Agents warrant a separate assessment of the CPU execution layer; demand still depends on task workloads

The higher forecast reflects a change in the vendor's demand assessment

In an article published on 2026-05-07, AMD forecast that the server CPU market would exceed $120 billion in 2030. By its official investor roundtable on 2026-07-23, the forecast under discussion had reached $220 billion. This change deserves attention: AMD raised its expectations for the size of the server CPU market and included agent execution in its explanation of demand. But this remains a vendor's forward-looking estimate, based in part on customer interviews and observations of internal workflows. It cannot establish that actual purchases have already reached the corresponding scale. The higher forecast offers a demand hypothesis to test; deployment and usage evidence is still needed.

Evidence: [1] [2]

Execution tasks give CPUs a separate role to assess

The demand hypothesis has a concrete technical basis. AMD identifies orchestration, tool calls, enterprise applications, permissions, security and state handling as CPU work. NVIDIA's technical article also discusses sandbox execution, data retrieval and orchestration. These stages show that assessing agent resources requires more than counting the accelerators needed for model inference. Anthropic's Claude Tag provides a product example: users can delegate tasks in Slack that use authorized tools, Claude works asynchronously, and the product supports scheduling future tasks. However, a product's ability to execute work in the background is not enough to prove that users will keep using it, and the scale of CPU workloads cannot be inferred directly from that capability.

Evidence: [1] [4] [6]

A separate execution layer is an architectural choice

The WeChat public account argues that agents will create an “Agentic CPU” layer that can scale independently. AMD's materials support a narrower judgment within that claim: a distinction can be made between the host CPU that supplies data to accelerators and the CPU compute layer that executes tool and application tasks. The former emphasizes timely data supply, memory bandwidth and coordination across the system; the latter handles execution tasks proposed by the model. This suggests that a separate CPU layer may have value when execution work consumes substantial resources and can be provisioned separately from inference nodes. Here, “Agentic CPU” is primarily a division of work. The available evidence is not yet sufficient to treat it as a new hardware category with uniform specifications and established demand.

Evidence: [1] [3] [7]

Core counts cannot stand in for task completion capability

The key to determining how much of this resource layer is needed is how tasks consume it. NVIDIA discusses code with extensive branching and memory-sensitive retrieval. AMD also explicitly states that actual capacity and throughput depend on workloads, models, memory, software, orchestration and system configuration. Core or thread counts therefore cannot be directly converted into the number of agent tasks that can be completed. My judgment is that assessment should focus on task completion latency, throughput and resource use: execution work that can run in parallel may benefit from more cores, sequential steps may depend more on single-core speed, and memory access may also constrain overall performance. Determining which configuration works requires measurements of specific tasks; it cannot be inferred directly from the number of agents.

Evidence: [4] [3] [1]

Scheduling merits research; funding has yet to prove its effects

Gimlet Labs' company press release confirms that Andreessen Horowitz led its $300 million Series B funding round. The company's stated approach includes matching GPUs, specialized accelerators and CPUs to inference stages, along with orchestration across nodes. This provides a concrete example of heterogeneous scheduling receiving financial backing. A further inference is possible: if different stages suit different hardware, scheduling efficiency will affect the entire system's ability to complete tasks. But funding itself cannot prove that customers have already achieved cost reductions, nor can it prove that the benefits of coordination across nodes are sufficient to cover the costs of data transfer and waiting.

Evidence: [5]

The mechanism is supported; demand and efficiency still need evidence

The available technical materials are sufficient to support one judgment: retrieval, orchestration and execution tasks in agents make CPU resources worth assessing separately. Stronger market conclusions still depend on unresolved premises—whether these tasks are widely used, how much CPU resource they actually consume, and how much demand remains after software and configuration optimization. If subsequent deployment measurements show that CPU execution persistently limits task completion, the case for independent scaling will strengthen. If the main constraints lie elsewhere, or optimization substantially reduces execution resource consumption, inferred CPU demand will need to be revised downward. AMD's market forecast should be tested against this evidence.

Evidence: [1] [2] [3] [4]

What to watch next

• Whether real deployments disclose CPU execution time, memory wait time and task completion latency in enough detail to determine whether the CPU is a primary constraint.
• Whether a separate CPU layer shows consistent throughput gains or cost reductions when completing the same tasks as other configurations.
• Whether AMD's subsequent market forecasts gain support from actual deployment and purchasing evidence.
• Whether Gimlet Labs publishes comparable customer task results that show the benefits of heterogeneous scheduling and the costs of coordination across nodes.