Every budget VPS from about $3 to $8 per month is sold with a fair-use CPU clause. Almost nobody reads it, and almost everybody is surprised by it eventually — usually the week a cron job, a crawler, or a traffic spike pushes the server past a threshold nobody told them existed. The clause is not a scam. It is how a host can put a four-vCPU plan on a shared node and still keep the node responsive for everyone else. But “fair use” is vague by design, and vague clauses are enforced at the host’s discretion.
This guide covers how fair-use caps actually bite on low-end plans, how to detect throttling with tools already installed on most servers, and what to ask before you buy so the clause is a known quantity rather than a surprise.
What a Fair-Use CPU Clause Actually Does
A fair-use clause is enforcement of a sustained CPU budget, expressed either as a percentage of a core, a percentage of the assigned vCPUs, or a “reasonable use” statement with no number at all. Three enforcement mechanisms are common on KVM budget plans:
| Mechanism | How it behaves | Typical budget-plan limit | Visible in |
|---|---|---|---|
| cgroup CPU quota (cpu.max) | Hard ceiling; processes are throttled once quota is exhausted per period | 100-200% of one core on a 4-vCPU plan | /sys/fs/cgroup/…/cpu.max, cpu.stat |
| Host-level steal / scheduling share | Soft; you compete for physical cores and lose | No stated number | top %st, vmstat st column |
| Policy-based suspension | Human or automated review after sustained load | “Reasonable use”, often 24h sustained | Support ticket, AUP email |
The first mechanism is honest and measurable. The second is the one that hurts, because there is no number to plan against. The third is the one that gets servers suspended at 3 a.m.
How It Bites in Practice on a $3-$8 Plan
A $4 plan commonly advertises “4 vCPU” — four virtual cores sharing a physical core or two on a node with twenty other tenants, so sustained throughput lands between 0.8 and 1.5 physical cores’ worth. Fine for a web server or a small database; not fine for compiling or a CI runner.
The failure mode is not a clean “you have been throttled” message. It is a server that benchmarks at full speed for sixty seconds and then quietly halves its throughput, which looks like an application problem rather than a hosting limit.
- Burst then flatline. First 30-60 s at 100% CPU, then a sustained plateau at 50% or 25%. Classic cgroup quota.
- Steal-time creep. Your CPU sits at 60% but %st climbs past 10-20%, meaning the hypervisor is giving your time slices to other tenants.
- Evening slowdowns. Consistent degradation between 19:00 and 23:00 local to the datacentre indicates node contention, not a fair-use cap.
- Sudden suspension. Automated AUP enforcement for “sustained 100% CPU over 24 hours” — usually triggered by a runaway process you did not notice.
Detecting Throttling: Four Commands You Already Have
You do not need a monitoring stack to prove a fair-use cap. These four checks are enough, and all of them run on a stock Debian or Ubuntu image.
1. Read the cgroup quota directly
On a cgroup v2 system, cat /sys/fs/cgroup/cpu.max returns two numbers: quota and period. A result of 200000 100000 means 200 ms of CPU per 100 ms period, i.e. a hard cap of two cores regardless of how many vCPUs the plan advertises. If the plan says four vCPU and this says two, you have found the cap before you found the problem.
2. Count throttled periods
cat /sys/fs/cgroup/cpu.stat includes nr_throttled and throttled_usec. Run a load for five minutes, then read the counters. If nr_throttled is rising by thousands per minute, you are being capped hard and continuously. If it is flat, your limit is elsewhere — likely steal time.
3. Measure steal time under steady load
vmstat 5 12 prints a st column. Anything above roughly 5% sustained is contention; above 15% means the node is oversold enough to affect latency-sensitive work. Pair it with top, where the same figure appears as %st in the CPU summary line.
4. Run a steady-state stress test, not a burst
Single-shot benchmarks lie on budget VPS because they finish inside the burst window. Run a fixed load for at least ten minutes and plot throughput per minute. A flat line at the advertised core count is honest capacity. A line that starts high, drops within a minute and stays low is a fair-use cap doing its job. A line that wobbles by 30% minute to minute is a contested node.
A Five-Minute Throttle Test You Can Run Today
| Step | Command | What to record |
|---|---|---|
| Baseline quota | cat /sys/fs/cgroup/cpu.max | quota/period → implied core cap |
| Reset counters | note nr_throttled from cpu.stat | starting value |
| Apply load 10 min | stress-ng --cpu $(nproc) --timeout 600s | — |
| Read steal | vmstat 5 12 | mean st % |
| Read throttle delta | cpu.stat again | nr_throttled change |
Interpretation: throttle delta near zero and steal under 5% means the plan delivers close to its advertised CPU. Throttle delta in the thousands with low steal means a clean, honest cap — you know your real ceiling. High steal with no throttling means an oversold node: unpredictable rather than merely limited.
What to Ask Before You Buy
Four questions, all answerable by a competent host in a single ticket. The quality of the answer tells you as much as the content:
- Is there a cgroup CPU quota on this plan, and what is it in cores?
- What is the vCPU-to-physical-thread ratio on the node type I will be placed on?
- What triggers fair-use enforcement, and is it automated suspension or throttling?
- If I am throttled, does it show in my metrics or do I need to measure it myself?
A host that answers “yes, 200% hard cap, 2:1 threads, we never suspend for CPU, only throttle” is selling you a known quantity. That is worth more than a plan with two extra vCPUs and no answer at all. When you compare budget VPS plans by CPU policy and price, put the answered quota next to the advertised vCPU count — the gap is the real spec.
Matching Workloads to Fair-Use Budgets
- Static site / reverse proxy: bursty; a 100% cap is plenty. Fair use will never bite.
- Small database (< 20 GB): needs sustained single-core performance. A 200% cap on fast cores beats 400% on throttled ones.
- Container host: many light services; watch steal time rather than quota, since contention hurts tail latency.
- CI runner / builds: sustained 100% for minutes. Almost always violates budget fair use — use a dedicated or higher tier instead.
Bottom Line
Fair-use CPU policies on budget VPS plans are not hidden traps, but they are hidden numbers. The fix is mechanical: read cpu.max, watch nr_throttled, measure steal time, and run a ten-minute steady-state test instead of a thirty-second benchmark. Then ask the host to state the quota in writing before you pay. You will end up with either a plan that honestly delivers its spec or a plan you correctly declined — and either is a better outcome than discovering the cap during an outage.

