AI chip shortages such as GPUs are already widely known, but there is also talk these days that CPUs are in short supply. That could leave the software industry facing infrastructure bottlenecks it has not experienced before.

Caitlin Rese (케이틀린 레세), who leads Anthropic's platform, stressed that software development organisations need to focus on making the software stack itself more efficient as a way to respond to a CPU supply shortage.

She recently said on social media platform X (Twitter), "In the past, it was enough to plan ahead only for databases or GPUs. CPUs could be borrowed from the cloud right when needed, but now CPUs also need to be secured in advance. If you are a team running software at scale, you should factor time into your roadmap in advance to respond to the CPU supply shortage."

She said server order lead times have increased to six months from 1 to 2 weeks, and prices have risen 10 to 20 percent since March. She also expects supply shortages to persist for several quarters.

The cause is agents. Agents run repetitive loops to gather context, call tools and interpret results. The longer a session is maintained, the more CPU use rises. Running code or manipulating a browser also requires sandboxes or containers.

She said, "In an earnings announcement, Intel said the CPU-to-GPU ratio in AI data centres fell to 4-to-1 from 8-to-1, and it could go to 1-to-1 as agent workloads increase. AMD took a similar position."

The problem becomes clearer in the supply structure. Most logic chips are made by Taiwan's TSMC, followed by Intel and Samsung. Memory chips are handled by three companies: SK Hynix, Samsung and Micron. As AI demand surges, these companies cannot increase output without years of investment and tens of billions of dollars.

Rese said, "TSMC production lines are limited. GPUs, CPUs, Apple, Qualcomm and Broadcom all have to split volume here. The same is true for SK Hynix, Samsung and Micron. HBM and standard DRAM compete for the same wafers."

Looking only at CPU companies, AMD has no in-house fabs and depends on allocations from TSMC. Intel has fabs but is experiencing yield problems and is shifting some PC chip production lines to servers. DRAM prices are also rising as production tilts toward HBM.

Rese said it appears difficult to find a solution to the CPU shortage within the next few quarters. The situation could also be a major variable for software development teams. Her proposed response measures can be summarised into four main points.

First, teams should start by checking where CPUs can and cannot be used. She said, "Workloads that require isolation cannot share hardware with other tenants, reducing efficiency. Some workloads must run only in specific locations or regions, and some require specific machine types. You need to identify and optimise these constraints in advance."

Second, teams should reduce the time it takes to buy servers and make them usable in practice.

She said, "It does not end with securing servers. It takes anywhere from a few days to several months to set up a new cluster and put it into actual service. How efficiently you build this preparation process matters."

Third, teams should first check the utilisation of existing infrastructure. She said, "Average CPU utilisation in Kubernetes remains around 10 percent even across the industry as a whole. That is because many set aside a large buffer to match an expected peak and often do not revisit the estimate. There is significant room to find waste here."

Fourth, teams should raise the efficiency of the software stack itself. She said, "Work is needed to refine code and structure so that less CPU is used not only in the API service layer but also in the agent layer."