also points out that the ratio of CPUs to GPUs in Agentic AI deployments will change from the current 1:4 to 1:8 to 1:1 to 1:2.

Morgan Stanley further predicts that by 2030, Agentic AI may bring an additional demand of $32.5 billion to $60 billion to the data center CPU market, and the entire server CPU market size will exceed $100 billion.

To put it simply, GPUs are the "engines" in the AI era, but the Agent era also needs a "dispatching center", which is the CPU. This is exactly the position that ARM's AGI CPU is targeting.

ARM's data center products including the AGI CPU | Image source: ARM

ARM's AGI CPU is not just a PPT product. Its list of initial customers reads like a "Hall of Fame" in the AI industry - Meta is a joint - development partner and the first - launch customer, and OpenAI, Cerebras, Cloudflare, SAP, and SK Telecom have all signed commercial contracts. Server manufacturers such as Lenovo, Supermicro, and Quanta have started to offer complete systems equipped with the AGI CPU.

Santosh Janardhan, the head of Meta's infrastructure, said that Meta and ARM jointly developed this chip and will deploy it in conjunction with Meta's self - developed MTIA accelerator to optimize the computing density of its "gigawatt - scale" AI data centers. Sachin Katti, the head of industrial computing at OpenAI, said that the AGI CPU will become an important part of OpenAI's infrastructure, responsible for coordinating the orchestration layer of large - scale AI workloads.

What really excites Wall Street is the growth rate.

ARM revealed in its earnings conference call in early May that the committed demand for the AGI CPU from customers in FY2027 - 2028 has exceeded $2 billion - And this figure was only $1 billion six weeks ago when the product was just launched. Doubling in six weeks, this kind of demand acceleration is extremely rare in the hardware industry.

Meanwhile, ARM's traditional business is also accelerating. In Q4 of the fiscal year 2026, its revenue reached $1.49 billion, a year - on - year increase of 20%. Among them, licensing revenue soared by 29% to $819 million, setting a record for the best single - quarter performance.

ARM's Re - valuation

What really triggered this market rally was a heavyweight research report released by Bernstein analyst David Dai on May 19.

Dai initiated coverage of ARM with an Outperform rating and directly set a target price of $300. His core logic is clear - AI is shifting from chatbots to Agentic AI, which will bring a structural increase in CPU demand, and ARM will be the biggest winner due to its energy - efficiency advantage. He predicts that by 2030, the server CPU market will quadruple to $137 billion, ARM's annual revenue will exceed $26 billion, and its earnings per share will reach $9.83.

Bernstein is not alone. Jefferies raised its target price from $210 to $290, TD Cowen from $165 to $265, and Bank of America set it at $300. Citi also confirmed the management's target for 2031 - a total revenue of $25 billion and earnings per share of $9.

This is not just the optimism of one institution but a collective reconstruction of ARM's valuation framework by the entire Wall Street. The market is no longer pricing ARM as a "chip IP licensing company" but is starting to value it as an "AI infrastructure platform".

Interestingly, in the same month when ARM launched its AGI CPU, NVIDIA also announced that it would sell its ARM - based Vera CPU as an independent product. On May 19, NVIDIA vice - president Ian Buck personally delivered the first batch of Vera CPU systems to Anthropic, OpenAI, Oracle, and SpaceXAI. A GPU giant and an IP licensing giant entered the CPU market in the same month - This itself is the most powerful endorsement of the judgment that "Agentic AI is reshaping CPU demand".

Looking back at ARM's own business logic. In the past, by selling blueprints, it received royalties ranging from 5 cents to $2 per chip. Now, by manufacturing and selling its own chips, the selling price of a data center CPU is in the thousands of dollars. Jumping from the bottom to the middle - upper reaches of the value chain, the unit revenue has increased by thousands of times.

Of course, the risks are also huge. ARM's current price - to - earnings ratio is close to 300 times, and the price - to - sales ratio is close to 167 times. This valuation is not for ARM in 2026 but for ARM in 2030 or even further in the future. If the implementation of Agentic AI is slower than expected, if there are execution errors in the process from IP design to large - scale chip production, or if the self - developed chips offend existing licensing customers and lead to a deterioration of the relationship - any problem in these links may cause a sharp correction.

It is worth noting that ARM insiders have sold approximately $31.9 million worth of stocks in the past three months. The management is using actions to tell you that they also think the current price is not cheap.

However, the market is clearly willing to pay for this story. Because what ARM is doing is transforming from a "tax collector" to an "infrastructure provider". If Agentic AI really explodes as everyone expects and the CPU demand really quadruples, then today's ARM may just be at the beginning.

For the first time in 35 years, ARM is holding not just a blueprint but a real chip in its hands. Whether this chip can support a brand - new ARM may not be known until after the mass - production ramp - up in 2028.

But at least for now, Wall Street has cast its vote of approval in advance.

This article is from the WeChat official account “GeekPark” (ID: geekpark), author: Hualin Dancer, editor: Jingyu. Republished by 36Kr with permission.