What a datacenter IP tells the target
Every public address belongs to an autonomous system, and that mapping is published. The fastest check any site can run is to look up the ASN behind a request: an address registered to a hosting provider is a server, one registered to a consumer ISP is a household. That single lookup, not the individual address, is what separates the two classes.
This is worth stating plainly because it sets the boundary of what datacenter capacity can do. Rotating between addresses inside the same hosting ASN does not defeat ASN-based classification β you have changed the address, not the class. Where a target scores address class, no amount of datacenter IP diversity substitutes for residential capacity.
When datacenter is the right answer
For a large share of real workloads it is, and buying residential instead is the most common overspend we see.
Datacenter proxies are the correct choice when the target is not aggressively fingerprinting address class: internal tools and staging environments, uptime and availability checks, documentation and reference sites, public APIs, and most price or catalogue sources that publish data openly. They are faster than residential because they sit on server-grade connections, and they are billed per IP rather than per gigabyte, so heavy predictable volume has a predictable cost.
They are the wrong choice when the target actively scores address class β retail, travel, ticketing, social and sneaker sites almost always do. If you are seeing blocks on a defended consumer site, moving up to residential is the fix; if you are paying for residential to fetch open documentation, moving down will halve your bill without changing your success rate.
Measure rather than assume: run a few hundred real requests on the cheaper option and compare. Choosing a proxy type walks through it.
Dedicated or shared
A dedicated address is yours alone, so its reputation is a product of your traffic and nobody else's. That matters whenever a target is likely to associate the address with an account or a pattern of use, and it is the reason dedicated capacity behaves more predictably over long-running jobs.
Shared addresses cost less and are perfectly reasonable for stateless, low-sensitivity work, but you inherit whatever else has been run through them. If a job matters and you cannot explain a sudden drop in success rate, shared IP reputation is the first thing to rule out.
Throughput, concurrency and what actually limits you
Datacenter capacity is fast enough that your own client is usually the bottleneck before the network is. Start lower than you think β around ten to twenty concurrent requests per proxy β then raise it in steps while watching two numbers: success rate and average response time. If success rate falls as you add threads, the target is rate limiting you rather than the proxy failing. If response time climbs steeply, you have saturated something and are queueing.
Reuse connections with keep-alive rather than opening a fresh TCP and TLS handshake per request, skip assets you never parse, and cache anything you fetch more than once. Most scrapers can halve their thread count without losing throughput.
The failures we see most
429, or success rate sliding as volume grows. Rate limiting. Slow down, add jitter so requests are not perfectly evenly spaced, and back off after errors rather than retrying immediately.
403 on a fresh address. Usually address-class detection rather than a bad IP, which means rotating harder will not help. Either the target needs residential capacity or your client fingerprint is contradicting itself β the five detection layers explains which is which.
A 200 containing a challenge page. The most costly failure, because unvalidated bodies mean you are recording blocks as successes. Check for a marker you expect on a real page before counting a fetch as good.
Our five-step test isolates a proxy problem from a target problem in a couple of minutes.
Acceptable use
A proxy changes the network path, not the rules. A robots.txt disallow is a refusal whether or not anyone enforces it, and using our capacity to circumvent an explicit refusal falls outside our terms, along with the targets set out in our acceptable use rules. What we log, and for how long, is documented in our logging policy.

