Datacenter, ISP, Residential or Mobile: Picking a Proxy Type by Target

Webshare will sell you a dedicated datacenter address for about three cents a month. Bright Data's discounted residential rate is $4 a gigabyte. Both numbers came off the vendors' own pricing pages on 27 July 2026, and the distance between them is the largest cost decision in most scraping budgets. Almost every guide to that decision is organised by product: one page per proxy type, each describing the type in isolation, each ending in a list of vendors. Engineering teams do not start from a type. They start with a target list and a block rate, and they want the cheapest thing that clears it.

This guide runs the other way, from four classes of target back to the type that handles each one, with a ladder to climb when the cheap option stops working. The proxy networks category is the largest on this directory, and every price below was read off the vendor's own pricing page on 27 July 2026. Proxy pricing moves constantly, so treat each figure as an observation with a date on it rather than a fact.

The four types, and what you are actually buying

The billing column matters as much as the detection column, and it is the part vendor comparison pages skip. Datacenter and ISP are sold per address with traffic included. Residential and mobile are sold per gigabyte. The cost curve therefore inverts depending on how heavy your pages are. A crawl pulling small JSON payloads across many addresses is cheap on datacenter and expensive on residential. A crawl pulling a few dozen image-heavy rendered pages can be the reverse. The proxy pricing guide covers how to normalise quotes across those units before you compare them.

Target class one: open catalogue and static content

Sitemaps, RSS feeds, open JSON endpoints, documentation sites, government and academic portals, and the long tail of catalogues running no bot defence at all. This is a much larger share of the web than the proxy industry's marketing suggests, because the industry has no reason to mention it.

Datacenter IPs are correct here. Nothing on these targets is trying to identify you, so the addresses only need to spread load widely enough to stay under a naive per-address rate limit.

Webshare is the easiest place to establish that, mostly because you can test before paying. Its pricing page on 27 July 2026 offered 10 proxies free with no card, then dedicated datacenter addresses from $0.0299 per proxy per month at 100 proxies, falling to $0.0179 per proxy at 60,000 or more. That is roughly three cents an address. There is no per-gigabyte product in this market within two orders of magnitude of it.

Rayobyte is the other self-serve pick and publishes the whole ladder without a sales call. On the same date its pricing page listed static datacenter from $1/IP, rotating datacenter at $0.30/GB on the 1 to 50 GB starter tier, static ISP from $5/IP, rotating ISP at $3.75/GB, and rotating residential from $3.50/GB. Having the rungs priced side by side on one page is genuinely useful when you are trying to work out what escalation will cost before you commit to it.

Who this is wrong for: anyone whose target list is mostly marketplaces, social platforms or sites behind a CDN with bot management switched on. Buying several thousand datacenter addresses and discovering they are rejected the moment they touch a defended site is a common and expensive mistake. It is also wrong for teams with no per-target success measurement, because the entire argument for datacenter rests on knowing which of your targets are actually undefended. If you cannot name them, you do not have the argument.

Target class two: logged-out marketplaces and light rate limiting

Sites that run no full anti-bot system but do throttle hard, flag repeat visitors, or quietly serve degraded content to addresses they distrust. Mid-size retail, job boards, real estate portals, most listing sites of any size.

Two answers here, and session length decides between them.

If the crawl is stateless and you are just fetching pages, cheap rotating residential is the pragmatic choice. Evomi advertised residential at $0.49/GB on 27 July 2026, the lowest headline rate among the vendors checked for this guide, with a one-day trial. Decodo, the 2025 rebrand of Smartproxy, headlines residential at $2/GB, though its own pricing page on the same date showed the smallest monthly plan at 3 GB for $11.25 plus VAT, which works out to $3.75/GB, and pay-as-you-go at $4.00/GB. That gap between headline and entry price is the single most reliable pattern in proxy pricing, and it is worth checking before you build a budget on a number from a comparison table.

If the crawl holds a session, ISP is the better tool. A static residential proxy gives you an address that reads as a consumer ISP but behaves like a datacenter box: stable, fast, yours for the month. IPRoyal listed ISP proxies from $1.80 per proxy on 27 July 2026, datacenter from $1.39 per proxy, and residential from $1.75/GB with, in its own words, "traffic that never expires". Non-expiring traffic is unusual in this market and it matters more than it sounds for lumpy workloads, because the normal alternative is buying bandwidth monthly and forfeiting whatever you did not burn.

Webshare sits in this tier too, with static residential from $0.30 per proxy per month at 20 addresses falling to $0.225 at 10,000, and rotating residential from $3.50/GB at 1 GB down to $1.40/GB at 3,000 GB. Two comparable products, two different billing units, and no way to compare them without knowing your own bytes per page.

Who ISP is wrong for: anyone who needs pool diversity. ISP pools are small by construction. If your access pattern requires a fresh identity per request, you will exhaust an ISP allocation and start recycling addresses the target has already seen, which is worse than not rotating at all because it teaches the target exactly which addresses belong to you.

Target class three: CDN-fronted sites with bot management

On this tier the address you buy decides whether you get a response at all, which is why the vendor explanation of what happens here is worth checking against primary sources. The standard claim, repeated on every vendor page, is that anti-bot systems classify by network at the edge and reject hosting ranges before they look at anything else. What the documentation says is more specific and slightly different in shape. Cloudflare's own page on its bot score describes four detection engines: heuristics, machine learning, anomaly detection and JavaScript detections. Its description of the machine learning engine names "headers, session characteristics, and browser signals" as inputs. Network origin is not listed among the documented bot score inputs at all.

What Cloudflare does document, in its Rules language field reference, is that site operators can write their own rules against ip.src.asnum, the autonomous system number of the client address, alongside cf.threat_score and the TLS fingerprint fields cf.bot_management.ja3_hash and cf.bot_management.ja4. Both pages were read on 27 July 2026.

The distinction matters. Network origin appears in the documentation as something a site operator filters on directly, in a rule they wrote, alongside your TLS fingerprint, rather than as an input the vendor scores for you. The outcome at the edge is the same, which is why the folk explanation survives, but the shape of the fix is different: you are up against a rule about your address and possibly a second rule about your handshake, and buying more expensive addresses only ever answers one of them. The blocked-scraper diagnostic guide walks through telling them apart from the response.

For the address half of the problem, rotating residential is the answer. Rotating proxies from a large pool give you a fresh consumer network per request, which is the one signal a defended site cannot cheaply reject without also rejecting its own customers.

At the top of this tier sit Bright Data and Oxylabs, both selling all four types with network and carrier-level targeting. On 27 July 2026 Bright Data advertised "Over 400M+ monthly ethical residential IPs from 195 countries" at $4.00/GB pay-as-you-go, marked as a 50% discount off a regular $8/GB. Oxylabs advertised "175M+ highest-quality proxies" with an entry tier at $6/GB. The Bright Data vs Oxylabs comparison covers where else they differ. Both are wrong for a team that wants to swipe a card and start this afternoon, and both attach a sales conversation to anything beyond the entry tier.

Below them, IPRoyal, Decodo and Evomi all sell rotating residential self-serve at rates that make experimentation affordable. NodeMaven takes a different angle, filtering its pool for address reputation and selling residential at $2.20/GB on a monthly plan or $4.50/GB pay-as-you-go, with a $3.50 trial covering 750 MB usable across its residential and mobile pools. The trial is the point. It is cheap enough to test against your actual targets in an afternoon, which is worth considerably more than the 99.54% success rate the same page advertises. That figure is the vendor's own and is not independently audited. Neither is anyone else's.

Who residential is wrong for: high-volume crawls of undefended pages, where you pay per-gigabyte rates for detection you never needed. It is also wrong when the block is fingerprint-driven rather than address-driven.

Target class four: carrier-gated and mobile-only

Mobile app backends, platforms that gate on carrier attribution, and a small set of consumer services that treat mobile addresses as inherently more trustworthy. The mechanism is carrier-grade NAT: thousands of real subscribers share one public address, so blocking it has a real cost to the site. That is a short list, and most target lists never touch it.

Proxidize is the specialist, running mobile proxies on real carrier connections with carrier and city targeting from one dashboard and API. Its mobile pricing on 27 July 2026 was $2/GB flat across the standard plans, from a $50 25 GB tier up to $400 for 200 GB, with enterprise volumes quoted from $0.50/GB. Its origins were modem hardware for teams building their own proxy farms, and that self-hosted line still exists alongside the managed service. NodeMaven sells mobile at the same $2.20/GB as its residential product, which is unusual: mobile normally carries a premium.

The cost picture is not uniform, and that is the interesting part. Proxidize's own residential line starts at $1/GB against its $2/GB mobile, so mobile is roughly double within that one vendor. Rayobyte prices rotating mobile at $0.50/GB but only inside a $250 trial block covering the first 500 GB. IPRoyal does not sell mobile per gigabyte at all, listing it from $117 a month. Three vendors, three billing shapes, no comparable unit. Anyone quoting you a single market rate for mobile has not looked.

Mobile is wrong for anyone who has not first confirmed that residential fails on the specific target. "Mobile is the strongest type" is true about trust and almost never true about cost per successful request.

Which vendors sell which type

Checked against each vendor's own pricing and product pages on 27 July 2026. Lineups shift quarterly, so verify before you sign anything.

The practical read: if you expect to escalate between types mid-project, buying from a vendor that sells all four saves a second procurement cycle and a second integration. Rayobyte, IPRoyal, Decodo and Evomi all cover the full range self-serve. If you already know you need exactly one type, a specialist is usually cheaper per gigabyte, and Proxidize is the clearest example of that on the mobile side.

Two entries in that table are worth reading as warnings rather than gaps. Webshare has no mobile product, so a target that turns out to need carrier addresses means a second vendor. NodeMaven has no datacenter product, so it cannot serve the cheap bottom rung of the ladder at all. Neither is a defect. Both are reasons to check the whole ladder against your target list before you standardise on one account.

Pool size is close to useless as a buying signal

The advertised figures do not describe the same quantity. Bright Data says "over 400M+ monthly" residential addresses. Oxylabs says "175M+". Proxidize says "millions of real IPs" across 195 or more countries and declines to be precise. Webshare says "80m+". None of these is audited, and the qualifiers differ in ways that matter: a monthly count of every address that appeared at least once is a different number from addresses reachable in your target country during your crawl window, and vendors are free to headline whichever is largest.

There is no route from a pool figure to a decision. The number you actually need is cost per successful request against your own target list, and every vendor named above offers a free tier, a trial or pay-as-you-go, which makes that measurable in a day. We publish no success-rate figures of our own on this site, and you should treat anyone selling proxies who publishes theirs as reporting a marketing number until your own retry logs agree with it.

The fallback ladder

Assume the cheap option works until it does not, and build the escalation before you need it.

1. Datacenter. Cheapest per request by a wide margin. If success holds, stop here. Most teams never test this rung and overpay permanently as a result.
2. Datacenter with a real browser fingerprint. Before spending more on addresses, check whether the block is fingerprint-driven. A residential address paired with a bare HTTP client still presents a TLS handshake that no browser produces. This rung costs engineering time rather than money and it resolves a surprising share of blocks.
3. ISP static. Adds consumer ISP registration while keeping datacenter speed and per-address billing. Right when sessions are long.
4. Rotating residential. Adds pool diversity and genuine consumer networks. The standard answer for defended targets, and the point where your bill starts tracking bandwidth instead of address count.
5. Mobile. Carrier attribution and CGNAT cover. Reserve it for targets that demonstrably need it.

Each rung costs meaningfully more per successful request than the one below, and the gap between rungs one and four is the largest in the ladder. Skipping rungs is how proxy budgets get out of hand.

The ladder belongs in the crawler rather than in a runbook. Classify the failure, not just the status code. A 403 served instantly by an edge network is a decision about your address and warrants a rung upward. A 429 arriving after fifty successful fetches is a rate limit and warrants more addresses on the same rung. A challenge page returning 200 with no useful body is a third case entirely and needs a browser, not an address. Teams that collapse all three into "blocked, retry on residential" end up routing an entire crawl through the most expensive tier because a handful of endpoints rate-limited them in week one, and because the crawl then works, nobody revisits it. Keeping per-target success and per-target cost in one table is what makes the escalation decision arguable rather than reflexive.

Proxy type is one layer of several

The most common misdiagnosis in this category is treating the address as the whole problem. A defended site can filter on your network origin and separately fingerprint your TLS handshake, and Cloudflare's field reference shows both are available to operators as ordinary rule inputs. Two independent checks, so upgrading one axis moves nothing if the other one is what caught you: a residential address driving a default HTTP client still fails a TLS-aware rule, and a real browser on a datacenter address still fails a network rule. If a proxy upgrade does not change your block rate, stop buying addresses and look at your client. That is work for an antidetect browser or hosted browser infrastructure, not a proxy contract.

There is also the option of not managing this layer at all. A web scraping API bundles proxy selection, fingerprinting and retry logic behind one endpoint and prices per request, which removes the type-selection question entirely at the cost of control and unit economics at scale. Several vendors above sell one alongside their proxies, which is a reasonable way to buy the hard 5% of your target list while running the easy 95% on datacenter addresses yourself.

Bottom line

Start from the target, not the type. Test datacenter first, because a large share of the web still has no defence and the price difference is not marginal. Move to ISP when you need a stable identity, to residential when the target rejects you on the basis of where your address lives, and to mobile only after you have proven residential fails. Verify the block is address-driven before you climb a rung. And treat every pool figure and headline per-gigabyte rate, including the ones quoted above with a date attached, as a marketing number until your own logs contradict it.