Overview
The blockchain trilemma, popularized by Ethereum's Vitalik Buterin, states that a blockchain can only optimize for two of three properties: scalability, security, and decentralization.
Most chains make explicit tradeoffs: Ethereum prioritizes security and decentralization (low throughput). Solana prioritizes scalability and security (high hardware requirements reduce decentralization). Rollups prioritize scalability and decentralization (sequencer centralization challenges).
Cerberus's Approach, and what has actually been tested
Radix's Cerberus consensus argues the trilemma is a limitation of specific consensus designs rather than a fundamental law. The Cerberus whitepaper proposes to get there by braiding consensus across shards, so that a transaction involves only the shards it touches, which would give:
- Scalability – throughput scaling with shard count rather than with per-node capacity
- Security – full BFT security per shard, with braiding supplying the cross-shard guarantee
- Decentralization – commodity hardware nodes, with capacity added by adding nodes
That is a design, not a deployment. Braiding has never run in production: not on Radix Babylon, which runs Cerberus unsharded and so has nothing to braid, and not in the Radix Foundation's Hyperscale reference implementation either. The trilemma claim above should therefore be read as the protocol's argument rather than as a demonstrated result.
The Hyperscale 500k TPS test is frequently offered as that demonstration, and it is not one. It sustained over 500,000 TPS and peaked above 700,000 on commodity AWS instances, with more than 590 community nodes in the public phase, but what it measured was the Foundation's Hyperscale implementation, which the lead developer of the Xi'an production candidate states "never really used Cerberus". The figures are a property of the software that was tested, not evidence about the protocol specified in the whitepaper. A sharded network that answers the trilemma on Radix is still forthcoming work, and the current candidate is hyperscale-rs, whose per-shard consensus is HotStuff-2 derived rather than Cerberus.