How do blockchains reach consensus? Proof of work, proof of stake and what finality really means

Who this is for: Treasurers, fund operations teams, founders and policy staff who need to decide how long to wait before treating a blockchain transfer as settled, and who want to compare proof of work, proof of stake and BFT chains on security, decentralisation and energy using real numbers.

Every payment system needs one answer to “who owns what right now”. Banks get it from a ledger kept by one trusted party and a legal rule on when a payment is final. Blockchains get it with no trusted party: thousands of independent machines run the same rules, and a consensus protocol decides which of many possible histories counts. The hard part is not agreeing when everyone is honest, but when some participants lie, go offline, or try to spend the same coin twice.

The stakes are higher in 2026 than when Bitcoin launched. Tokenised US Treasury products held $14.8 billion as of 6 October 2026 according to rwa.xyz, and the BIS Annual Economic Report of June 2025 made settlement finality the central test for any tokenised money system. An institution that treats a transfer as settled before the chain does carries an unpriced risk; one that waits longer than necessary leaves capital idle.

This guide covers how proof of work and the main proof of stake families work, what finality means on each, where decentralisation and energy numbers stand, which attacks have actually happened, how MEV reshaped block production, and what Ethereum and Solana are shipping in late 2026 and 2027, ending with a worked example sizing the finality wait for a $10 million tokenised bond.

Blockchain consensus by the numbers

The problem: double spends and Byzantine faults

A digital coin is data, and data can be copied. If Alice sends the same coin to Bob and to Carol, both see a valid signature, and without a central ledger there is no way to decide which spend came first. This is the double spend problem.

The deeper issue is Byzantine fault tolerance, named after a 1982 paper by Lamport, Shostak and Pease about generals coordinating an attack while some send contradictory messages. A Byzantine participant does not merely crash; it behaves arbitrarily. The classical result is that safety holds only while fewer than one third of participants are faulty, and every proof of stake protocol in use today inherits that bound.

Classical BFT protocols assumed a known, fixed set of participants, but on an open network anyone can create unlimited identities. Bitcoin made votes cost computation; proof of stake makes them cost locked capital. Both make an open network behave enough like a closed one for the classical maths to apply.

Nakamoto consensus and probabilistic finality

Miners bundle transactions into a block and search for a number that makes the block’s hash fall below a difficulty target. Finding one takes about ten minutes network-wide, and difficulty adjusts every 2,016 blocks to keep it there. A valid block pays the miner new bitcoin (3.125 BTC since the April 2024 halving) plus fees. Nodes follow the chain with the most accumulated work, and that one rule resolves disputes: if two miners find blocks at once, whichever branch gets the next block wins and the other is orphaned.

Why there is no final block, only a less likely reversal

Because the heaviest chain always wins, no Bitcoin block is strictly final: an attacker who builds a heavier private chain can replace public blocks. What protects a transaction is that each block on top makes replacement exponentially harder. Satoshi Nakamoto’s whitepaper (October 2008) works this out in section 11: the chance that an attacker with share q of hashpower catches up from z blocks behind falls as (q/p)^z, where p is the honest share. With q at 10 percent, the chance of overturning a transaction five blocks deep is 0.09 percent, and after ten blocks 0.00012 percent; with q at 30 percent it takes 24 blocks to get below 0.1 percent. The familiar “six confirmations” comes from that table (about 0.024 percent against a 10 percent adversary) and is a convention, not a rule: exchanges raise it when a chain is under attack, as Kraken did when it moved Monero to 720 confirmations in August 2025.

What makes the attacker’s share expensive is the network’s size. Hashrate stood at about 968 EH/s with a difficulty of 132.7 trillion on 6 October 2026 per mempool.space, after first touching one zettahash in 2025 according to The Block. Even 10 percent means sourcing about 97 EH/s of specialised hardware and the power to run it, against miners who earned about $17.2 billion in 2025. Proof of work security is electricity and chips converted into a probability.

Proof of stake: the main families

Proof of stake replaces hashpower with locked capital. Validators deposit tokens, propose and vote on blocks in proportion to stake, earn rewards for honest behaviour, and lose deposit (slashing) for provable misbehaviour such as signing two conflicting blocks. For the staking side, see what crypto staking is. Four designs secure most of the industry’s value.

Ethereum: Gasper (Casper FFG plus LMD-GHOST)

Ethereum’s consensus, live since the Merge on 15 September 2022, has two parts. LMD-GHOST is the fork choice rule: follow the branch with the most recent stake-weighted attestations. Casper FFG is the finality gadget on top. Time runs in 12-second slots and 32-slot epochs (6.4 minutes). At each epoch boundary validators vote on a checkpoint; when two thirds of staked ETH backs it the checkpoint is “justified”, and when the next checkpoint is also justified the earlier one is “finalised”. Reverting a finalised block requires a third of total stake to sign conflicting votes, a slashable offence: with 43.7 million ETH staked in October 2026, roughly 14.6 million ETH destroyed. Normal finality takes two epochs, 12.8 minutes, in practice 13 to 19 minutes depending on where in an epoch a transaction lands. If finality stalls for more than four epochs, an “inactivity leak” drains non-voting validators’ stake until the active set again holds two thirds.

Tendermint and CometBFT: instant finality, smaller validator sets

Tendermint, described by Jae Kwon in 2014 and maintained today as CometBFT, powers the Cosmos Hub and over a hundred app-chains. Each height goes through propose, prevote and precommit rounds, and when more than two thirds of voting power precommits, the block is final within seconds. There are no reorgs by design; if more than a third of validators go offline, the chain halts rather than forking. Because every validator must talk to every other each round, these chains typically run 100 to 200 validators.

Solana: Tower BFT today, Alpenglow next

Solana’s current consensus, Tower BFT, uses Proof of History as a verifiable clock and has validators cast votes as on-chain transactions with lockouts that grow as they stack votes. A transaction is “confirmed” once a supermajority has voted, typically within a second or two, and “finalized” after 31 further blocks, about 12.8 seconds (see what is Solana).

Alpenglow, unveiled by Anza in May 2025 and approved as SIMD-0326 in September 2025 with 98.27 percent of participating stake voting yes, replaces Tower BFT. Its voting component, Votor, exchanges signed votes off-chain: if 80 percent of stake approves in one round the block is final (about 100 ms); if 60 to 80 percent approves, a second round at 60 percent finalises it (about 150 ms). It tolerates 20 percent malicious stake plus 20 percent offline. On the public test cluster, 96 percent of blocks took the fast path at an average slot time of 214 ms as of August 2026 per Solana Compass. Testnet deployment began on 23 September 2026 (see our report). It did not activate on mainnet on 28 September despite calendar entries saying so; Anza research head Roger Wattenhofer said there was “no Alpenrush”, and the next tentative activation window in the Agave v4.4 schedule is 9 November 2026. The second phase, Rotor, replaces Turbine propagation and is scheduled separately.

Avalanche: Snow family, repeated sampling

Avalanche’s Snowman protocol reaches agreement by repeated sampling: each node asks a random stake-weighted sample of 20 validators which block it prefers, adopts that preference when 15 or more agree, and treats the decision as final after 20 consecutive successful polls. Finality arrives in one to two seconds and message overhead does not grow with validator count, which is how the primary network sustains more than a thousand validators.

Finality times compared

“Practical finality” below is the point at which a cautious operator should treat a transfer as irreversible under the protocol’s own assumptions.

Speed and decentralisation are in tension: chains that finalise in a second or less run far fewer independent validators than Ethereum or Bitcoin, and move finality risk from the protocol to the stake distribution.

Validator economics and decentralisation metrics

Stake concentration

The Nakamoto coefficient is the smallest number of entities that together control enough of a network to break it: 51 percent of hashpower, or 33 percent of stake on most proof of stake chains. Trackers compute it differently, so compare within one source. Nakaflow’s 6 October 2026 readings put Avalanche at 23, Sui at 18, Cardano at 15, Cosmos Hub at 3 and Hyperliquid at 3; Solana Compass puts Solana at 18 across 669 staked validators, up from 12 two years earlier. A single-digit coefficient means a handful of exchanges or staking providers could halt the chain.

Ethereum’s concentration is measured by operator. Lido, the largest liquid staking protocol, peaked at 32.5 percent of staked ETH in May 2022 and had fallen to 20.7 percent (8.6 million ETH) by June 2026 according to Four Pillars. Coinbase reported 4.5 million ETH, 12.17 percent of staked ETH, in its Q1 2026 validator report. The informal limit is that no operator should pass one third, the point at which it could block finality; the two largest together still approached that line in 2026. How correlated large operator sets are showed on 30 September 2026, when a security disclosure prompted about 523,000 ETH across 17,000 MetaMask-linked Lido validators to begin exiting, stretching the exit queue to 13 days (see our coverage).

Client diversity

Failures can come from software, not stake. If a bug in one client leads a majority of validators to sign an invalid block, the chain can finalise something wrong, and reverting it costs the stake of everyone who signed. The safe threshold is that no client runs more than a third of validators. As of April 2026, stakefish put consensus-layer shares at Lighthouse about 43 percent, Prysm about 31 percent and Teku about 14 percent, and execution-layer shares at Geth about 50 percent, Nethermind about 25 percent, Besu about 10 percent and Reth about 8 percent. Both layers still had one client above the line; the seven-block Beacon Chain reorg of 25 May 2022, caused by an uneven client rollout of a fork choice change, shows why that matters.

What validators earn

Ethereum’s staking yield was about 2.6 percent in October 2026 per validatorqueue.com. Solana validators pay about 1.1 SOL a day in on-chain vote fees; Alpenglow moves votes off-chain and replaces that with a Validator Admission Ticket of about 0.8 SOL a day under SIMD-0357, activated on 22 July 2026. Bitcoin miners now earn about 99 percent of revenue from the block subsidy, with fees down from roughly 7 percent to about 1 percent over 2025 per The Block, which makes the security budget a question at every halving.

Energy use: proof of work versus proof of stake

The Cambridge Centre for Alternative Finance’s April 2025 Digital Mining Industry Report estimated Bitcoin’s electricity use at 138 TWh a year, about 0.5 percent of global consumption, with 52.4 percent from sustainable sources (42.6 percent renewables, 9.8 percent nuclear) and emissions of 39.8 million tonnes of CO2 equivalent, from a survey covering 48 percent of hashrate. Hashrate has grown since, so the 2026 figure is likely higher. The energy is the security: the only way to attack is to spend comparable energy.

Proof of stake relies on capital at risk instead. Ethereum.org estimates the post-Merge network at about 0.0026 TWh a year, a 99.988 percent reduction from the roughly 21 TWh used before September 2022, which is why allocators with ESG mandates have found staking easier to approve than mining exposure. It is not a security argument in either direction: both models have held for years at scale and fail in different ways.

Attacks and reorgs that actually happened

A 51 percent attack does not let an attacker steal from arbitrary addresses or mint tokens; signatures must still be valid. It allows double spending: deposit coins at an exchange, withdraw something else, then publish a heavier private chain in which the deposit never happened. Every real case has hit a small proof of work chain whose hashrate could be rented or redirected from a larger chain with the same algorithm.

1. Bitcoin Gold, May 2018. About 388,000 BTG, roughly $18 million at the time, was double spent against exchanges; a second attack in January 2020 took about $72,000.
2. Ethereum Classic, January 2019 and August 2020. About $1.1 million was double spent in January 2019; two attacks within a week in August 2020 caused estimated losses of $5.6 million and $1.68 million, prompting the ETChash algorithm change that November.
3. Ethereum Beacon Chain, 25 May 2022. A seven-block reorg occurred before the Merge, caused by uneven client rollout of proposer boost rather than an attack; no finalised block was affected.
4. Monero, August and September 2025. The Qubic project paid miners to point hashrate at Monero, growing from under 2 percent in May 2025 to a claimed 51 percent by 12 August 2025, when a six-block reorg occurred. On 15 September 2025 the chain suffered its deepest reorg in 12 years: 18 blocks, 36 minutes and 118 confirmed transactions invalidated. Kraken paused deposits, then required 720 confirmations.

No comparable attack has succeeded on Bitcoin or Ethereum mainnet. The lesson for anyone accepting deposits is that confirmation policy should scale with how much of a chain’s hashrate or stake is rentable, and be revisited when that changes.

MEV and proposer-builder separation

Consensus decides who proposes a block, not what goes in it. Since DeFi took off, transaction ordering within a block has become valuable in itself: arbitrage, liquidations and front-running make up maximal extractable value (MEV). Ethereum’s response was proposer-builder separation (PBS): specialised builders assemble the most profitable block and bid for it through relays, and the proposing validator signs the highest bid without seeing its contents. Flashbots launched MEV-Boost to do this out of protocol in 2022, and most Ethereum blocks have flowed through it since.

The result is efficient but concentrated. Relayscan data for the 24 hours to 6 October 2026 shows Titan building 53.35 percent of MEV-Boost blocks, Quasar 20.63 percent and BuilderNet 16.60 percent, with the two largest relays handling two thirds of payloads. Builders do not control consensus, but one with half the blocks has large influence over which transactions are included and when. Ethereum’s answer is to bring PBS into the protocol (EIP-7732) so relays are no longer trusted intermediaries, and to let validators force inclusion of transactions builders leave out (FOCIL, now slated for the Hegota fork). Solana handles the same problem through the Jito client’s off-chain bundle auction.

Ethereum’s 2026 roadmap: Glamsterdam and faster finality

Ethereum shipped Pectra on 7 May 2025 and Fusaka on 3 December 2025, which introduced PeerDAS so nodes verify blob data by sampling. Those upgrades scale layer 2 data capacity but did not change finality.

Glamsterdam is the consensus-relevant fork. Its headline EIPs are EIP-7732 (enshrined proposer-builder separation) and EIP-7928 (block-level access lists, which declare each transaction’s state footprint upfront to allow parallel execution), with 17 further EIPs. Pencilled for the first half of 2026, it slipped as ePBS proved hard to implement across ten client implementations. The Ethereum Foundation announced on 17 September 2026 that Glamsterdam would activate on Sepolia at epoch 353,024 on 6 October 2026, with Hoodi and mainnet dates “TBD”. Mainnet in late 2026 or early 2027 is a working assumption, not a commitment.

Faster finality is further out. Single-slot finality would finalise a block in the slot it is proposed, but needs signatures from the whole validator set aggregated within 12 seconds, which ethereum.org still calls a multi-year research item. The more concrete path is the “lean consensus” track Justin Drake set out on 31 July 2025: a redesigned beacon chain with finality in seconds, hash-based post-quantum signatures and a smaller rotating committee, with three-slot finality the favoured intermediate step. None of this is in Glamsterdam or Hegota, so institutions should plan on 13 to 19 minute finality through at least 2027.

How we got here: a timeline

Worked example: settling a $10m tokenised bond

How to evaluate a chain’s consensus: a checklist

• What does finality actually mean here? Probabilistic (Bitcoin), economic with slashing (Ethereum) or classical BFT (Tendermint, Alpenglow). A good answer names the condition for irreversibility and what breaking it costs.
• How long until that condition is met, in the worst normal case? Use the upper bound (19 minutes on Ethereum, not 12.8) when writing operational policy.
• What is the Nakamoto coefficient, and who are those entities? Below 10 means a few exchanges or staking providers could halt the chain. Check which tracker and method produced the figure.
• Is any single client above one third of validators? If so, a client bug can finalise an invalid chain. Ethereum’s execution layer still had Geth near 50 percent in April 2026.
• Can the hashrate or stake be rented? Every successful 51 percent attack hit a chain sharing an algorithm with a much larger one. For proof of stake, ask how much stake is liquid and borrowable.
• What happens when a third of participants go offline? Tendermint chains halt; Ethereum leaks inactive stake and keeps finalising; Bitcoin slows but continues.
• Who builds the blocks, and can they censor? Builder concentration (Titan above 50 percent of Ethereum MEV-Boost blocks in October 2026) is a censorship question even when consensus is healthy.

Risks and open questions

The first open question is whether proof of stake’s economic finality holds under a real adversary rather than a bug. Ethereum’s slashing has only ever been triggered by operator mistakes, so the claim that a third of stake would be destroyed remains a model result, and if a cartel did finalise a bad chain the community would almost certainly coordinate a manual fork, a “social layer” the protocol maths does not capture.

The second is concentration through convenience. Liquid staking, exchange staking and ETF-held stake pool voting power with a few operators; the September 2026 exit of 523,000 ETH from one operator group showed how correlated large validator sets are, and on faster BFT chains a smaller validator set is the price of speed. The third is proof of work’s security budget as subsidies halve, with fees at about 1 percent of miner revenue in 2025 and the next halving due in 2028. The fourth is upgrade risk: Glamsterdam’s slip and Alpenglow’s phantom 28 September date show that consensus changes are the hardest software in the industry to ship.

What to watch next

Glossary

Byzantine fault

A participant that behaves arbitrarily, including lying or sending conflicting messages, rather than simply crashing. Protocols tolerate fewer than one third such participants.

Double spend

Spending the same coin twice by getting two conflicting transactions accepted, the core problem consensus exists to prevent.

Reorg

A chain reorganisation, when nodes abandon recent blocks for a competing branch, invalidating transactions that were only in the abandoned blocks.

Finality

The point at which a transaction cannot be reverted under the protocol’s security assumptions. Probabilistic on Bitcoin, economic on Ethereum, immediate on BFT chains.

Slashing

Destroying part or all of a validator’s stake for provable misbehaviour such as signing two conflicting blocks.

Nakamoto coefficient

The smallest number of entities whose collusion could break a chain’s consensus, typically a third of stake or a majority of hashrate.

MEV

Maximal extractable value, the profit available from choosing and ordering transactions within a block, such as arbitrage and liquidations.

Proposer-builder separation

Splitting the job of assembling a block (builder) from signing it (proposer), done today via MEV-Boost relays and to be enshrined in Ethereum by EIP-7732.

Why it matters

The settlement assurance a bank gets from a central bank ledger and a legal definition of finality, a blockchain has to earn from mathematics and incentives. The BIS argued in June 2025 that tokenised finance only works if settlement finality is unambiguous; the figures here show it can be, but the answer differs on every chain and shifts as hashrate, stake and software evolve. An hour on Bitcoin, a quarter of an hour on Ethereum and a fraction of a second on Alpenglow-era Solana are each defensible, provided the operator knows which adversary the number assumes.

For tokenised real-world assets and the custodians that hold them, consensus sets the speed at which capital can be reused, the risk to reserve against, and the counterparties, from Lido to Titan, an institution is implicitly trusting. The upgrades landing in late 2026 and 2027 will move the numbers again, so the right habit is to write confirmation policy per chain, source each number, and date it.

Sources

1. Bitcoin.org: Bitcoin: A Peer-to-Peer Electronic Cash System (section 11, Calculations), October 31, 2008
2. Ethereum.org: Gasper, accessed October 6, 2026
3. Ethereum.org: Single slot finality, accessed October 6, 2026
4. Ethereum.org: Energy consumption, updated July 28, 2026
5. Ethereum Foundation: Glamsterdam Testnet Announcement, September 17, 2026
6. Ethereum Foundation: Fusaka Mainnet Announcement, November 6, 2025
7. Ethereum Foundation (Justin Drake): Lean Ethereum, July 31, 2025
8. Solana Foundation: Alpenglow upgrade page, accessed October 6, 2026
9. Avalanche: Avalanche Consensus (Snowman), accessed October 6, 2026
10. BIS: Annual Economic Report 2025, Chapter III, June 24, 2025
11. Cambridge Judge Business School: Sustainable energy rising in Bitcoin mining, April 28, 2025
12. mempool.space: Mining hashrate and difficulty API, October 6, 2026
13. validatorqueue.com: Ethereum validator queue and staking statistics, October 6, 2026
14. Relayscan: MEV-Boost relay and builder statistics, October 6, 2026
15. Nakaflow: Nakamoto coefficient tracker, October 6, 2026
16. Solana Compass: Solana decentralization statistics, accessed October 6, 2026
17. rwa.xyz: Tokenized Treasuries, October 6, 2026
18. Four Pillars: Lido’s share of staked Ethereum has fallen by more than ten percentage points from its peak, June 2026
19. KuCoin News: Coinbase Q1 2026 validator report, 4.5m ETH staked, 12.17 percent of network, May 13, 2026
20. stakefish: The State of Ethereum in 2026, April 21, 2026
21. Solana Compass: What Solana’s Alpenglow upgrade changes, August 9, 2026
22. FinanceFeeds: Solana’s Alpenglow did not go live on 28 September, September 30, 2026
23. The Block: 2026 Bitcoin Mining Outlook, December 30, 2025
24. FXStreet: Monero suffers deepest-ever blockchain reorganization, invalidating 118 transactions, September 15, 2025
25. The Block: Monero faces reorg fears after Qubic claims 51 percent of hashrate, August 12, 2025
26. Cointelegraph: 51 percent attack on Monero prompts proposals to overhaul consensus, August 20, 2025
27. Barnabé Monnot: The 7-block reorg on the Beacon Chain, May 29, 2022
28. Wikipedia: Ethereum Classic, accessed October 6, 2026
29. Wikipedia: Bitcoin Gold, accessed October 6, 2026

Disclosure: This guide is for education only and is not investment, legal or tax advice.