Every week someone asks for a server with "enough bandwidth" without having worked out what enough means. The honest answer is that most websites never come close to saturating the cheapest port on the market, and the handful of workloads that do saturate it are obvious from the first line of the brief.
So the real question is not how much bandwidth does a server need in the abstract. It is which of two ceilings — the monthly transfer allowance, or the physical port speed — your workload runs into first. They are different limits, they fail in different ways, and only one of them shows up on an invoice.
Everything below is either measured from a named source or computed from inputs you can change.
Start with the two numbers you already have
Monthly transfer is pageviews multiplied by page weight. That is the entire formula.
For page weight, the Web Almanac 2024 page weight chapter measured a median of 2,652 KB on desktop and 2,311 KB on mobile in October 2024. Use your own number if you have one — open the browser network panel, hard-reload, read the transferred total — because a median is a starting point, not your site.
Take a store doing 100,000 pageviews a month at the mobile median:
100,000 × 2,311 KB ≈ 231 GB a month.
Spread evenly across a 30-day month that is 231 GB × 8 ÷ 2,592,000 seconds ≈ 0.71 Mbps. On a 1 Gbps port that is seven hundredths of one percent of the line.
That result is the point of this post. If you run a website — even a busy one — bandwidth is almost certainly not your constraint, and a provider selling you a bigger port to fix a slow site is selling you the wrong thing. Go and look at CPU, database queries and time-to-first-byte instead.
Six workloads, with the arithmetic shown
Average throughput is the monthly total divided by the month. Peak is what the port has to survive while the work is actually happening, and peak is the number that decides which port you buy.
| Workload | Monthly transfer | Average over the month | Peak sustained | Port that fits |
|---|---|---|---|---|
| Brochure site, 50k pageviews | ~116 GB | 0.36 Mbps | ~4 Mbps | 1 Gbps, unnoticed |
| Store, 100k pageviews | ~231 GB | 0.71 Mbps | ~7 Mbps | 1 Gbps, unnoticed |
| SaaS API, 2M calls/day at 20 KB | ~1.2 TB | 3.7 Mbps | ~15 Mbps | 1 Gbps, unnoticed |
| Video origin, 200 concurrent 1080p streams, 6h/day | ~97 TB | 300 Mbps | 1.2 Gbps | 10 Gbps |
| Nightly 2 TB backup in a 4-hour window | ~60 TB | 185 Mbps | 1.1 Gbps | 10 Gbps |
| Game patch, 50,000 downloads × 8 GB over 48h | ~400 TB | — | 18.5 Gbps | 20 Gbps or more |
The 1080p figure comes from a 6 Mbps stream, which is a reasonable mid-range bitrate; 200 × 6 Mbps = 1.2 Gbps. The patch row is 400,000 GB × 8 ÷ 172,800 seconds = 18.5 Gbps. Change the inputs and the answers move, but the method does not.
The backup row is the one that catches people
Look again at the nightly backup. Two terabytes is not a large backup, and four hours is a generous window. It still needs 1.1 Gbps sustained, which a 1 Gbps port physically cannot deliver.
What happens is not an error. The backup simply takes longer: 2 TB is 16,000 gigabits, so at 1 Gbps line rate it needs 16,000 seconds — 4 hours 27 minutes at a theoretical maximum you will never quite hit. On a 10 Gbps port the same copy takes about 27 minutes.
Your backup window is set by your port, and almost nobody works that out before buying. If nightly jobs are overrunning into the morning, the fix is usually the port, not the disks.
Port speed and transfer quota are two different ceilings
| Transfer quota | Port speed | |
|---|---|---|
| What it limits | Total gigabytes in a billing month | Bits per second at any instant |
| How it fails | An overage charge, or a throttle | Queuing, latency, slow transfers |
| Who notices | Your finance team | Your users, immediately |
| Fixed by | Buying more quota | Buying a faster port |
A 5 TB monthly quota sounds generous until a patch day moves 400 TB. A 100 Mbps port sounds fine until a backup window closes. Work out both numbers; they rarely point at the same upgrade.
On a 1 Gbps port running flat out for 30 days the physical ceiling is 324 TB — 1 Gbps × 2,592,000 seconds ÷ 8. That is a useful sanity check: if a provider offers you an unmetered 1 Gbps port, they are offering a ceiling of 324 TB, and if your plan needs more than that the port is what you have to change.
What eats bandwidth that you did not plan for
- Crawlers and scrapers. On small sites automated traffic routinely rivals human traffic. Check your logs before you check your theory.
- Cache misses on a CDN. A CDN cuts egress from the origin only for what it caches. Personalised HTML, checkout and API responses usually come straight from you.
- Replication and off-site copies. A second copy to another datacenter doubles the egress you were already paying for. Sizing that is its own job — see Offsite Backup Storage: Sizing and Costing the Third Copy.
- Attack traffic. A volumetric flood fills the port regardless of whether the requests are real. Network-edge filtering exists so the flood is dropped upstream rather than counted against you.
- Your own deploys. Container images and artefacts pulled on every release are real gigabytes.
Stop estimating and measure
Estimation is for the pre-purchase conversation. Once a server exists, measure it:
- Monitor network bandwidth and traffic on a server to get the shape of a normal week, including the peaks an average hides.
- Test network throughput with iPerf to check the port actually delivers what the invoice says it does.
Then size for the peak plus headroom, not the average. An average is a number that has already thrown away the only hour that mattered.
What this means when you go shopping
If the arithmetic put you in the first three rows of that table, buy on CPU, memory and disk, and treat bandwidth as solved. Every NexonHost dedicated server ships with an unmetered 1 Gbps port as standard — eight Intel Xeon and AMD EPYC configurations from €129 a month as of September 2026 — and a normal website will never find the edge of it. A Linux VPS is enough for most of them.
If you landed in the last three rows, you have a genuine bandwidth requirement and two things to decide. First, whether you want a transfer quota at all: unmetered dedicated servers remove the monthly allowance entirely, so the port speed becomes the only ceiling. The trade-offs are worked through in Unmetered Bandwidth vs Metered. Second, how fast that port needs to be: a 10 Gbps dedicated server moves 3.24 PB a month at line rate, with ceilings up to 200 Gbps on the larger EPYC builds.
Pricing and configurations here are as of September 2026 — check the product pages for current figures.
Sources
- Web Almanac 2024 — Page Weight, HTTP Archive. Median page weight measured October 2024. Retrieved 21 September 2026.
- NexonHost dedicated server configurations and pricing. Retrieved 21 September 2026.
- NexonHost 10 Gbps dedicated servers — port ceilings by build. Retrieved 21 September 2026.
