Choosing between a 32, 64 and 128 vCPU account sounds like a technical decision, but for most people it is really a budget decision in disguise. Buy too small and your jobs will not run. Buy too big and you have paid for headroom you never touch.
Here is a practical way to think about it.
First: what is a vCPU quota?
Your vCPU quota is the ceiling on how much compute you can run at once in a given region — not how much you can run in total over time. It is a concurrency limit. If your quota is 32 vCPUs, you can run one 32-vCPU instance, or four 8-vCPU instances, but not both simultaneously.
This matters more than people expect, because the limit bites exactly when you scale out.
32 vCPU: the sensible default
Good for: web applications, APIs, staging environments, small-to-medium data processing, CI/CD runners, and most solo-developer or small-agency work.
Realistically, this covers the majority of projects. You can run a comfortable production stack, a handful of worker nodes, and still have room to experiment. If you are unsure and not doing heavy ML or rendering, start here.
64 vCPU: for when you scale out
Good for: larger container clusters, parallel data pipelines, mid-size model fine-tuning, batch processing, and teams running multiple environments at once.
The typical trigger for moving up to 64 is not a single big instance — it is running many things concurrently. If your CI pipeline, your staging environment and your production workload all compete for the same quota, 32 starts to feel tight fast.
128 vCPU: genuine heavy lifting
Good for: large-scale ML training, high-throughput analytics, video rendering farms, simulation workloads, and enterprise deployments with strict throughput targets.
This is a real cost commitment, and you should only buy into it if you know your workload profile. If you are guessing, you are probably not there yet.
A simple way to decide
- List everything that runs at the same time — production, staging, CI, background workers.
- Add up their peak vCPU usage, not their average.
- Add roughly 30% headroom for spikes and for the deployments where old and new instances briefly overlap.
- Round up to the next tier.
If that maths lands you under 32, take the 32 tier and stop worrying about it.
Do not forget the region
Quotas are per-region. A 64 vCPU quota in one region does not give you anything in another. If you plan to deploy across multiple regions, factor that in before you buy — it is one of the most common mistakes we see.
Still unsure?
Describe your workload to us and we will tell you honestly which tier fits, even when that means recommending the cheaper one. You can see all the available tiers here, or ask us directly.
Leave a Reply