Intel's stock surged this year. Trading at $39.4 on January 2 (U.S. Eastern time), the stock soared to $132.2 on the 23rd of this month. It more than tripled in half a year, and the investment industry believes Intel still has room for further gains. On the 11th of this month, Bank of America (BofA) raised its investment rating on Intel by two notches, from sell to buy. BofA forecast that the artificial intelligence (AI) agent boom would drive growth in Intel's central processing unit (CPU) business.
On the 1st of this month, Nvidia announced at "GTC Taipei 2026" that it is mass-producing "Vera," a CPU for data center servers, and will ship the product this autumn. Nvidia's blueprint is to expand its product ecosystem, which has been built around graphics processing units (GPUs), and reinvent itself as a full-stack AI company. Nvidia CEO Jensen Huang predicted, "The CPU market for AI agents will grow much larger than before."
The CPU is rising again. Since the ChatGPT moment in 2022, the protagonist of AI infrastructure has unquestionably been the GPU. Whether for foundation models or services, creating AI required securing GPUs first. AI model developers such as Google and OpenAI, as well as global conglomerates and startups seeking to internalize AI technology, all jumped into the battle to secure GPUs. GPU demand exploded, and an asymmetry of insufficient supply persisted for some time.
The reason the GPU's value emerged in the early days of the AI era lies in the correlation between AI training and the GPU's performance characteristics. The AI development process is broadly divided into two stages. The first is the training stage, in which large volumes of data are fed into an AI model to repeatedly accumulate knowledge. The second is the inference stage, in which an AI model draws conclusions about newly input data based on its existing learned knowledge.
When the priority is to create a usable AI model—in other words, when AI training matters—the GPU is in the spotlight. The GPU is designed to execute complex computations in parallel and pursue efficient work. The GPU is essential when performing large volumes of identical commands simultaneously. The same is true for AI model training. Because rapidly training large volumes of data is important, parallel computation is needed, and the GPU has been put to good use here. Yoo Eung-jun, CEO of June AI Consulting and former head of Nvidia Korea, explained, "In the AI training stage, parallel computation that processes hundreds of millions to hundreds of billions of pieces of data simultaneously is important," adding, "Data center GPUs are equipped with thousands of computing cores, optimized for AI training."
It is not that the CPU lacks computing functions. Comparing only the performance of cores—the components that directly execute computations—the CPU's core performance is superior to the GPU's core. However, the CPU specializes in serial computation rather than parallel computation. In the first place, the CPU is a product designed to process the many commands issued inside a computer in sequence. It directly handles all the tasks needed to correctly execute all of a server's software. It is a so-called "do-everything-yourself" style. However, from the perspective of needing to perform large-scale data training simultaneously, this way of carrying out tasks is less efficient. Naturally, the AI industry's attention turned to the GPU.
Four years have passed since the ChatGPT moment. The market's gaze, which had been concentrated on training, has gradually shifted to inference. As the level of AI training reached a certain orbit, the importance of inference—the stage of executing AI services—came to the fore. In particular, the AI agent boom became the catalyst that elevated the value of the CPU. In the past, AI sufficed if it gave one answer to one question, like a chatbot. Now, multiple AI agents each carry out their own commands, synthesize them to complete a mission, and even perform post-verification on their own. It is the era of agentic AI. In this process, AI agents cross between online and offline and even utilize external tools. This is where the true value of the CPU shines.
Park Jeong-guk, chief technology officer (CTO) of Elice Group, which operates a modular data center business, analyzed, "The CPU prepares data, distributes work to the GPU, and controls memory, storage, networks, and external programs," adding, "It is a kind of control tower, and as AI services grow more complex, the importance of the CPU is growing again."
The problem is that the control tower's performance falls short of the AI industry's expectations. The GPU resources stationed in data centers must be operated as efficiently as possible, but the CPU's control speed cannot keep pace with the countless work commands. The CPU has begun to be pointed to as a cause of the AI bottleneck. Lee Yong-hyuk, CTO of Klusch, an AI infrastructure platform company, noted, "To use an analogy, the AI bottleneck is a situation where GPUs sit idle because they cannot receive work," pointing out, "Delays in CPU data pre-processing, inefficient GPU task distribution, and the narrow bandwidth of the passage (interconnect) between the CPU and GPU correspond to the AI bottleneck."
Demand for high-performance CPUs suited to the AI era has triggered a new industry trend. Tech giants that had not previously developed CPUs have one after another thrown down the gauntlet, declaring they will build CPUs themselves. The news that drew the most attention was Nvidia's Vera announcement. CTO Lee assessed, "Nvidia's proprietary CPU brand 'Vera' is a product that has all the key CPU improvements the AI industry wants," adding, "Its technical ambition to fundamentally eliminate the AI bottleneck is evident, by combining it with Nvidia's next-generation GPU 'Rubin' through chip-to-chip connection."
Besides Nvidia, Qualcomm and Arm also plan to enter the CPU market. Qualcomm had previously developed CPUs for mobile devices and personal computers (PCs), but recently announced its data center server brand "Dragonfly" and is preparing to ship CPUs for data centers. Arm, which had made semiconductor design intellectual property (IP) licensing its main business model, also unveiled its proprietary CPU product "Arm AGI CPU" this year. This product was created for the purpose of working with Meta's AI training and inference accelerators.
Behind the ignition of Big Tech's CPU competition also lies concern over AI data center operating costs. The longer the AI bottleneck lasts, the more AI companies not only fall behind in AI development speed but also have to bear expensive data center electricity bills in the meantime. They end up failing to produce results while only draining costs. Accordingly, the common view in the industry is that for some time, general-purpose CPUs from Intel and AMD and Big Tech companies' proprietary CPUs will coexist in AI data centers.
CEO Yoo emphasized, "Electricity bills account for 30 to 40 percent of data center operating costs," adding, "Just as Nvidia and Amazon Web Services (AWS) adopted Arm-based designs when developing CPUs for their overwhelming energy efficiency, CPU competitiveness leads directly to cost competitiveness." CTO Park said, "Going forward, global Big Tech companies will focus on the performance and power efficiency of the entire system that combines CPU, GPU, memory, and network."
