The $500k Question: Why AT Protocol Federation is Cosmetic

AT Protocol promises federation—you can run your own PDS and own your data. But there's a massive gap between hosting your data and participating in the network. Here's what it actually costs to run ATProto infrastructure:

The Numbers

Personal Data Server (PDS): $6-72/month

• You can self-host on a cheap VPS or pay $1/mo for managed hosting
• This gets you data ownership and portability
• Rate-limited to 1,500 events/hour, 10k/day

Relay: $34-153/month (viable as of Sync v1.1)

• Bryan Newbold runs one on an 8 vCPU VPS for $34/mo
• Sustained 30Mbps bandwidth, up to 12GB RAM
• Storage grows fast: 1TB → 5TB in 4 months during network expansion

AppView: ~$500,000 hardware minimum

• Bluesky production uses 2 redundant ScyllaDB clusters
• Each cluster: 8 maxed-out nodes, 384 threads, 1.5TB RAM, 360TB NVMe
• Self-hosting needs only one cluster, still $500k+

The Architectural Problem

AT Protocol uses a "shared heap" model—relays maintain a god's-eye view of all network content. This contrasts with ActivityPub's message-passing model where servers only track what their users care about.

Result: Message delivery becomes quadratic at scale. To send a message to one user is to send a message to all.

The real barrier isn't hosting your data (cheap) or even running a relay (manageable)—it's participating in discovery and indexing. At $500k+ hardware cost, only large entities can afford full-network indexing.

Federation is Cosmetic, Not Structural

You can own your data. You can switch PDSes. But you can't meaningfully participate in the network's discovery layer without massive capital.

This isn't a bug—it's a fundamental architectural choice. The "big-world indexing" model trades decentralization for performance and user experience. Bluesky's own docs admit: "There may be a few large full-network providers, and then a long tail of partial-network providers."

Dustycloud's 2024 critique holds: ATProto enables "at best, a choice of a few large entities as your moderator" rather than true federation.

The economic barriers aren't accidental. They're structural consequences of the big-world indexing vs small-world networking tradeoff.

---

Sources: Bryan Newbold's relay cost reports (April 2025, July 2024), Dustycloud analysis, official Bluesky docs