Technology
Agent infrastructure is making CPU capacity a planning problem again
Katelyn Lesse argues that software teams need to plan general-purpose compute capacity rather than assume cloud autoscaling will always absorb demand. Agent sandboxes, browsers, build systems, databases and orchestration all consume CPUs even when model inference runs on accelerators. Her essay links this demand to competition for chipmaking and memory capacity, then focuses on practical capacity constraints within a software fleet.
Editorial summary · Publisher article
What matters
- Count the CPU work around inference, including concurrent sandboxes and growing CI workloads.
- Regional, isolation and machine-type requirements can make nominal spare capacity unusable.
- Hardware delivery is only the start: cluster bring-up time and excessive reservations also constrain usable capacity.
Where the engineering work is
The essay suggests forecasting demand, improving cluster deployment, revisiting overprovisioned services and making both application and agent runtimes more efficient. These are capacity-planning questions, not an argument to abandon workload isolation. Its reported server lead times and price increases are the author's observations and cited industry accounts, not a representative market survey.
Context & caveats
- This is practitioner analysis. Supply, price and utilization estimates were not independently measured by this newspaper.
Why this made the edition
Full supplied article reviewed, with source limitations and attribution retained. Checked against prior editions for a genuinely new development or distinct substantive analysis.
Source material