What the CPU shortage means for software teams
Katelyn Lesse
Software teams are hitting a CPU shortage as server lead times stretch to about 6 months. AI agents are driving it, and even plain old apps now need more compute.
Based on reporting by Katelyn Lesse — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
If you build and run software at scale, the new bottleneck may not be GPUs. It may be CPUs. Server quotes are now landing around six months out, not a week or two, and prices are up roughly 10% to 20% since March. Intel’s CFO even said demand had blindsided the company. For teams used to treating general-purpose compute as elastic cloud plumbing, that’s a rude little wake-up call.
The reason is broader than AI hype, but agents are a big part of it. Every service still runs on CPUs: APIs, web servers, queues, workers, workflow engines, cron jobs, databases, CI. Agents add more on top of that. The model call may happen on an accelerator, but the rest of the agent is a pile of CPU work: collecting context, looping, calling tools, parsing replies, keeping sessions alive, running code in sandboxes, opening browsers. Once those sessions pile up, the sandboxes themselves become a serious drain.
Intel said on its Q1 earnings call that AI-heavy datacenters have shifted from about one CPU for every eight GPUs to one for every four, with agentic workloads potentially pushing that to one-for-one. AMD has said something similar. That matters because more accelerators do not mean fewer CPUs. It means more of both, and the rest of the software world is still competing for the same chips.
The supply side is cramped in several places at once. Leading-edge fabs cost more than $20 billion and take three to five years to build. TSMC makes most of the world’s logic chips, while memory comes from SK Hynix, Samsung, and Micron. GPUs are fighting CPUs for TSMC capacity. HBM is fighting regular DRAM for memory wafers. Intel has its own fabs, but it is working through yield problems and shifting some capacity from PC chips to server chips. CPUs also need DRAM, and that memory has gotten more expensive because so much production is moving toward HBM.
For software teams, the practical shift is simple: CPU capacity now needs planning. Not just databases, and not just accelerators. Capacity constraints, regional placement, machine types, and the slow bring-up of new clusters all start to matter more when the industry’s average Kubernetes CPU utilization is supposedly around 10%. That is a lot of expensive caution baked into fleets, and a lot of room for waste to get exposed.
My take — AI-written commentary, not fact-checked reporting
This is what happens when everyone decides “just autoscale it” was a business strategy. The cloud-native era trained teams to ignore CPU until the bill arrived; now the bill is showing up before the machine does. The awkward truth is that AI didn’t just eat compute — it taught normal software teams how wasteful they’d been all along.
Read more about this at: Katelyn Lesse