Intel: CPU value needs re-evaluation in the Agent era as AI compute ratio trends toward 1:1

Intel held the Intel Connection technology innovation and industry ecosystem conference in Suzhou, where the core judgment given was: the value of the CPU in AI systems needs to be revalued in the Agent era. Supporting this judgment is a ratio change—from one CPU paired with four GPUs (1:4) to one CPU paired with one GPU (1:1).

The attribution of this statement needs to be made clear: this is a judgment given by Intel, a vendor-side claim, not an industry conclusion verified by a third party. The report also points out that the 1:1 ratio corresponds to the agent inference deployment trend emphasized by Intel, and the specific ratio still depends on the task and system configuration. That is, even in Intel's own wording, 1:1 is not a hard-and-fast formula that applies everywhere.

The report also mentions that Intel's stock price surged 12% the day before the conference, and says this indirectly confirms the market's recognition of the CPU's importance in the Agent era. This stock price figure comes from the report's retelling; the report does not give the specific date, baseline, or reason, so it can only be viewed as the reporting party's statement and cannot be taken as independent verification of the above technical judgment.

To understand this ratio judgment, one must first look at how the Agent works. Zhang Ya-qin, foreign academician of the Chinese Academy of Engineering and founding dean of Tsinghua University's Institute for AI Industry Research, explained at the conference that an Agent can plan, act, trial-and-error, and iterate around a goal until the task is completed. This means the model's answer becomes the basis for subsequent actions, and the action results enter the next round of processing.

Chen Baoli, vice president of Intel's Data Center Group and general manager of Intel China, used the task of 'organizing news' to illustrate: the user only needs to set it to collect AI news from the previous day every morning at 8 o'clock and complete a summary, but behind the scenes the Agent needs to go online to find information, and the content brought back may be web pages, PDF reports, or videos. Information in various formats needs to be extracted and processed first before it can be handed to the large model for analysis. In Chen Baoli's view, much of this preprocessing work is suitable for the CPU to handle.

Chen Baoli also simplified the data center into three mutually cooperating parts: CPU servers or clusters responsible for the Agent's execution and work orchestration, GPU and other accelerated computing resources undertaking model computation, and storage systems saving the data involved in tasks, conversations, and newly generated documents. According to the report's wording, if bottlenecks occur in the processing and moving of data, expensive computing resources may sit idle, and the completion speed of the entire task will also be affected. This causal judgment is the reporting party's inference, not a measured conclusion.

Bringing the perspective down to the procurement level, the questions become more specific. Feedback Chen Baoli and his team obtained from exchanges with domestic software partners and internet vendors is that customers care both about how many agents can run and about the capability and reliability of each agent.

The report gives two scenarios: some businesses need to start a large number of sandboxes in a short time to provide runtime environments for agents, and the system must respond promptly while already-running tasks continue to be processed; when multiple Agents run simultaneously on the same set of equipment, as the number increases, whether each Agent can still get a response within the required time—this figure is what determines whether the service can actually be used.

Therefore, the tests and optimizations Chen Baoli introduced on site examine both quantity and the performance of individual agents, and the references for CPU evaluation include concurrency capability, response time, and support for real tasks. These test and optimization results are all on-site introductions from Intel's side, and the report does not provide publicly available quantitative figures.

Facing the above problems, Intel's approach is to put the CPU and accompanying software into actual tasks and optimize them together. Chen Baoli introduced that Intel hopes to use the built-in acceleration capabilities of Xeon processors to speed up data preprocessing-related work and help customers reduce waiting.

Another entry point is storage and caching. Chen Baoli introduced Intel's KV Cache optimization solutions, including using the processor's built-in capabilities for lossless compression and accelerating data transmission. He also mentioned that Intel is working with domestic software partners to build high-performance storage systems, with the goal of delivering data faster to the computing resources that need it and reducing CPU and GPU idling.

The actual effects and deployment scope of these solutions are not given with data in the report; the progress and landing time of the high-performance storage system cooperation with domestic software partners are also unclear. Therefore, they can currently only be understood as directions introduced by Intel's side, not as verified benefits.

A similar division of labor also exists on personal devices. Gao Song, vice president of Intel and general manager of Intel China for the Client Computing and Physical AI Group, introduced that the CPU handles the Agent's planning, decision-making, and other work requiring timely response, the GPU handles matrix-computation-type AI tasks, and the NPU supports long-running audio and video processing, computer vision, and other tasks with lower power consumption.

This division of labor corresponds to a very practical requirement: the Agent must respond promptly and may also need to continuously perceive the environment. Different tasks have different requirements for performance and power consumption, and software needs to allocate work to the appropriate computing unit. The device-side collaborative division of labor described by Gao Song is an introduction from Intel's side, and the report does not provide third-party verification.

Putting the above together, Intel's claim has a relatively complete logical chain: the Agent brings more execution, scheduling, and data processing work, these tasks are suitable for the CPU to handle, so the CPU's value in the system needs to be re-measured, and the ratio accordingly moves from 1:4 to 1:1.

But there are several points on this chain that need to be clearly marked. First, whether the ratio change from 1:4 to 1:1 is a general industry trend or a one-sided judgment by Intel, the report does not provide third-party verification. Second, the report does not explain how the CPU-to-GPU ratio in actual enterprise procurement will change with task types. Third, in a video address, Chen Liwu stated that as the cornerstone of modern computing, x86 continues to drive computing toward intelligent applications—this is a vendor's own statement, not an external conclusion.

Therefore, a more prudent reading is: Intel has proposed a judgment about the CPU's role in the Agent era and given its own product and software path; whether this judgment holds and to what extent depends on whether publicly available quantitative results can be produced later, and how customers configure CPU and GPU in actual deployments.