GPU Fades, CPU Rises: 'Agentic AI' Sparks the Second Semiconductor Crisis

Increasing CPU Workloads in the Agentic AI Era

Advanced Foundry and Backend Processing Already Fully Booked

Supply Shortages Drive Benefits Across the FC-BGA Value Chain

There are forecasts that the main driver of this year's semiconductor supercycle will shift from graphics processing units (GPU) to central processing units (CPU). If the race to secure GPUs marked the first cycle, system optimization and the efficiency of CPU deployment are now being analyzed as the key to the second cycle. As artificial intelligence (AI) evolves from simply providing answers to becoming "agentic AI"—capable of planning and executing tools on its own—the CPU is being reevaluated as a core resource that determines data center performance.

Agentic AI Era: The Return of the CPU as 'Conductor'

On April 27, market research firm Growth Research stated in a report that "the bottleneck in AI infrastructure is shifting from the GPU to the CPU." The firm predicted that the future landscape of the AI semiconductor market will not be determined by the number of GPUs, but rather by how efficiently the CPU orchestrates overall operations.

So far, GPUs have dominated the market by handling the core computations for AI training and inference. In traditional chatbot-type AI architectures, the CPU was limited to an auxiliary role—organizing user input and passing it to the GPU. However, with the advent of agentic AI, this situation has been reversed. Agentic AI does not stop at one-off answers; it repeatedly performs web searches, database queries, API calls, and code execution, and if the results are insufficient, it automatically retries the process. The CPU has thus taken on the role of 'orchestrator,' managing and coordinating these complex workflows.

According to a study published by the Georgia Institute of Technology in November of last year, the CPU processing stage accounts for 50–90% of total latency in agentic AI workflows. No matter how fast the GPU’s computations are, if the CPU’s task allocation speed cannot keep up, the overall system performance declines.

Consequently, the composition of CPUs and GPUs in a single AI server is also changing. Whereas the CPU-to-GPU ratio used to be around 1:4 or 1:8, in recent agentic AI servers, this ratio has narrowed to between 1:1 and 1:2. This marks a structural shift in which, based on the same number of GPUs, the required number of CPUs can surge up to eight times.

Supply Chain Bottlenecks: "Six-Month Wait Even If You Have Money"

Despite this explosive demand, supply is falling short. High-performance server CPUs rely on advanced processes such as TSMC's 2–3 nanometer (nm; one billionth of a meter) nodes, which are already crowded with global big tech companies seeking to produce AI GPUs and mobile application processors (APs). As of the first quarter of 2026, the lead time (from order to delivery) for new orders on TSMC's 3nm process has reached 52 to 78 weeks.

Another variable is the priority allocation by foundries (semiconductor contract manufacturers). Foundries like TSMC are prioritizing production lines for high-margin AI GPUs and mobile APs, leaving server CPUs relatively lower in priority. In addition, advanced packaging facilities (such as CoWoS and SoIC), which are essential for connecting multiple chiplets, are also focused on GPU production, further exacerbating the CPU supply shortage.

The shortage is already impacting both prices and delivery times. Whereas CPU lead times used to be sufficient at 1–2 weeks, they have now increased to 8–12 weeks, with some products requiring a wait of more than six months. As a result, major design firms such as Intel and AMD are raising product prices by about 10–15%. Intel’s surprise earnings report for the first quarter of 2026 (sales of $13.6 billion) reflects this recovery in CPU demand and rising unit prices.

Yonghee Han, a researcher at Growth Research, explained, "Because CPUs are designed to fit the process technology of a specific foundry, it is structurally impossible to immediately transfer a CPU made at TSMC to Samsung Electronics or Intel Foundry. Changing the process requires redesigning, which alone takes more than two years and costs over $500 million."

The Domestic Semiconductor Value Chain: FC-BGA and Backend Processing Poised to Benefit

The Korean semiconductor industry is expected to find opportunities not in direct CPU design, but in the surrounding value chain affected by the shortage. In particular, the FC-BGA (flip-chip ball grid array) substrate sector—which is essential for high-performance CPUs—is likely to benefit first.

Researcher Han noted, "Server and AI-use FC-BGA substrates are more than three to four times larger in area and have more than twice as many layers compared to those for general PCs, making the technical barriers much higher. Therefore, the few companies capable of reliably mass-producing FC-BGA, such as Samsung Electro-Mechanics and Daeduck Electronics, are poised to see concentrated benefits."

In addition, advanced packaging materials, memory, and OSAT (outsourced semiconductor assembly and testing) sectors are also expected to benefit in succession. As the CPU supply shortage is resolved, advanced process and packaging bottlenecks are likely to stand out first, and, when finished products begin shipping in earnest, test and backend processing volumes are expected to rise in tandem.

This content was produced with the assistance of AI translation services.