Bandwidth
The volume of data a proxy can move, and the unit residential and mobile proxies are typically billed by.
Bandwidth is how much data a connection can move, and on a residential proxy it is also the invoice. Those pools are billed by the gigabyte because their supply is metered by real underlying connections, so traffic rather than addresses is the resource you are actually buying. That single fact changes how you should write a scraper.
The change is that page weight becomes a line item. A modern product page is commonly one to three megabytes, and the overwhelming majority of that is images, fonts, video, tracking scripts and stylesheets. The HTML carrying the data you actually want is frequently under a hundred kilobytes of it. If you fetch the page the way a browser does, you are paying residential rates to download photographs you immediately discard.
Which makes bandwidth discipline the highest-leverage cost control available, and it needs no negotiation with anyone. Request compression and honour it. Fetch the document without automatically pulling every referenced asset. If you are driving a real browser, block image, media and font requests at the network layer, since a headless browser fetches everything by default exactly as a real one does. Prefer a site's JSON endpoint over its rendered page when one exists, because it is usually a fraction of the size and easier to parse.
Retries deserve their own mention because they are the quiet multiplier. A failed request that already transferred most of a response still cost you that traffic, so an aggressive retry policy against a flaky target can double a bill without collecting anything extra. Failing fast is cheaper than failing repeatedly.
It is also worth separating bandwidth from latency, because they are independent and people routinely conflate them. Bandwidth is how much fits through per second; latency is how long a single round trip takes. A connection can have plenty of one and little of the other. Bulk collection is governed by bandwidth and concurrency, while anything interactive is felt as latency first, and buying more of the wrong one fixes nothing.
Finally, not every family is billed this way. Datacenter and ISP proxies are usually sold per address with generous or unmetered traffic, because the supply is a machine in a rack rather than somebody's metered home line. If a workload is genuinely bandwidth-heavy and the target does not screen by network type, moving it to a datacenter proxy is often a larger saving than any amount of request tuning.
How HProxy handles it
Because residential is billed per gigabyte, the cheapest advice we can give is often to use less of it rather than to buy more: fetch what you parse, and move genuinely bandwidth-heavy work that does not need residential trust onto datacenter addresses instead.
Frequently asked questions
How do I reduce proxy bandwidth costs?
Stop downloading what you do not parse. Enable compression, avoid pulling images, fonts, video and stylesheets unless you need them, and if you drive a headless browser block those request types explicitly, because it fetches everything by default. Prefer a JSON endpoint over a rendered page where one exists. And keep retries tight, since a retried request pays for the traffic it already moved.
Why is my bandwidth usage higher than the page size suggests?
Usually because you are measuring the HTML and paying for the whole page. Every image, font, script and stylesheet a browser pulls counts, and so do redirects, failed attempts and retries that transferred a partial response before giving up. TLS handshakes and headers add a little on top. Measure what actually crossed the wire rather than the size of the document you kept.
Is unlimited bandwidth real?
It exists, and it generally comes attached to a different constraint rather than to no constraint. Unmetered traffic tends to be sold per address, per port or with a concurrency ceiling, which suits datacenter and ISP supply where the cost is a machine rather than metered consumer traffic. Read what the limit actually is, because there is always one somewhere.
Bandwidth or latency, which matters for scraping?
Bandwidth and concurrency govern bulk collection: total volume moved in a window. Latency governs anything interactive or sequential, where each round trip must finish before the next begins. If your job is a long queue of independent fetches, more concurrency helps and lower latency barely registers; if each step depends on the last, the reverse is true.
Back to the full glossary.