What does Meta Muse mean for GPU and CPU?
With the launch of personal AI agents such as Meta Muse, AI is evolving from passive, responsive chatbots to autonomous agents running around the clock, which may drive up the consumption of CPU, GPU, memory, and network infrastructure resources.
According to the latest research report released by Citi on October 6th, AI agents represented by Meta Muse are pushing market expectations for computing hardware to new heights. The bank predicts that by 2030, the total addressable market (TAM) for global CPUs will expand from $29 billion in 2025 to $300 billion, with a compound annual growth rate as high as 60%.
On the GPU side, the rise of agents has also brought about a significant increase in structural demand. Citi estimates that if Meta Muse reaches 100 million daily active users, approximately 200,000 to 390,000 Blackwell-grade GPUs will be needed, potentially bringing Nvidia a one-time revenue of $7 to $19 billion.
This trend directly benefits core computing power suppliers. The report points out that since Meta is one of AMD's largest server business customers, AMD will be a major beneficiary of the CPU renaissance and its price target has been raised to $800. Meanwhile, Nvidia will also continue to benefit from surging GPU demand.
The Rise of Agents: CPU Becomes the New Bottleneck
Before the explosive growth of agent AIs, CPUs were mainly limited to traditional workloads and acted as "Head Nodes" for AI applications. In head nodes, CPUs were only responsible for “management,” sending user requests to GPUs and returning results, while heavy matrix multiplication and inference tasks were handled by GPUs.
However, agents have changed this division of labor. Citi notes in its report, “Compared to traditional chatbots, we believe agent AIs are a potential magnitude-level driver of compute demand.” Agents need to handle orchestration, inference loops, data processing, and security components, making the CPU the new bottleneck.
As AI shifts from model training to inference, and then to autonomous agent workflows, the ratio of CPUs to GPUs is changing significantly. In the model training phase, the ratio is 1:8; at the inference stage, it’s 1:4; but with agent AIs, the ratio is evolving to 1:1 or even higher. The bank predicts that by 2030, CPUs dedicated to agents will grow at a compound annual growth rate of 247% and account for 52% of the entire CPU market.
Meta Muse’s GPU Compute Ledger
Meta Muse is the first consumer AI agent released at a true social network scale, making its computation footprint one of the most critical variables in today's AI infrastructure. Unlike standard chatbots, Muse functions as an autonomous agent that runs continuously, with its underlying logic resulting in far greater GPU consumption.
Citi has created a bottom-up GPU demand model. In the base case, assuming a typical user makes 4 simple requests and 12 agent tasks daily (such as finding a restaurant, checking a calendar, or drafting an invitation), each agent task entails about 8 model invocations. Because the model does not retain memory between invocations, it must reread the ever-increasing conversation context each time.
Based on this, the bank estimates that each Muse user would require roughly 0.0020 units of a GB200-grade GPU. In simple terms, one GPU can serve about 500 users. “Structurally, agent AIs have a much heavier inference workload than chatbots,” the report emphasizes. Under the same framework, the GPU capacity required for chatbot users is only 12.5% to 25% of what’s required for agent users.
Cascading Effects on Memory and Network
The agent workflow not only sets higher requirements for logic chips, but also triggers cascading effects in memory and networking. On the memory side, CPUs have a high additional ratio for LP, DDR, and SSD. Micron recently pointed out that agent workflows, as exemplified by Meta Muse, are enabling consumers to extract more value, and CPUs are placing increasing constraints on DRAM.
On the network side, agent AIs have markedly increased network traffic. Nvidia’s management told Citi, “Agents can run for hours or even uninterruptedly… thus driving much more token generation for consumer and enterprise workloads.”
This continuous operation means that more and more infrastructure access comes from agents rather than humans. This is driving the growth of “north-south” data center traffic (e.g., storage access, security, configuration, and other services), creating new acceleration opportunities for DPUs (such as BlueField 4) and Spectrum-X Ethernet.
