A new trend of CPU shortages
The Pragmatic Engineer
CPU shortages are hitting cloud buyers now, not just GPU buyers. Spot deals are fading and even big companies are being told to wait.
Based on reporting by The Pragmatic Engineer — 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
Cloud teams are running into a new bottleneck: CPUs. A recent dinner with CTOs and infrastructure leaders surfaced the same complaint over and over — getting CPU capacity from cloud providers is getting harder, and the cheap spot deals people once relied on are disappearing.
That old bargain mattered. Spot pricing could cut CPU bills by as much as 90% because providers were offloading machines they couldn’t keep full. Now, according to the people in the room, it’s often impossible to get spot instances without long-standing provider relationships, and even reserved capacity has to be planned months ahead. In some cases, providers are simply saying no because they do not have enough CPUs, or not the right ones.
Simon Eskildsen, CEO of turbopuffer, said the problem is tied to AI workloads. Reinforcement learning uses a lot of CPUs, and so do agents that search, run software, compile code, test code, and lint it. He said even large companies are competing for allocations, and expects things to get worse before they get better. A VP of Engineering at a large inference provider told the same story from another angle: they have money to spend and want more capacity, but the cloud providers have nothing left to sell them.
The squeeze is structural. Katelyn Lesse, head of platform engineering for Claude Platform, pointed to factory bottlenecks at TSMC, SK Hynix, Samsung, and Micron. GPUs and CPUs are competing for production lines at TSMC, while DRAM manufacturing is giving way to high-bandwidth memory. Intel has its own fabrication plants, but it has been dealing with yield problems and shifting capacity toward server chips. The result is a tighter supply chain on both the compute and memory sides.
The knock-on effect is already visible in day-to-day planning. CPU orders are taking around six months to be fulfilled instead of one to two weeks, prices are up by 10% to 20%, and capacity planning now needs to happen up to 12 months in advance. AI agents are part of the reason: Uber’s agent requests were up ninefold over six months, and the cloud ratio for AI data centers has shifted from 1 CPU to 8 GPUs toward 1:4, with 1:1 not looking impossible. The old habit of autoscaling into CPU capacity on demand is starting to look optimistic.
My take — AI-written commentary, not fact-checked reporting
The industry spent years treating CPU as boring plumbing, and now it’s the thing everyone forgot to reserve. AI did not just eat the GPU buffet; it dragged the whole kitchen into the same room and started counting spoons. Companies that still think capacity planning is for other people are about to learn what a six-month wait feels like.
Read more about this at: The Pragmatic Engineer