If you've spent any time on IRC, you've seen a netsplit. One moment your channel has fifty people in it. The next, half of them vanish simultaneously, their quit messages reading something like *.net *.split. A few minutes later, they all flood back in. It's one of those uniquely IRC experiences — watching the network's topology become visible through its failures.
But the term "netsplit" has a second, more permanent meaning in IRC history. It refers to the deliberate, political fractures that split the original unified IRC network into the dozens of separate networks we know today. These weren't technical failures. They were divorces.
What Is a Netsplit, Exactly?
To understand the wars, you need to understand the architecture. IRC networks use a spanning tree topology — servers are linked in a tree structure with no loops. Every server connects to one or more other servers, and messages route through this tree to reach all users.
A netsplit occurs when a link between two servers breaks. The tree splits into two (or more) separate sub-trees. Users on each side can still talk to each other, but the two halves can't communicate. From a user's perspective, everyone on the "other side" appears to quit simultaneously.
Before netsplit:
ServerA -- ServerB -- ServerC -- ServerD
|
ServerE
After ServerB--ServerC link breaks:
[ServerA -- ServerB] [ServerC -- ServerD]
|
[ServerE]
Two separate networks until the link is restored.Temporary netsplits are normal operational events — a flaky network connection, a server restart, a routing hiccup. They're annoying but self-healing. The permanent splits were something else entirely.
The First Great Split: EFnet and the Anarchy Model
The original IRC network — what would eventually become EFnet — had no central authority. Any server administrator could link their server to the network if they could convince an existing server admin to accept the link. There were no formal policies, no governing body, no terms of service. It was anarchy in the most literal sense.
This worked fine when the network was small — a few dozen servers, mostly at universities. But as IRC grew through the early 1990s, problems emerged. Some servers were poorly maintained. Some server administrators abused their access. Some servers were linked over unreliable connections that caused constant netsplits, disrupting the entire network.
The fundamental question was: who decides which servers can link? In a decentralized system with no hierarchy, the answer was "whoever has the political clout to convince the right people." This led to endless arguments on the IRC operator channels about linking policies, server standards, and administrative responsibility.
EFnet — the name stands for "Eris-Free network," a reference to a specific incident involving a server called eris.berkeley.edu that was delinked after disputes — settled into a model of rough consensus among server administrators. It worked, more or less, through a combination of social pressure, tradition, and the occasional forceful delink of problematic servers. But it was messy, and not everyone was happy with the anarchy.
The Undernet: A Different Philosophy
In October 1992, a group of IRC users and administrators who were frustrated with EFnet's chaos decided to start fresh. They created the Undernet, which took a fundamentally different approach to network governance.
Where EFnet was anarchic, the Undernet was structured. They established formal committees for server linking, created written policies, and — most importantly — developed a centralized service system. The Undernet's "X" and "W" bots (later CService) provided channel registration, meaning you could "own" a channel name and have it protected even when you weren't online.
This was revolutionary. On EFnet, channel ownership was enforced by bots — programs that sat in channels and maintained operator status. If your bot crashed or got killed, you could lose your channel to anyone who happened to be there. The Undernet's approach was that channel registration should be a network service, not a user responsibility.
The Undernet also pioneered the concept of channel service applications — you had to apply to register a channel and demonstrate that it was a legitimate, active community. This was bureaucratic, but it reduced the chaos of channel takeovers that plagued EFnet.
DALnet and the Rise of Services
DALnet was founded in July 1994 by dalvenjah (Ryan Smith). If the Undernet was a reaction against EFnet's anarchy, DALnet was a reaction against both — it aimed to be the most user-friendly IRC network possible.
DALnet introduced several innovations that other networks eventually adopted:
- NickServ: Nickname registration and protection. You could "own" your nickname and prevent others from using it.
- ChanServ: Channel registration with persistent settings, access lists, and automatic operator status.
- MemoServ: An offline messaging system — leave a message for someone who wasn't currently online.
- Longer nicknames: DALnet increased the maximum nickname length, which seems trivial but was actually controversial at the time.
DALnet's services model — NickServ, ChanServ, MemoServ — became the template that almost every other network eventually followed. Today, services packages like Anope and Atheme implement these concepts as standard features. But in 1994, it was genuinely novel.
At its peak around 2003, DALnet had over 100,000 simultaneous users. Then the DDoS attacks came. For months, DALnet was hammered by massive distributed denial-of-service attacks, likely from botnet operators who used IRC for command and control and didn't appreciate DALnet's aggressive anti-abuse stance. The network never fully recovered its user base.
IRCnet vs EFnet: The Transatlantic Divorce
The biggest split in IRC history happened in July 1996. It wasn't about server policies or services. It was about protocol changes.
European IRC administrators wanted to implement "channel delay" and "nick delay" — mechanisms that would prevent the channel and nickname collisions that occurred during netsplits. On EFnet, when a netsplit healed, both sides' versions of a channel would merge, often leading to chaos as conflicting operator lists and bans collided. Channel takeovers during netsplits were a constant problem.
The proposed solution was TS (TimeStamp) — a way to track when channels were created and when users joined, so the network could determine which "side" of a split had the legitimate state. American administrators favored one approach to timestamps. European administrators favored another — specifically, they wanted nick delay and channel delay features that would temporarily lock nicknames and channel names after a netsplit to prevent abuse.
The disagreement was both technical and political. Neither side was willing to adopt the other's solution. After months of arguments on the oper channels and mailing lists, the European servers simply... disconnected. They formed IRCnet, taking roughly half the network's users with them.
From that point on, EFnet and IRCnet were separate networks with separate cultures. EFnet remained primarily American, kept its anarchic traditions, and eventually implemented its own timestamp system. IRCnet was primarily European, had no services (NickServ, ChanServ — none of it, by deliberate design philosophy), and maintained that users should manage their own channels through bots.
I was on the European side when it happened. One day there was one network. The next day there were two. Channels were duplicated, communities split along geographic lines, and nothing was ever quite the same.
QuakeNet and the Gaming Wave
QuakeNet was founded in 1997 as a network for Quake players to organize games. It grew to become one of the largest IRC networks in the world, peaking at over 240,000 simultaneous users around 2005.
QuakeNet took yet another approach to services. Instead of NickServ (nickname registration), they implemented "Q" — an authentication service where you had an account name separate from your nickname. You could use any nickname you wanted, but your identity was tied to your Q account. This was arguably a better design than NickServ's approach, where your identity was tied to a specific nickname string.
The gaming focus also meant QuakeNet attracted a younger, less technical user base than EFnet or IRCnet. This shaped the network's culture — less about Unix and programming, more about gaming clans and social chat. The administrators adapted with stricter abuse policies and more automated moderation tools.
Lessons from the Splits
Looking back at three decades of IRC network politics, a few patterns emerge:
Decentralization without governance doesn't scale. EFnet proved that a purely anarchic model works for small communities but creates unsustainable political dynamics as the network grows. Every other network that followed implemented some form of structured governance.
Technical disagreements are often political disagreements in disguise. The EFnet/IRCnet split was ostensibly about timestamp implementations, but the real conflict was about who had the authority to make protocol decisions for the network. The technical question was just the battlefield.
Services are inevitable. IRCnet held out the longest against NickServ/ChanServ-style services, but even they eventually implemented some basic nickname enforcement. Users want identity persistence, and if the network doesn't provide it, they'll implement it themselves with bots — which is worse for everyone.
Fragmentation has a cost. Every split reduced the total value of IRC by dividing the user base. A hundred networks with ten thousand users each is worth less than one network with a million users. The splits may have been necessary, but each one weakened IRC's position against future competitors.
The netsplit wars are over now — not because anyone won, but because the stakes have changed. IRC's remaining networks have found their niches and coexist peacefully. The battles that matter now are between IRC and the proprietary platforms that learned from its strengths while (mostly) avoiding its governance mistakes.
Frequently Asked Questions
What was the biggest IRC network ever?
QuakeNet holds the record for peak simultaneous users on a single network, reaching approximately 240,000 users around 2005. EFnet and IRCnet were also extremely large in the late 1990s and early 2000s, each hosting over 100,000 simultaneous users at their peaks.
Why did EFnet and IRCnet split?
The split in July 1996 was primarily over how to handle nickname and channel collisions during netsplits. European administrators wanted nick delay and channel delay features, while American administrators preferred a different timestamp-based approach. Neither side compromised, and the European servers disconnected to form IRCnet.
Do netsplits still happen?
Yes, temporary netsplits still occur on IRC networks when server-to-server links fail. However, modern IRC servers are more resilient, and the large-scale political splits of the 1990s are no longer common. Networks have established governance structures that resolve disputes before they escalate to full splits.
Which IRC networks still exist from the split era?
EFnet, IRCnet, Undernet, DALnet, and QuakeNet all still exist and operate, though with significantly reduced user counts compared to their peaks. Libera.Chat (successor to Freenode) and OFTC are among the more active networks today.
FAQ
What is an IRC netsplit?
A netsplit occurs when the connection between two IRC servers breaks, dividing the network into isolated segments. Users on different sides cannot see each other until the servers reconnect.
Why were netsplits used for channel takeovers?
During a netsplit, attackers could create an empty version of a channel on their side. When servers reconnected, the attacker's operator status could override the legitimate ops, depending on the timestamp resolution.
How did IRC networks solve netsplit abuse?
Networks implemented timestamp-based channel protection (TS6 protocol), services like ChanServ to restore ops, and server-side channel modes that prevent desyncs during reconnection.