IPv4, why there are only 4.3 billion addresses and why that's a problem

Field note. Why your host can't just give every customer a unique IP, why most servers share one IP via different ports, and why a dedicated IP costs extra.

In 1981, when the IPv4 standard was finalized, the world had a few thousand computers and growing. The protocol allowed for 2^32 addresses, just over 4 billion. The designers thought this would last forever.

It didn't. We ran out of IPv4 addresses in the 2010s. IPv4 is now a scarce, traded commodity. Networks pay tens of dollars per address on resale markets.

This is the story of how 4.3 billion became too few, and what the world has done about it.

The math

IPv4 addresses are 32-bit numbers. That gives 2^32 = 4,294,967,296 possible addresses. About 4.3 billion.

Not all are usable. Various blocks are reserved:

• Private ranges (RFC 1918): about 17 million addresses, can't be used on the public internet.
• Multicast: another 268 million.
• Loopback, link-local, reserved, etc.: another 100 million.
• Various other administrative reservations.

Usable public IPv4 addresses: roughly 3.7 billion. Still seemed like enough in 1981.

The world's population in 2026: about 8 billion. Devices per person: 5 to 10 in developed countries. Servers in data centers: hundreds of millions. The math doesn't work.

When we ran out, exactly

February 3, 2011. IANA, which holds the global IPv4 pool, gave out its last /8 blocks (each containing ~16 million addresses) to the regional internet registries. There was a ceremony.

After IANA exhaustion, the regional registries (ARIN, RIPE, APNIC, etc.) had their own pools to work through. They ran out at different times:

• APNIC (Asia-Pacific): April 2011.
• RIPE (Europe): September 2012.
• LACNIC (Latin America): June 2014.
• ARIN (North America): September 2015.
• AFRINIC (Africa): exhausted in 2017, with some controversy over remaining allocations.

After exhaustion, RIRs moved to small-allocation policies (newcomers can get small blocks for specific transitional purposes) and to facilitating a transfer market: networks with surplus IPv4 can sell to networks that need it, with the RIR brokering.

How the early decisions concentrated addresses

In the 1980s and early 1990s, IPv4 was allocated in classful blocks:

• Class A: /8 (16.7M addresses).
• Class B: /16 (65K addresses).
• Class C: /24 (256 addresses).

Early adopters got large blocks. MIT, Stanford, IBM, GE, AT&T, the U.S. Department of Defense, and many others received /8 blocks (16.7 million addresses each). At the time, this seemed fine.

Today, those legacy /8 holders have far more addresses than they need. A /8 represents 0.39 percent of the entire IPv4 space. Holding one is like owning a city block in 1850 manhattan.

Some legacy holders have sold portions of their blocks (MIT sold half of theirs to Amazon Web Services in 2017). Others sit on them. There's no mechanism to force them to give addresses back.

In 1993, classless addressing (CIDR) replaced the rigid class boundaries. After that, addresses were allocated in arbitrary sizes (/19, /22, etc.). This was much more efficient but came too late to undo the early bulk allocations.

The price of IPv4 today

IPv4 addresses are tradeable. Prices have climbed steadily.

• 2014: about $7 per address.
• 2018: about $20 per address.
• 2021: about $50 per address.
• 2024-2026: $40 to $50 per address, with some volatility.

A /24 (256 addresses) costs about $12,000 at current rates. A /22 (1024 addresses) costs about $50,000.

For perspective: a typical hosting customer needing 5 IPs for their server costs the host $250 in IPv4 alone. This is why some hosts now charge an "IPv4 surcharge" for additional addresses, or push customers toward IPv6.

What the workarounds look like

The IPv4 shortage has been managed (not fixed) through several techniques.

NAT

Already covered in earlier articles. Network Address Translation lets many devices share one public IP. A home network with 30 devices uses one public IP, with the router translating between the inside and outside.

NAT has saved IPv4 from total collapse but it has real costs:

• Inbound connections to specific inside devices require port forwarding.
• Some protocols (older ones, P2P, SIP) break through NAT.
• Multiple layers of NAT (NAT inside NAT) cause issues.

CGNAT

Carrier-grade NAT. The ISP runs a giant NAT, with many customers sharing one public IP. Mobile networks especially use CGNAT.

CGNAT means you don't have a unique public IP at all. The IP your services see is shared with thousands of other customers. Many features (port forwarding, hosting from home) don't work.

CGNAT is one of the reasons why hosting a game server from a residential mobile or cable connection has gotten harder. CGNAT also makes "what's my IP" answers unhelpful (it's a shared IP).

Cloud and CDN consolidation

Modern services often share IPs at the application layer. A single Cloudflare IP can serve traffic for millions of different websites; the right one is identified by the HTTP "Host" header (or TLS SNI). Same idea for many cloud platforms.

This is efficient for HTTP but doesn't help for protocols that don't carry a hostname (raw TCP/UDP services like games).

The transfer market

Existing IPv4 holders sell unused space. Brokers facilitate. Prices set by supply and demand.

This works but is regressive: it favors networks that can afford to buy IPv4, and lets early holders extract rent indefinitely.

What hasn't worked

Several proposed fixes haven't materialized:

Reclaiming legacy blocks. Politically and legally difficult. Legacy holders have property-like claims to their blocks.

Forcing IPv6. Tried in many ways. IPv6 adoption is slowly increasing but has taken 30 years to reach ~40 percent global traffic. Forcing-style mandates haven't worked.

Reservation reclamation. A few small blocks have been reclaimed (240/4 used to be "reserved future use" but is now allocated in some forms). Drop in the bucket.

What it means for users

Day-to-day:

Your home connection probably has one public IPv4. Some ISPs offer additional IPs for a fee. CGNAT customers don't have a dedicated one at all.

You may share an IP with strangers. CGNAT is increasingly common, especially on mobile. Your "IP" might be shared with hundreds or thousands.

Hosting a server from home is harder than it used to be. No public IP means no inbound connections. Port forwarding requires a non-CGNAT public IP.

Services that ban "by IP" sometimes affect innocent users. A spammer on a CGNAT'd ISP can get the shared IP blocked, affecting everyone else on it.

Geographic IP databases are imprecise. With IP transfers and CGNAT, the IP-to-location mapping is increasingly noisy.

What it means for hosting

For game hosting and similar services:

IPv4 is a real cost. Every public IP a host gives you represents $40 to $50 in capital, plus operational overhead. This is reflected in pricing.

IPv6 reduces the cost. Hosts that aggressively use IPv6 internally (or push customers to IPv6) save on IPv4. Many haven't because customer demand for IPv4 remains.

Anycast multiplies effective addresses. A single IPv4 advertised from multiple locations (anycast) reaches more customers efficiently. We have a dedicated anycast article in Series B.

Reverse DNS, SPF, and reputation matter. When you share an IPv4 block with other customers, one bad neighbor can taint the block's reputation for everyone. Higher-quality hosts manage this carefully.

The transition that won't end

For nearly 20 years, "we'll all be on IPv6 soon" has been the technical consensus. It's true and it's slow.

As of 2026:

• Global IPv6 traffic: ~40 percent of total internet traffic.
• Top services (Google, Facebook, Netflix): ~50 to 70 percent over IPv6.
• ISP support: ~60 percent of broadband connections globally have IPv6 in some form.
• Hosting: many providers offer IPv6 alongside IPv4 but few have IPv6-only options.

The transition is happening. It's just glacial. The economic incentives are weak (NAT works "well enough"), and every device, app, and service has to support IPv6 before you can move off IPv4 entirely.

Realistically: IPv4 will be relevant for another 10 to 20 years. IPv6 will become dominant gradually. Neither will fully disappear in our lifetimes.

Why the world didn't see this coming

The 1981 designers weren't stupid. They optimized for the technology of their day:

• A 32-bit address was a single machine word on common CPUs.
• It fit in memory tables of the era.
• The cost of more bits would have been real.

What they didn't predict:

• The PC revolution (millions of personal computers).
• The mobile revolution (billions of phones).
• The cloud (vast server farms).
• The IoT (sensors, fridges, lightbulbs).

It's easy to criticize in hindsight. At the time, 4 billion was generous. Engineering decisions live with the world they're made in.

Coming up

IPv4 ran out. IPv6 is the answer. What does IPv6 actually look like, and why has it taken so long? That's the next article.