Proxy AI reshapes computing power requirements! Citi: The 2030 CPU market is worth 300 billion US dollars, and I am optimistic that AMD (AMD.US) will reach 800 US dollars

The Zhitong Finance App learned that Citibank's latest research report indicates that with the successive launches of personal AI agents such as Meta Muse, proxy AI is becoming an order of magnitude driver of computing power demand. Citi raised the potential market size (TAM) of the 2030 CPU to 300 billion US dollars and raised the target price of AMD (AMD.US) from $575 to $800, maintaining a “buy” rating. Citi believes AMD is the main beneficiary of the explosion in CPU demand because Meta is one of the largest customers in the AMD server business. Meanwhile, Citi maintained Nvidia's (NVDA.US) “buy” rating and a target price of $315 due to increased demand for GPUs.

Proxy AI reshapes computing power requirements, and the CPU market size is 300 billion US dollars

Citi said in the report that recently, various personal AI agents such as Muse, Dots, Instinct, GrokBot, and GeminisPark have been released one after another. Unlike chatbots that are largely idle between user interactions, “always-on” AI agents such as Muse can run continuously, driving significant increases in computational, memory, and network consumption related to inference, covering enterprise and consumer workloads.

Citi has updated its CPU model and believes that compared to traditional chatbots, proxy AI is a potential order of magnitude driver for computing power requirements. Citi raised CPU TAM from $29 billion in 2025 to $300 billion in 2030, corresponding to a 60% compound annual growth rate. Citi still expects AMD to be the main beneficiary of the CPU demand explosion, with Intel being the secondary beneficiary.

Why is the CPU a new bottleneck?

In proxy AI, the CPU has become the new bottleneck. The ratio of CPU to GPU is going from 1:8 during model training, to 1:4 during inference, to 1:1 or higher in proxy AI. Citi believes that as the scale of proxy AI expands, this ratio will increasingly lean towards CPUs.

Traditionally, CPUs are mainly limited to two major applications: traditional workflows, and as head nodes for AI applications. In the head node, the CPU is only responsible for “management” — sending user requests to the GPU and then returning output to the user. GPUs, on the other hand, take on heavy tasks, such as running matrix multiplication to train large language models, or doing inference when querying chatbots.

The third new application is proxy AI. Citi defines a proxy CPU as a CPU that processes the orchestration, inference loops, data processing, and security components of autonomous AI agents. This new application has exploded over the past year as AI moved from chatbot reasoning to agent-based autonomous workflows.

GPU demand estimation for consumer AI agents

Meta's Muse is the first consumer AI agent launched on a real social networking scale. Its computing power requirements are one of the major unknowns in AI infrastructure.

Under Citi's benchmark, each user requires approximately 0.0020 GB200-class GPUs. Based on 100 million daily active users, this corresponds to about 200,000 to 390,000 Blackwell-class GPUs, about 2,800 to 5,500 NVL72 racks, and about $7 billion to $19 billion in one-time Nvidia revenue (including horizontally expanded networks). In terms of each user, every 1 million new daily users of the AI agent will bring Nvidia about 70 million to 190 million US dollars in revenue if a new server is used.

Memory and network: CPU drives DRAM and network requirements

In terms of memory, the CPU has a high adhesion rate to LP, DDR, and SSD. Micron expects server units to achieve high double-digit growth in CY26/27, and believes that another type of logic chip will also benefit from AI trends. Micron also recently stated that CPUs are putting more strain on DRAM, and Meta Muse is an example of consumers getting value from proxy workflows.

On the network side, proxy AI significantly increases network traffic as AI agents increasingly access infrastructure, which creates new acceleration opportunities for DPU and Spectrum-X Ethernet.