← back to blog

Carrier Grade NAT: Why Thousands Share One Mobile IP And Platforms Accept It

Put a SIM in a phone, switch Wi-Fi off, and look at the phone’s IP address in its network settings. Then open a “what is my IP” page in the same phone’s browser. On a lot of mobile networks the two won’t match. The phone’s own address will often sit in a range starting with 10 or with 100.64 to 100.127, and the page will show a completely different public address. Those two numbers are the whole mechanism behind the mobile proxy argument. The first is the address your carrier gave your phone inside its network. The second is the address the carrier shares between your phone and a large number of other subscribers when their traffic reaches the internet. That sharing is called carrier grade NAT. It’s the reason platforms treat mobile addresses gently, and it’s also misunderstood more than almost anything else in this field. This piece covers how it actually works, why platforms put up with it, and the specific places where that tolerance stops.

The shortage that created it

IPv4 has about four billion addresses. That sounded like plenty when the protocol was designed. It wasn’t enough for a world with several connected devices per person, and the regional registries that hand out addresses ran through their free pools during the 2010s.

Carriers were hit hardest. A mobile network may have tens of millions of subscribers, each with a phone that expects to be online all the time. Buying a public IPv4 address for every one of them became impossible or ruinously expensive.

The answer was to do at carrier scale what your home router already does. At home, your router has one public address and gives every device behind it a private one. When your laptop connects to a site, the router rewrites the source address to its public one and keeps track of which connection belongs to which device. That’s network address translation, NAT.

Carrier grade NAT is the same idea placed inside the carrier’s network. Phones get addresses from a range that only works within the carrier, and big NAT gateways translate their traffic onto a pool of public addresses on the way out. A block of address space, 100.64.0.0/10, was set aside specifically for this purpose, so carriers wouldn’t collide with the private ranges people use at home. Some carriers use that block, and some use the ordinary private 10 range internally.

The result is that many subscribers share each public address, at the same moment, all day long.

How the translation works

It’s worth being concrete, because the details explain what platforms can and can’t see.

When your phone opens a connection, it sends a packet from its internal address and some source port. The carrier’s NAT gateway picks one of its public addresses and a public source port, rewrites the packet, and records the mapping. Replies from the site come back to that public address and port, and the gateway uses its table to send them on to your phone.

Many carriers allocate ports in blocks. Each subscriber gets a range of ports on a particular public address for a period, and all of that subscriber’s connections use ports inside that range. That makes the NAT easier to manage and to log.

Logging matters here. Because so many people share an address, a public IP and a timestamp don’t identify a subscriber on their own. In many countries carriers are required to keep enough records to answer a lawful request, which in practice means the public address, the port and the time. There’s even a long standing recommendation that servers log source ports for this reason, so a site’s records can be matched to a subscriber by the carrier if a court asks.

The practical consequence for everyone else is simple. A platform looking at its own logs sees a public mobile address. It can’t turn that into a person. Only the carrier can, and only through a legal process.

How many subscribers share one address varies by carrier, by time of day and by how the gateway is designed. It can be hundreds or far more. Nobody outside the carrier knows the ratio for a specific address, and it changes constantly.

CGNAT is not anonymity

One misreading is worth clearing up early, because it leads people to the wrong conclusions about what a mobile address does.

Sharing an address with a crowd makes it hard for a platform to single out one subscriber from its own logs. It doesn’t make anyone untraceable. The carrier knows exactly which subscriber held which public address and port range at which moment, because it has to, both to run the network and to meet its legal obligations. In many countries those records are kept for a set period so they can be produced in response to a lawful request.

So the protection CGNAT gives is narrow and specific. It’s protection against crude blocking by a platform that only has the address to go on. It isn’t protection against the carrier, a court order, or an investigation that asks the carrier who was behind a connection. Anyone treating a mobile address as a way to hide from the law has misunderstood the technology, and that isn’t what this channel is about in any case.

For legitimate operators the distinction is practical rather than philosophical. You’re buying a network layer that looks like ordinary consumer traffic, which is exactly what it is. The accounts on it still belong to someone, they’re still subject to the platform’s rules, and the connection is still attributable to a real SIM with a real registered owner.

That last point is a reason to keep your mobile supply clean. SIMs registered properly, used for work you’d be comfortable explaining. The crowd hides your accounts from lazy blocking. It doesn’t change who you are responsible to.

What a platform sees on a mobile address

From the platform’s side, a busy mobile address looks unusual by the standards of a home connection.

It sees many different devices arriving from it within a short window. Different phone models, different operating system versions, different apps. It sees many different accounts logging in from it, most of them unrelated to each other. It sees traffic at all hours, because the population behind it is large. And it sees the address change hands over time, since carriers rotate subscribers across their pool.

On a home broadband address, that pattern would be alarming. A single household address with dozens of unrelated accounts and devices looks like a proxy or a farm. On a mobile address, the same pattern is normal, because that’s what CGNAT produces.

Platforms learn which addresses are mobile in a few ways. The address belongs to a network registered to a mobile carrier, which an ASN lookup shows immediately. Commercial IP intelligence databases label connection types, and mobile is one of the labels. And the platform’s own traffic reveals it, since a mobile address shows exactly the many devices, many accounts behaviour described above.

Once an address is known to be mobile, the platform can expect that pattern, which is where the tolerance comes from.

Why blocking it is expensive

The core of the argument is about cost. When a platform blocks or heavily penalises an address, everyone on that address pays for it.

A datacenter address typically carries no ordinary consumers at all. Blocking it costs the platform almost nothing. A home address carries one household. Blocking it costs one household’s worth of goodwill. A busy mobile address carries a large and constantly changing crowd of real customers, most of whom have never done anything wrong. Blocking it locks all of them out, generates support tickets, loses sales and makes the product look broken to people who pay for it.

So platforms avoid treating mobile addresses as strong evidence against an account. They’re far more likely to use softer tools on mobile ranges. A captcha. A verification prompt. A rate limit on a specific action. And they shift the weight of their decisions onto signals that belong to the account or the device rather than the address.

This is the part people get right about mobile proxies. The address is perfectly visible. What protects it is the crowd around it, which the platform has strong reasons not to punish as a whole.

It’s also the part people take too far. “Platforms won’t block a mobile IP” is roughly true. “Platforms won’t act on an account that uses a mobile IP” isn’t true at all, and the gap between those two statements is where most of the rest of this piece lives.

The weight moves somewhere else

When a platform can’t lean on the address, it leans harder on everything else. That’s the most important consequence of CGNAT for anyone running accounts, and it’s the one least discussed.

On a datacenter address, a platform can often act on the address alone. The account barely gets examined. On a mobile address, the address is weak evidence, so the account gets examined more carefully.

That means device signals matter more. The browser fingerprint or the phone’s own identifiers. Whether the device has been seen with other accounts. Whether the device looks real.

Account signals matter more. Its age, its history, its verification status, its contacts and connections, whether its behaviour looks like a person’s.

Behaviour matters more. What the account does in a session, how fast, how regularly, and whether it moves in step with other accounts.

So a mobile proxy doesn’t reduce the total scrutiny an account receives. It moves the scrutiny from the layer you’ve paid to improve onto the layers you still have to get right yourself. An operator who buys mobile proxies and neglects the device and behaviour layers has bought protection for the one layer that was no longer carrying the weight.

That’s not a criticism of mobile proxies. It’s what they’re for. They take the network out of the list of easy reasons to act. The remaining reasons are still there.

Where the tolerance ends: the account itself

The first limit is the one covered above, and it’s worth restating plainly. A platform that won’t block a mobile address can still restrict, verify or remove an account that happens to use it.

Crowd protection works at the address level. It stops a platform from saying “this address is bad, so everything on it is bad.” It does nothing about a platform saying “this specific account looks automated” or “this account shares a device with an account we removed.”

Inside the crowd, each account is still judged on its own evidence. An account on a mobile address that signs up, immediately follows hundreds of profiles and posts identical links looks like what it is, whatever address it uses. The address being mobile only means the platform didn’t have an easy shortcut to the answer.

So the question to ask of any account on a mobile address is whether it would survive if the address were perfectly neutral. If the answer depends on the address shielding it, the account is relying on protection the address doesn’t actually provide.

Where the tolerance ends: the address’s own baseline

The second limit is subtler. A mobile address carries a crowd, but platforms can learn what that crowd normally looks like.

Every busy address has a baseline. A certain rate of new signups, a certain mix of device types, a certain proportion of logins that later get flagged. Platforms that see a lot of traffic can model those baselines per address or per range.

If one address suddenly produces far more new accounts than its baseline, or a burst of accounts that all behave the same way, the anomaly stands out against the address’s own history. The platform doesn’t need to block the address. It can add friction to new signups from it for a while, or scrutinise accounts created from it during that window.

This is why a burst of registrations from one mobile address can still go badly even though the address is shared. The crowd is real, but the crowd doesn’t normally create dozens of accounts in an hour. The anomaly is measured against what’s normal for that address, not against zero.

I can’t tell you what baselines any particular platform models, or how sensitive they are. The mechanism is standard anomaly detection, and it’s cheap to build for anyone with the traffic.

Where the tolerance ends: IPv6

The third limit is structural and very often missed. CGNAT is an IPv4 technique. IPv6 doesn’t need it.

Many mobile carriers now give phones native IPv6 alongside a shared IPv4 path. On IPv6, each phone typically gets its own prefix, commonly a /64 of its own. Nothing is shared. When a phone connects to a platform over IPv6, the platform sees an address block that belongs to that one subscriber for as long as the carrier keeps the assignment.

Big platforms almost all support IPv6. A phone on a dual stack mobile network will often reach them over IPv6 by preference. So for a lot of real mobile traffic, the crowd that protects the IPv4 address simply isn’t there. The platform sees one device’s own prefix.

That has two consequences for operators. First, it’s normal. Real phones on those networks look like this, so a mobile identity with its own IPv6 prefix isn’t strange. Second, it means the “thousands share this address” protection only applies to whichever traffic actually goes over IPv4. If your mobile proxy exits over IPv6 on a carrier that assigns per subscriber prefixes, several accounts sharing that proxy share a single subscriber’s prefix, with no crowd around them.

Some carriers are IPv6 only inside their networks and provide IPv4 through translation, which adds more variety. Whether a given mobile connection exits over IPv4, IPv6 or both is decided by the carrier and the specific SIM and modem, and the only way to know is to test through that connection.

Where the tolerance ends: rotation and timing

The fourth limit comes from how mobile proxies are usually run.

A mobile proxy is typically a modem or phone with a SIM, connected to the carrier. When it reconnects to the network, it usually gets a new public address from the carrier’s pool. Many proxy services let you trigger that on demand, which is called rotation.

Rotation gives you a different address in the same crowd. It doesn’t change the device behind it, or the accounts using it. If several accounts share one modem, they all move to the new address together, at the same moment. A platform that sees a set of accounts leave one address and reappear together on another, every time, has found a group without needing either address to be suspicious.

The same goes for timing in general. Several accounts on one modem are several accounts in the crowd, but their sessions start, pause and end around the same times, because they share one connection that goes up and down together. When the modem drops for a minute, every account on it drops.

Real subscribers in the crowd don’t move like that. Each phone connects and disconnects on its own schedule.

So the practical rule is to keep the number of accounts per modem modest and to avoid rotating in response to every small event. How modest depends on the platform and the value of the accounts, and nobody can give you a universal number. Watch your own results and tighten if you see groups getting flagged together.

A worked example on one modem

Here’s a hypothetical that pulls the limits together, since most mistakes come from several of them at once.

An operator has one mobile modem on a large carrier. Eight accounts share it. Every account runs in its own browser profile with its own fingerprint, so the device layer is fine. The operator rotates the modem’s address every ten minutes, because rotation sounded safer.

Look at it from the platform’s side. Every ten minutes, eight accounts that were active on one mobile address disappear from it and reappear on another mobile address within seconds of each other. Neither address is suspicious. Both carry crowds. But the pattern of eight accounts moving together, over and over, draws a clear line around those eight. The platform doesn’t need the addresses at all. It has the choreography.

Now suppose the carrier gives that modem IPv6 as well, and the browser prefers IPv6 for the platform. The eight accounts now share one subscriber’s IPv6 prefix, with no crowd on that path. That’s a second, simpler link.

Add one more fact. On a Tuesday the operator creates five new accounts on the same modem within half an hour. For that address, the signup rate jumps well above anything its normal crowd produces.

Each of those problems has an easy fix. Fewer accounts on the modem, split across more modems. Rotation only when there’s a reason, such as a session ending, rather than on a timer. IPv6 checked and either kept in the story deliberately or disabled. New accounts spread across addresses and days.

None of this required a bad address. The address did its job. The setup around it gave the game away.

Dedicated still means shared

Mobile proxy providers often sell “dedicated” lines, and it’s worth being precise about what that word can and can’t mean under CGNAT.

A dedicated mobile proxy usually means the modem and its SIM are yours alone. No other customer of the proxy service sends traffic through that device. That’s valuable. It means nobody else’s accounts are moving in step with yours on the same connection, and nobody else’s behaviour is tied to your device’s timing.

It can’t mean the public IPv4 address is yours alone. The carrier decides that, and on a CGNAT network the public address is shared with ordinary subscribers who have nothing to do with you or the proxy service. That’s the crowd, and it’s the whole point.

So a dedicated mobile line gives you your own device inside a shared crowd. A shared mobile line from a proxy pool gives you a device that other customers also use, inside the same crowd. The address layer looks similar in both cases. The difference is who else is on your side of the NAT, doing things at the same moments as you.

On IPv6 the picture changes again. If the SIM gets its own prefix, a dedicated line gives you a prefix no other customer touches, and a shared line gives every customer on it the same prefix.

When you compare providers, ask them to explain their use of the word. Whether the modem is yours, whether the SIM is yours, whether rotation is under your control, and whether other customers can ever land on the same device. Those answers matter more than the label.

Home broadband uses CGNAT too

It’s worth knowing that CGNAT isn’t only a mobile thing. Some fixed broadband providers, especially newer ones and those that ran short of IPv4 space, put home customers behind CGNAT as well. Fixed wireless and home 5G routers commonly do.

For an operator this cuts two ways.

If you’re buying residential proxies, some of those addresses may be CGNAT addresses on a fixed line provider. That gives them some crowd cover, but a smaller and more stable crowd than a large mobile pool, and platforms may treat them differently.

If you’re running a fleet from home, you might discover your own connection is behind CGNAT. You can check by comparing the WAN address your router reports with what a “what is my IP” page shows. If the router’s WAN address is in 100.64.0.0/10 or a private range, you’re behind CGNAT. That doesn’t make your home address a mobile address in the platform’s eyes. The ASN still says it’s a fixed line provider, and the crowd behind it is typically smaller.

It does explain some odd experiences, like a home connection getting more captchas than expected because other subscribers sharing the address have done something that raised its score.

Shared reputation cuts both ways

Sharing an address means sharing its reputation. The crowd gives you cover, and it also hands you the consequences of whatever the crowd does.

If other people on the same carrier address are running aggressive scrapers, spam or automated signups, the address’s reputation drops. Platforms can’t block it outright for the reasons above, but they can add friction to everyone using it. You might see more captchas and verification prompts on a mobile address than on a quiet home line, simply because of what others behind the same gateway are doing.

That’s especially relevant when many proxy operators draw on the same carrier. If a carrier’s ranges in one country become popular with proxy services, the proportion of automated traffic on those ranges rises. Platforms may respond with more friction for the whole range. Nobody’s address is blocked, but everyone’s experience gets a little worse.

You can’t control this directly. You can notice it. If accounts on one carrier start hitting more challenges than accounts on another, with similar setups and behaviour, the carrier’s range may have drifted. That’s useful information when deciding where to put new accounts.

It’s also a reason not to judge a mobile proxy by one bad day. The crowd’s behaviour changes, and so does the friction it attracts.

How to tell whether an address is CGNAT

If you want to confirm what you’re dealing with, there are a few simple checks.

For a mobile connection you control, compare the device’s own interface address with the public address a web page reports. An interface address in 10.x or 100.64 to 100.127 with a different public address means CGNAT on IPv4.

For any address, run an ASN lookup. It tells you which network owns the address. A mobile carrier’s network is a strong hint the address carries mobile traffic, although carriers also run fixed services from some of their networks.

IP intelligence services label connection type, typically as mobile, residential, business or hosting. Their labels are estimates, but they’re the same kind of labels a platform’s vendors produce.

Check IPv6 separately. Load an IPv6 test through the connection. If it shows an IPv6 address, note that traffic over IPv6 won’t share the CGNAT crowd.

For a mobile proxy you’re considering buying, ask the provider which carrier the SIMs are on and whether exits are IPv4, IPv6 or both. Then test it yourself. A provider that can’t answer is telling you something too.

What to do with this

If you use mobile proxies, or are thinking about them, here’s how the mechanism translates into decisions.

Treat the mobile address as removing one easy reason for a platform to act, not as a shield. Build every account so it would pass if its address were neutral.

Put your effort into the layers that now carry the weight. A consistent device per account, sensible behaviour, a believable account history.

Keep accounts per modem modest, and spread their activity so they don’t start and stop together. Rotate for a reason, not by reflex.

Check whether your mobile exits use IPv6, and if they do, remember that IPv6 traffic isn’t in a crowd.

Pace new account creation per address so it stays within what that address’s crowd could plausibly produce.

Watch for carrier level drift in challenges, and diversify across carriers where you can.

Keep records. Which accounts sit on which modem, which carrier, when each was rotated. When something goes wrong, those records are how you tell an address problem from an account problem.

The honest limit

Carrier grade NAT is real, and the tolerance it creates is real. It’s the most solid reason mobile addresses behave better than datacenter ones for account work. It isn’t a guarantee of anything.

I can describe the mechanism with confidence, because it’s how carrier networks are built. I can’t tell you how any specific platform scores mobile addresses today, how many subscribers share a particular address, or where a given platform’s thresholds sit. Those are internal to the carrier and the platform, and they change.

What I can tell you from running mobile proxy hardware is that the address is only ever one layer. The accounts that do well on mobile are the ones that would have done well anyway, with the mobile address taking away a reason to look at them too closely. The accounts that do badly are usually the ones that were counting on the crowd to hide something the crowd was never going to hide.

I run Singapore mobile proxy and cloud phone farms myself, on real SIMs behind real carrier NAT, so this is the mechanism I work with every day. If you want the rest of the network picture, from proxy types to how many accounts a line can carry, start on the Multi Account Ops home page.

Get new guides and videos first — join the Telegram channel.

need infra for this today?