How decentralized networks changed year by year

The history of decentralized networks is a sequence of five dated shifts, from Usenet in 1980 to Threads in 2024. Each one answered the problem the previous shift created, and each answer moved work and risk onto the member.

What to take away

1980: Usenet and the shared message format

Usenet began in 1980, written at Duke University and the University of North Carolina by Tom Truscott, Jim Ellis and Steve Bellovin. UUCP passed its articles between machines over phone lines. Two sites agreed on a message format and a delivery rule, and anyone could implement both.

Reachability is worth more than a headcount suggests. That is the observation behind Metcalfe's law, and why interoperation was worth the trouble from the start.

NNTP, published as RFC 977 in 1986, carried the same articles over TCP/IP and made joining cheaper.

The cost arrived with the benefit. A system anyone can join is a system anyone can join, and the first unwanted bulk messages followed within a few years.

1990s: the open relay era and the first refusals

Email's SMTP protocol, standardized as RFC 821 in 1982, let any server relay mail for any other. That openness made the network work and made it abusable. Spam volume grew through the 1990s until operators began refusing traffic from servers they did not recognize.

The first mass commercial flood came in April 1994, when two lawyers posted an advertisement for green card services across thousands of Usenet groups. Paul Vixie started the MAPS Realtime Blackhole List in 1997, a shared list of servers to refuse.

Every answer to unwanted delivery is some form of selective refusal: refuse this sender, refuse this server, refuse anyone not already known. Federated systems answer at the server level because that is the only level with enough information to act.

Once refusal exists, so does politics about who refuses whom. The notes on addressing models trace the same pressure through private messaging, where it produced walls instead of blocklists.

Sender authentication stacked up after that: SPF in 2003, DomainKeys in 2004 and DMARC in 2012. Each one asked the domain owner to publish a record, which moved work onto the person who held the domain.

2008: StatusNet, identi.ca and portable identity

StatusNet, an open-source microblogging server, launched in 2008 under the name Laconica. Evan Prodromou opened its best-known instance, identi.ca, in July 2008, and it grew to hundreds of thousands of accounts by 2009.

StatusNet supported OpenID, which let a member's identity resolve to a domain they controlled rather than to the server holding their account. The domain name system resolves that identity to whichever server currently serves it.

The burden moved to the member. You had to own and renew a domain, which is a bill and a skill, and most people did not. A .com registration typically costs $10 to $20 a year at registrars such as Cloudflare Registrar, Namecheap or Porkbun, as of 2026.

2010 to 2016: Diaspora, GNU social and the identity split

Diaspora ran a Kickstarter in 2010 and raised over 200,000 dollars from more than 6,000 backers. It promised portable identity across pods and delivered a partial version of it. GNU social, the renamed StatusNet codebase, carried the older model forward.

GNU social took its name in 2013, when StatusNet merged with Free Social. Friendica had started in 2010 with its own account portability, Misskey arrived from Japan in 2014, and Hubzilla split off in 2015.

Three identity answers emerged in this period and all three survive:

Each answer moves a burden onto the member. Problems get solved by transferring work and risk from operators to individuals.

2017 to 2019: Mastodon and the blocklist wars

Mastodon launched in 2016 and grew sharply in April 2017. By 2019, server-level blocking was the default answer to abuse, and the fights over it were public.

Blocking a whole server is the only instrument that scales, and it is far too blunt: it removes thousands of people who did nothing to stop one. Shared blocklists let operators pool work, but one operator's judgment propagates to servers that never reviewed it.

Client-side labeling moves the decision to the reader, which respects autonomy but asks every reader to become their own moderator. None of these has settled, which is why removal power is one of the three questions that classify a network in the mechanism-based classification.

The same years added federated video and link sharing. PeerTube shipped from Framasoft in 2018, Pixelfed followed the same year, and Lemmy arrived in 2019.

2018: ActivityPub becomes a W3C Recommendation

The W3C published ActivityPub as a Recommendation on 23 January 2018. It standardized the protocol layer that Mastodon, Pleroma and later Threads and Flipboard use to exchange messages.

The Social Web Working Group started the specification in 2014. ActivityPub replaced OStatus, the Atom-based format that identi.ca and GNU social used to pass posts between servers.

Standardization did not solve abuse, identity or discovery. It made the first problem cheap to answer, which freed every operator to spend their time on the other four.

2020 to 2023: Bluesky, AT Protocol and the discovery gap

Bluesky began as a Twitter-funded project in 2019, incorporated in 2021, and opened its AT Protocol to federation in 2023. Its answer to identity is a domain handle, the same answer StatusNet tried in 2008.

Nostr arrived in 2020 with the opposite answer: identity is a key pair, and any relay carries your messages. The Damus app put that model on the iPhone App Store in early 2023.

Discovery remains the honest cost of the whole arrangement. A central service can index everything it holds.

A federated network holds nothing centrally, so a member's search covers whatever their own server received. The pillar notes on what decentralisation changes and what it does not state this plainly.

Partial indexes, opt-in discovery and hand-made directories each work. PeerTube's Sepia Search, run by Framasoft, covers only the videos its crawler is allowed to see.

FediDB and Fediverse Observer count instances and software rather than posts. None adds up to a search that answers where the people you care about are.

2024 to 2026: Threads, regulation and the liveness question

Meta's Threads opened limited ActivityPub federation in 2024, which put a centralized operator inside the federated address space. Flipboard federated its magazines the same year, and WordPress sites can join with the ActivityPub plugin that Automattic acquired in 2024.

Ghost began rolling out ActivityPub support in 2025. Managed hosting for a small instance typically costs $5 to $25 a month from providers such as Masto.host or Cloud68, while self-hosting costs a server and your own time.

The Online News Act and the CLOUD Act both bear on where Canadian user data sits and who can compel it, and the Office of the Privacy Commissioner is the authority for Canadian privacy complaints.

User counts across these networks are disputed. Mastodon's own instance list, third-party trackers and press figures disagree, and any number should carry its source and date.

The general argument for reading a category by its constraints rather than by its calendar applies here more strongly than anywhere else on this site.

Reading the sequence for a network in front of you

Ask which shifts it has answered and which it has inherited, in order.

A network that has answered the first three and not the last two is not behind. It is where the field is.

Common questions

Is the sequence the same for every federated network?

The order is, because each problem is created by the previous answer. The point a given network has reached differs, and so does which answer it picked at each step. Two networks at the same point with different answers behave very differently.

Does a newer answer replace an older one?

Rarely. They accumulate. A network can run server-level refusal, shared lists and client-side labeling at once, applying to the same message in sequence. When working out why something was not delivered, assume several mechanisms rather than one.

Where do bridges to centralized services fit?

They are an answer to the first problem, attempted late and from the wrong side. They inherit every later problem at once: abuse crosses the bridge, identity does not, and moderation authority stops at it. Bridges are useful and they are the most fragile thing in this category.

Which shift should a reader care about most?

Whichever one is unsolved in the network they are considering joining. For most people that is discovery, because it decides whether the place will feel empty. No amount of good design elsewhere compensates for a room you cannot find anyone in.