[Advanced] — proof-of-work, proof-of-stake, BFT, and the honest version of what “finality” and “51% attack” mean.

The kid version first

Thousands of computers all want to add the next page to the shared notebook. Only one page can be next. How do they pick, when anyone can join, nobody’s in charge, and some of them are lying?

Every answer has the same shape: make trying to write cost something real, and make it so the cheapest way to profit is to follow the rules. Proof-of-work burns electricity. Proof-of-stake risks money. The mechanism differs; the logic doesn’t.

What consensus must actually deliver

Three properties, and they’re in tension:

  1. Safety — no two honest nodes accept conflicting histories. “Nothing bad happens”
  2. Liveness — the chain keeps producing blocks. “Something good eventually happens”
  3. Sybil resistance — influence can’t be bought by creating identities

FLP says that in a fully asynchronous network, no deterministic protocol can guarantee consensus terminates if even one process may fail — so safety and liveness cannot both be guaranteed unconditionally. Every real design picks which one to sacrifice under partition — and that choice is the cleanest way to classify chains:

Under a network partitionBehaviourExamples
Favour livenessKeeps producing blocks; may temporarily fork, resolves later. Finality is probabilisticBitcoin, Ethereum’s fork-choice layer
Favour safetyHalts rather than risk a conflicting commit. Finality is instant and absoluteTendermint/Cosmos, and Ethereum’s finality layer

Ethereum post-Merge is a hybrid running both — a fork-choice rule that keeps producing, plus a finality gadget that periodically stamps blocks as irreversible. That combination is the main thing to understand about modern PoS.

Proof-of-Work

Find a nonce such that hash(block_header ‖ nonce) < target. There is no method but guessing, so the only strategy is to guess fast. Difficulty auto-adjusts to hold a target block time (Bitcoin: 10 minutes, retargeting every 2016 blocks).

Why it works: you can’t fake having done the work, and you can’t do it cheaply. Influence tracks hashrate, and hashrate costs hardware and electricity — things Sybil identities can’t conjure.

Fork choice: the valid chain with the most cumulative work wins. Not the longest — cumulative difficulty, which matters when difficulty changes.

Finality is probabilistic. Reversing k blocks requires out-working the network across all of them. Each confirmation makes reversal exponentially more expensive, never impossible. Six confirmations is convention, not a proof.

The honest costs: enormous energy use, whose magnitude is the security budget and can’t be optimised away; centralisation into pools and cheap-power regions; and hardware that must keep running forever to keep the chain secure → energy.

Proof-of-Stake

Replace “burn electricity” with “lock up capital.” Validators deposit the native asset (32 ETH on Ethereum); the protocol pseudo-randomly selects proposers and committees; misbehaviour is punished by slashing — destroying part of the deposit.

The key improvement over PoW is not energy. It’s that the punishment is in-protocol. A PoW attacker who fails loses only electricity and keeps their hardware; a PoS attacker who equivocates has their capital destroyed by the very chain they attacked. Attacks become directly, automatically expensive rather than merely wasteful.

Nothing-at-stake, solved: the early objection was that signing every competing fork costs a validator nothing. Slashing for equivocation makes it cost everything. That is the whole answer, and it’s why slashing is load-bearing rather than a nicety.

Ethereum’s structure specifically (post-Merge, 2022):

  • Time is divided into 12-second slots and 32-slot epochs
  • One validator proposes per slot; a committee attests
  • Attestations accumulate; when two-thirds of stake attests across two consecutive epochs, blocks are finalised — reversal would require destroying at least a third of all staked ETH. That is finality with a price tag attached, and the price is public
  • Fork choice between finality checkpoints is LMD-GHOST, which follows the subtree with the most attestation weight rather than simply the longest chain

The honest costs: wealth concentration compounds (stake earns stake); liquid-staking providers re-centralise what the design decentralised — Lido’s share of staked ETH has been a live governance concern for years; and validators must stay online or bleed value, which pushes them to the same handful of cloud providers.

Classical BFT — the third family

PBFT and descendants (Tendermint) run explicit voting rounds: propose → pre-vote → pre-commit, tolerating up to f faults among 3f+1 validators. Commit means committed — instant, absolute finality, no reorgs ever.

The trade: you must know the validator set, and message complexity is O(n²), so it doesn’t scale past low hundreds of validators. If the set can’t reach two-thirds, the chain stops — Cosmos chains have halted in production, by design rather than by failure. That’s safety chosen over liveness, explicitly.

Comparison

PoWPoS (Ethereum)BFT (Tendermint)
Sybil resistanceHashrateStaked capitalKnown validator set
FinalityProbabilistic~13 min, economicInstant, absolute
Under partitionKeeps going, forksKeeps going, stops finalisingHalts
Validator countUnbounded~1M (Ethereum, 2026)Low hundreds
Attack costRent hashrateBuy + lose stakeCorrupt ⅓ of a known set
EnergyEnormousNegligibleNegligible

The 51% attack, precisely

Majority control does not let an attacker steal coins or forge signatures — no amount of hashrate breaks ECDSA. What it lets them do is:

  • Reorder and censor transactions
  • Double-spend their own coins by publishing a longer chain that omits a transaction they already got paid for

That’s the entire threat, and it’s enough — it has repeatedly destroyed smaller chains. Ethereum Classic was 51%-attacked three times in 2020, with reorgs thousands of blocks deep. Security scales with the value securing the chain, so small chains are cheap to attack and rentable hashpower makes it a service you can buy.

Key insight

Consensus mechanisms are not competing for correctness — they’re competing on where they put the cost. PoW puts it outside the system (energy) so attackers keep their capital when they fail. PoS puts it inside (slashable stake) so failure is punished by the protocol itself. Read every consensus design by asking what does an attack cost, who does it cost, and is the punishment enforceable in-protocol — that single question separates the serious designs from the rest.

Source: [reference] — Aug 2026.

↑ Back to Web3 Foundations