SILICON VALLEY — "Nvidia will be worried about Qualcomm, and AMD and Intel will be bothered by Arm."
The two companies drawing the most attention in the semiconductor industry these days are Qualcomm and Arm. Qualcomm is a mobile chip powerhouse, ranked alongside MediaTek and Apple among the top three in the market for application processors (APs), the brains of smartphones. Arm effectively monopolizes the smartphone architecture market — architecture being the set of instructions that determines how a chip operates — with more than 90% of APs built on Arm technology. Arm's architecture is also showing its strength in physical AI, the physical ecosystems that incorporate artificial intelligence, including self-driving cars.
Qualcomm and Arm have long expanded their businesses around the mobile market, but their course has shifted sharply this year. Qualcomm announced the acquisition of a startup with technology that competes with Nvidia's software and is weighing the design of application-specific integrated circuits (ASICs) for AI companies. Arm unveiled its first in-house AI chip that can be installed in AI data center servers. Qualcomm's custom chips overlap with Nvidia's graphics processing units (GPUs), while Arm's processors for AI servers overlap with the central processing units (CPUs) of AMD and Intel.
As mobile chips are pushed aside by AI server chips in the semiconductor market and smartphone prices soar on memory bottlenecks, the two companies are reducing their dependence on mobile chips. For Nvidia, AMD and Intel, which have dominated GPUs and CPUs in the AI server market, formidable challengers have emerged. Arm Chief Executive Rene Haas said in the company's first-quarter fiscal 2027 earnings announcement that royalty revenue related to data center products more than doubled from a year earlier, and that demand for the new chip lineup was greater than expected.
Qualcomm said on June 24 that it would acquire AI chip startup Modular. Bloomberg, citing sources, reported that the deal was worth $4 billion, or 6 trillion won. Modular is a startup founded in 2022 by former Google developers Chris Lattner and Tim Davis. The $4 billion figure is more than double the $1.6 billion valuation the company was awarded when it raised $250 million in September of last year. The acquisition is set to close within the second half.
Modular develops AI compilers. A compiler is software that converts AI code into machine language so it can run on computer chips. Modular has developed Mojo, an AI programming language that allows AI models to run on AI chips without complex code modifications regardless of the manufacturer — Nvidia, AMD, Intel or Qualcomm — as well as MAX, an AI inference engine. With Nvidia dominating the AI chip market through CUDA, software tailored to its own GPUs, Modular could threaten Nvidia's empire with a general-purpose compiler that helps AI models achieve peak performance across a variety of infrastructure. Bloomberg analyzed that acquiring Modular would allow Qualcomm to take on Nvidia at the software level, beyond simple chip competition.
Qualcomm has been searching for an alternative target since its 2018 bid for Dutch automotive chipmaker NXP collapsed under opposition from regulators. A prime example is its $2.4 billion acquisition last year of Alphawave, a chip design company listed on the London Stock Exchange. Qualcomm is broadening its stage to AI chips by adding Alphawave's chip design technology to its own mobile chip development capabilities. Reuters reported on June 24, citing sources, that Qualcomm was in negotiations to provide custom chip (ASIC) design services to China's ByteDance, the parent company of TikTok.
Qualcomm's work with partners on server CPUs and custom inference chips has been read as a move into the data center market by a company that has led in high-end flagship smartphone chips and now seeks to overcome stagnation in the smartphone market. The Modular acquisition is also seen as a strategy aimed at Nvidia, which commands the data center through its combination of GPUs and CUDA. Qualcomm in fact released its AI200 and AI250 accelerators and server racks, optimized for next-generation data center AI inference, in October of last year.
Qualcomm has not hidden this ambition. In the press release announcing the Modular acquisition, the company said it expects the deal to strengthen its optimized AI computing layer capabilities and the software foundation for its data center strategy, supporting more efficient inference, orchestration and deployment while further solidifying relationships with model developers and hyperscalers, the operators of large-scale computing infrastructure. Qualcomm Chief Executive Cristiano Amon said that through Modular, Qualcomm would combine scalable and energy-efficient data center technology with an open-ecosystem approach and open the next chapter of AI. Modular Chief Executive Chris Lattner said the company would be able to continue advancing Modular's software platform as part of Qualcomm's broad strategy spanning edge computing to cloud computing.
For Intel and AMD, which lead the market for AI server CPUs, Arm is a thorn in the side. It has broken with the tradition of not competing with its customers and thrown down a challenge to fabless chip designers. Haas unveiled the Arm AGI CPU, an in-house CPU for AI data centers, for the first time on March 24 of this year. He called it a defining moment for the company, saying Arm could now support agentic AI infrastructure by offering partners more options built on Arm's high-performance, low-power computing foundation. At Hot Chips 2026, held at Stanford University in the United States from Aug. 23 to 25, Arm disclosed key specifications of the Arm AGI CPU, including the use of up to 136 Neoverse V3 cores.
Arm is the world's largest semiconductor intellectual property (IP) company. Its first in-house chip in the 35 years since its founding amounted to a declaration that, amid intensifying competition among big tech firms to develop their own chips, Arm too would move beyond a business model centered on selling blueprints and enter the chip manufacturing market. The chip is designed to run up to 136 cores — the components that handle instruction processing — at 300 watts, with a focus on improving energy efficiency at AI data centers that require enormous amounts of power. Haas stressed that the new product delivers more than twice the performance per rack — a rack being the structure that houses network equipment — at the same power compared with x86 platforms. X86 is the architecture Intel and AMD follow when building CPUs.
Meta, which took part in joint development of the chip, has been named the first customer, and mass production begins in the second half. Meta, the parent company of Facebook, plans to run its own data centers using both its in-house Meta Training and Inference Accelerator (MTIA) and Arm CPUs. OpenAI and Cerebras have also decided to use Arm CPUs, and SK Telecom (017670) said it would use the new product to expand into inference data center infrastructure. Arm said SK hynix (000660) also participated in the development as a memory solutions partner.
Arm, a British company, has stuck since its founding in November 1990 to a revenue model in which it draws blueprints for chipmakers using its IP and collects only fees. Once it hands over the architecture, major customers such as Samsung Electronics (005930), Apple, Qualcomm and Nvidia design chips from those plans. Even after capturing 90% of the AP architecture market, Arm kept to its principle of not developing chips itself, earning it the nickname "the Switzerland of the semiconductor industry."
But as the semiconductor market has reoriented around AI and demand for inference chips has exploded, that principle of more than 35 years has broken. Securing inference chips has become critical to developing agentic AI that judges and acts on its own, and the standing of CPUs, which handle sequential computation, has risen. As competition over inference AI models heats up, data centers are expected to need more than four times the CPU capacity per gigawatt than before. One analysis even projects that CPU growth will surpass that of GPUs by 2028.
Arm has been preparing for the inference era since SoftBank, led by Chairman Masayoshi Son, acquired it for $31.4 billion in 2016. After entering the data center chip IP market in 2018 and courting big tech firms, Google and Microsoft are now developing Arm-based AI chips. SoftBank also acquired U.S. chip design company Ampere for $6.5 billion last year.
Arm's move into its own chips is expected to send inevitable ripples through the AI semiconductor market. CPUs from Intel and AMD built on Intel's x86 architecture have dominated the data center market, but a formidable competitor has now appeared. Arm could press the chipmakers it supplies with IP to use its own CPUs. The head of one semiconductor startup said Arm can pose an enormous threat to the market with server CPUs alone, even without GPUs, adding that Arm is probably the company AMD and Intel worry about most.
