How We Model Server Memory & Pricing
Most developer cost calculators overlook operating system baselines and transient compilation spikes, leading directly to Out-Of-Memory (OOM) crashes in production. Here is our mathematical modeling criteria.
1. 750 MiB Ubuntu Linux Baseline
A standard Ubuntu 24.04 LTS server running systemd, journald, SSH, and the Docker engine allocates ~750 MiB of physical RAM before any user containers or model weights load. Sizing formulas that ignore this OS overhead result in immediate OOM failures on 1 GiB and 2 GiB instances.
2. Steady Runtime vs. Build Spikes
Framework compilers (like next build) consume transient memory spikes up to ~1.42 GiB (1,450 MiB) above runtime. We separate sustained runtime from compilation spikes, allowing builders to choose between provisioning physical RAM, offloading builds to external CI/CD, or absorbing temporary build spikes with persistent swap.
3. No Swap for LLM Inferences
Linux swapfiles are strictly forbidden for LLM inference (e.g. Ollama Llama models) and active database caches. Swapping model weights or transactional database pages onto disk causes catastrophic token latency stalls and query timeouts. We trigger hard warnings whenever sustained runtime memory exceeds physical RAM.
4. Shared vCPU Transparency
Budget cloud VPS providers use shared vCPUs. While cost-effective for background queues, automation webhooks, and web servers, sustained CPU-heavy LLM token generation may trigger hypervisor throttling. We document dedicated-performance alternatives (like Vultr High Frequency) for CPU-bound pipelines.
Tier Selection & Verdict Rules
Steady memory consumes ≤ 70% of physical RAM, and peak memory consumes ≤ 90% of physical RAM. Provides comfortable headroom for traffic bursts.
Steady memory consumes ≤ 85% of physical RAM, and peak memory consumes ≤ 100% of physical RAM. Operates safely within limits, but offers less burst buffer.
Steady memory exceeds 85% or peak memory exceeds 100% of physical RAM. Disqualified to avoid Linux OOM-killer process terminations.
Independent Pricing Verification
All cloud provider specs and monthly pricing (Contabo, Vultr, DigitalOcean, Railway) are checked directly against official provider storefronts as of November 2026. Outbound links are direct links to official pricing pages. Sizing algorithms and tier selections are derived entirely from modeled memory calculations.