Title: P2P at 100B Nodes: Real Scaling Limits
They list six hard requirements that no existing protocol seems to satisfy together:
- Anonymous unique auth — one-user-one-node without storing personal data or relying on certificate authorities/trusted third parties.
- Global, growing population coverage — 8B+ users today, 100M+ annual growth, plus deceased users (so growth is strictly monotonic).
- Strictly P2P — no supernodes; each user may run multiple devices, which fail, get lost, or go obsolete, pushing the network into the hundreds of billions.
- Cross-device account access — no passwords, tokens, seed phrases, or similar credentials.
- Sub-second lookup — any device locates any peer/data in under 1 second, with per-node routing state ideally under 100 KB.
- High replication + consensus recovery — ~100–1000x replication, verified recovery without central arbiters, and background sync under a few hundred bytes/sec per device.
They've reviewed Kademlia, Chord, BATMAN, GNUnet, I2P, and others, and conclude all of them break down on scalability or traffic efficiency at this scale.
Their own writeup is on GitHub under the name ikhrabry-spec — "A P2P Network Architecture for 100B Nodes." A patent is granted and applications are filed, but they're explicitly doubting whether their approach is truly novel and want pointers to anything they missed.
So the real question for practitioners: is 100B-node P2P with sub-second, low-state routing actually solvable today, or does this hit a fundamental scaling wall that DHTs can't cross?
For context, even well-regarded systems like Kademlia start showing measurable routing-table bloat and lookup latency once you scale beyond a few million nodes without aggressive pruning or supernode fallbacks — which violates their "no distinguished nodes" rule.
If anyone has tackled this kind of scale without trusted infrastructure, the community would want to see the approach.