ISP proxies

An ISP proxy — sold as a static residential proxy almost as often — runs on datacentre hardware but exits from an address registered to a consumer internet provider, and it is assigned to you alone. It suits work that needs one identity to persist for weeks and to move real volume while it does — a shape of work the other three categories serve badly.

Three properties, and only one of them is the hybrid

Property	Where it comes from	What it gives you
Registration to a consumer ISP	The address allocation the provider leases	Classified like a home line, not like infrastructure
Hosting in a datacentre	The physical machine and its uplink	A short, stable path that does not sleep
Exclusive assignment	The purchase model	A reputation built only from your own traffic

The first two rows are what the marketing sells. The third decides whether the purchase works. On a residential pool you inherit whatever the previous user of that address did an hour ago; on a shared datacenter address you inherit it from a stranger in real time. Here the history is yours, which makes your own behaviour the variable — and means a mistake is yours to live with rather than something rotation washes away.

One caveat sits under the first row. Registration says who holds the block, not where the hardware sits. An ISP can lease a block to a provider that racks it in another country, and geolocation databases update at their own pace. Operators can self-publish location for a block as a geofeed, a format defined in RFC 8805, discoverable from registry records by the method in RFC 9092 — but publication is voluntary, and your target may consult a database that ignores it. Treat the country you were sold as a claim to verify, not a fact.

Why per-address billing follows from the sourcing

The provider's cost here is a lease on an address block plus a machine to terminate it. Both are fixed monthly costs whether you send a byte or not, and datacentre bandwidth is cheap, so there is no reason to meter you and every reason to sell the durable thing: the address. That is datacenter pricing logic applied to a scarcer input, since getting a consumer allocation leased to you is the hard part of the business.

Because bandwidth is not the meter, the arithmetic that makes per-gigabyte plans painful does not apply. A rendering browser fetching every asset on a page costs what a bare HTTP request costs. So does a retry, a redirect and a challenge page you threw away. The cost climbs with the number of identities, not with the work each one does.

The shape of work this suits

Shape of work	Better served by	Why
Few identities, heavy bandwidth each	ISP	Traffic is unmetered; only identities are billed
A logged-in session held for weeks	ISP	The address never changes underneath the session
Rendering full pages with all their assets	ISP	Asset fetches cost nothing extra
Many identities, small responses each	Residential	Few bytes cross a wide pool
Broad crawling across many countries	Residential	Breadth is what a rented list cannot give
The target ignores network origin	Datacenter	The consumer registration buys nothing here
Mobile-only surfaces	Mobile	A fixed-line registration contradicts the context

The first three rows share a signature: your work is heavy in bytes, and the number of distinct identities is small and stable. That combination is where per-address billing wins decisively, and it is the case people miss by comparing headline rates rather than traffic shapes. The comparison guide works an example end to end.

The limit is breadth. An ISP allocation is a fixed list you rent, so the failure mode is not a rate — it is running out. Burn an address through careless volume and you have lost a paid asset until it is replaced, which is why rotation discipline matters more here than on a pool you draw from freely. Cycling a handful of static addresses is also readable as a pattern; drawing from a large pool is not.

Verifying what you bought, and what to ask first

Two of the three properties are checkable from your desk. The third is not checkable at all, which is why it belongs in the contract.

# 1. Registration: the holder should be a consumer ISP, not a hosting company
whois "$IP" | grep -iE 'orgname|netname|org-name|descr'

# 2. Announcing network, which is what a classifier reads first
whois -h whois.cymru.com " -v $IP" | tail -1

# 3. Stability: same address on every request, over a long run
for i in $(seq 1 50); do
  curl -s --max-time 20 -x "http://USER:PASS@$IP:8000" https://echo.example/ip
done | sort -u        # more than one line means it is not static

Exclusivity is the property you cannot test. Nothing observable from outside distinguishes an address held only by you from one quietly shared with two other customers, so get it stated in writing, ask whether it was quarantined between customers, and treat a vague answer as a no. Then ask:

• Is this address exclusively mine for the term, and was it reassigned from someone else?
• Which ISP holds the block, and can I see the registration before I pay?
• Which country is the registration in, and where is the hardware? See geo-targeting for why neither may match what a target believes.
• What is the replacement policy if the address becomes classified, and who judges that?
• Is traffic truly unmetered or capped, and what happens at the cap?
• How many addresses can I hold in the same country, and how many distinct subnets do they span? A rented list on one network shares one fate.
• Which authentication model? Address allow-listing fails silently when your own address changes; credentials survive that and end up in logs.

What we do not publish about this category

No prices per address, no throughput figures, no success rates and no counts of how many addresses a provider holds. Per-address pricing varies with the country and the block, and it moves, so a number here would be stale before it was useful. Throughput is a property of the route between your machine and the datacentre, so ours is not yours. And a success rate on an exclusive address largely measures how the previous tenant behaved.

Verify the two properties you can see, contract for the one you cannot, and measure your own targets from your own machine. Success rate sets out what a rate must state before it means anything, and the tools show what the far end sees.