Article

How to Choose a VPS Without Regretting It: A Homelab Buyer's Guide

How to Choose a VPS Without Regretting It: A Homelab Buyer's Guide

A $22-a-year box and a $200-a-year box can both advertise 2 vCPUs and 2 GB of RAM. One of them will compile your project while you make coffee; the other will stall opening a text editor at nine in the evening. The spec sheet is not exactly lying. It is just omitting everything that matters: how much CPU you are actually allowed to use, how crowded the network path is at peak hours, and what the same plan will cost when the first bill expires.

China's homelab community argues about this constantly. A recent LINUX DO discussion started with a buyer comparing three cheap US West Coast plans and asking which would stay fast in the evening; the replies turned into a general lesson: you are not buying hardware, you are buying a share of a machine, a route, and a billing relationship. This guide distills that decision process.

Start with the job, not the spec sheet

Every bad VPS purchase starts the same way: sorting plans by price, then picking the biggest numbers. Flip the order. Write down the one or two jobs the server will actually do for its first six months, because each job has a different bottleneck.

JobWhat actually mattersSensible floor
Personal blog or docs siteRAM for the web stack, stable network1 vCPU, 1 GB RAM, 20 GB disk
Code server or AI coding agent hostSustained CPU, RAM for the model context2 vCPU, 2–4 GB RAM
Media server or file syncDisk space, monthly transfer allowance2 vCPU, 4 GB RAM, 100 GB+ disk
Build box or CI runnerBurst CPU, fast disk2–4 vCPU, NVMe disk

Notice what is missing from the table: the headline port speed. A "1 Gbps port" means nothing if your plan's monthly transfer cap is 500 GB, or if the route to your users is congested every evening. The job decides which column to read first; the spec sheet comes second.

Four numbers that actually decide how a box feels

1. How much CPU you may really use. Budget VPS plans almost always share physical cores, and many providers cap sustained usage in their fair-use policy. BandwagonHost, for example, publishes per-plan CPU allowances in its knowledge base, with small plans limited to a fraction of one core on a one-hour rolling average. That is not a scam — it is how a $50-a-year plan stays profitable — but it means a plan that benchmarks brilliantly for five minutes can throttle under a sustained compile. If your job runs hot for hours, look for plans that explicitly promise dedicated cores or a service-level agreement.

2. RAM headroom, not RAM totals. A server running at 90 percent memory is one traffic spike away from the kernel's out-of-memory killer. Keep at least 30 percent headroom over what your stack uses at rest. A 1 GB box can run a static site and a small database; add a code server or an AI agent and you want 2 GB minimum, 4 GB to be comfortable.

3. Disk type and size. NVMe storage is now standard even on cheap plans, but size still bites: a 20 GB disk fills fast once you add Docker images, logs, and backups. Check whether snapshots count against your quota, and whether you can grow the disk later without migrating.

4. Port speed versus transfer allowance. Providers advertise the port ("1 Gbps", "2.5 Gbps") and bury the monthly transfer cap in the fine print. For a personal project, 1 TB a month is generous; for a media server it can be a week of streaming. Exceeding the cap usually means throttling or overage billing, not a friendly warning.

The route is the product

If your users — or you — sit in China, the network path matters more than the data center's city label. The same Los Angeles data center can feel fast or unusable depending on which carrier carries your packets home. Premium routes such as China Telecom's CN2 GIA, China Mobile's CMIN2, and China Unicom's premium transit avoid the congested public backbones, and they are the main reason a plan can cost several times more than a seemingly identical one next door. Forum members called a cheap plan without an optimized route the classic first-timer trap: fine at noon, painful at peak hours.

You cannot verify a route from a sales page alone. Before committing to a yearly plan, buy monthly first — a month of real evenings is worth more than any review. Test at the hours you will actually use the box, run a traceroute (mtr) from your own connection and check the carrier hops, not just the final ping, and compare candidates with the provider's looking-glass or test IP if one is offered.

The rule from the thread is blunt: if the route is wrong, no amount of CPU will save the experience; if the route is right, even a modest box feels fast.

The price traps everyone walks into

Cheap VPS shopping has three recurring traps, and all three showed up in the same discussion.

The renewal cliff. The headline price is often an introductory or first-term rate. Before you pay, find the renewal price in the cart or terms — it is frequently double, and it is the price you will actually live with. Forum users advised waiting for the Black Friday season rather than grabbing a mediocre deal now, since the best prices in this market are seasonal and recurring.

The yearly lock-in. Annual billing looks like a discount and acts like a bet: you are betting the provider, the route, and your own plans will not change for twelve months. For a first purchase from an unfamiliar provider, monthly billing is the cheap insurance. Switch to yearly only after a month of good evenings.

The expiring free tier. Several users in the thread mentioned starting on a cloud free tier and losing it when the trial ended — one had been running on AWS's free tier until it expired. Free tiers are a fine way to learn, but keep your setup portable: configuration in scripts or containers, not hand-tweaked server state. Our earlier guide to the free VPS hunt covers what the genuinely free options look like and where their limits bite.

A 30-minute test drive before you commit

Treat the first month as a trial with a checklist:

  1. Benchmark the box. Run YABS (Yet Another Bench Script), the community-standard script reporting CPU, disk, and network performance in one pass. A disk at a tenth of expected speed or a CPU from a decade ago is grounds for a refund.
  2. Load it like you mean it. Install your real stack, not a hello-world page. Run your heaviest job and watch sustained CPU and memory for an hour.
  3. Test the route at peak hours. Stream, sync, or SSH during the evening window and note latency and packet loss.
  4. Test support and the exit. File a trivial ticket and see how long a human takes. Check that backups or snapshots actually restore — a backup you have never restored is a rumor.
  5. Decide before the renewal date. Put the renewal date in your calendar the day you buy. The best time to switch providers is before you are charged, not after.

A sane shortlist, by budget

Free, if you can get it. Oracle's Always Free tier remains the standout: up to 4 Ampere A1 ARM cores with 24 GB of RAM and 200 GB of storage, permanently, plus two small AMD instances. The catch, widely reported, is capacity — popular regions regularly show "out of host capacity," so you may need patience or a less crowded region. It is the best first server because it costs nothing while you learn.

Budget, roughly $20–50 a year. This is where the forum thread lived: forum users compared a US West Coast plan reported at $21.99 a year against a mid-range option at about $5 a month. At this tier, expect shared CPU with fair-use caps and unoptimized routes unless the plan says otherwise — plenty for a blog or experiments, if you buy monthly first and test the evening route.

Premium route, roughly $50 a quarter and up. Plans with CN2 GIA or equivalent optimized routing, such as BandwagonHost's CN2 GIA-E line, trade higher prices and smaller transfer allowances for stable peak-hour performance. Buy this tier only if your use case actually needs the route; otherwise it is money spent on the wrong column.

Buy boring, then upgrade

The through-line is unglamorous: start small, pay monthly, test evenings, and upgrade when a real bottleneck appears. The buyers who regret their purchase are rarely the ones who bought too little — a $5-a-month box can be replaced in an afternoon. The regrets come from a year prepaid on the wrong route, or a plan whose renewal price nobody read. Pick the boring option, run your real workload for a month, and let the server earn the yearly commitment. Once it is running, a panel like 1Panel can take over day-to-day management, the part of self-hosting that eats your weekends.

Inspired by community discussion on LINUX DO, including the VPS comparison thread at https://linux.do/t/topic/3074546 and the budget VPS discussion at https://linux.do/t/topic/3074589. Provider details described here reflect the plans as publicly listed; always confirm current pricing and terms before paying.

Keep reading
One Menu Bar to Run Them All: The Switcher Tools Quietly Taking Over AI Coding
Read next →