The Real Meaning of Decentralization

Decentralization describes how widely control is spread, not how many machines run the software. A system is decentralized to the degree that no single person or group can change its rules, block transactions or shut it down, and that has to be checked layer by layer rather than accepted as a label.

Key takeaways

Vitalik Buterin's 2017 essay separates architectural (how many machines), political (how many controllers) and logical (one shared state or not) decentralization.
The Nakamoto coefficient is the smallest number of entities that together control a majority of a subsystem. In the month to October 20, 2026, three pools produced 62.9% of Bitcoin blocks, so the mining coefficient is 3.
Node counts say little on their own: in a Bitcoin node snapshot from the same day, 49.9% of the 24,701 reachable nodes were Tor onion addresses, and 15.8% carried a "Knots" suffix in their self-reported user agent.
The answer depends on how entities are grouped. Two Ethereum staking datasets put the one-third coefficient at 3 and at 4, and merging two related entries in one of them gives 2.
Legal definitions differ by rule. The SEC's March 2026 interpretation defines a "decentralized" crypto system as one where no person or group has operational, economic or voting control, and does not require decentralization for a token to be a digital commodity.

Three axes of decentralization

In a February 2017 essay, Vitalik Buterin proposed three separate questions:

Architectural: how many physical computers make up the system, and how many of them can break down at once while it keeps running.
Political: how many individuals or organizations control those computers in the end.
Logical: whether the interface and data structures look more like one monolithic object or an amorphous swarm. His simple heuristic is to cut the system in half, providers and users alike, and see whether both halves keep fully operating as independent units.

Buterin placed blockchains as politically and architecturally decentralized but logically centralized, because they maintain one commonly agreed state. The axes are independent, so "many nodes" (architectural) does not settle who holds the keys that matter (political); the concentration data below tests that political placement rather than assuming it.

The Nakamoto coefficient

A July 2017 essay by Balaji S. Srinivasan and Leland Lee turned the question into a number in three steps: list the essential subsystems (their example uses mining, clients, developers, exchanges, nodes and ownership), find the minimum number of entities needed to reach 51% of each, and take the lowest result as the system's score. A different choice of subsystems changes the result, as the authors note.

In their 2017 snapshot, geth ran 76% of Ethereum clients and Bitcoin Core ran most Bitcoin clients, so compromising one codebase would reach more than 51% of clients. That gave both Bitcoin and Ethereum a minimum coefficient of 1. For Ethereum, the authors named developer or mining concentration as the next bottleneck; two developers accounted for 51% of geth's commits.

Which threshold to use

The essay allows a "modified" coefficient when 51% is not the share that matters.

More than 50% (proof-of-work hash rate, or a simple majority of any subsystem): the original rule in the essay.
More than one-third (proof-of-stake finality): Ethereum needs two-thirds of stake to finalize, so a party above one-third can temporarily inhibit finality, until the protocol's inactivity leak penalizes it enough for the chain to finalize again.

Nakamoto coefficient = smallest k such that
  share(entity 1) + share(entity 2) + ... + share(entity k) > threshold
(entities sorted from largest to smallest share; threshold = 50% or one-third)

The result is only as good as the grouping behind it. Merging entities that share an owner lowers the coefficient; splitting a pool into the miners who point hardware at it raises it. A coefficient is neither a floor nor a ceiling on the coordination an attack would need, and it is not a forecast of one.

Concentration data, 2026

Each figure below is a dated snapshot; live sources will have moved since.

Bitcoin mining pools

Block counts come from the mempool.space mining pools page, which attributes blocks to the pool that published them. Shares for the rolling one-month window (4,488 blocks), viewed at 10:15 UTC on October 20, 2026:

Rank	Pool	Share of blocks	Cumulative
1	Foundry USA	25.4%	25.4%
2	AntPool	21.2%	46.6%
3	F2Pool	16.3%	62.9%
4	ViaBTC	9.5%	72.4%

The cumulative share first passes 50% at the third pool, so the coefficient is 3. The one-week (62.6% after three pools) and three-month (60.5%) windows give the same answer. Pool share measures who assembles and publishes blocks, not who owns the hardware, and miners can switch pools.

Bitcoin nodes and client software

The Bitnodes snapshot taken at 10:00 UTC on October 20, 2026 listed 24,701 reachable nodes. Of these, 12,318 (49.9%) were onion addresses, 11,175 (45.2%) IPv4 and 1,208 (4.9%) IPv6. Bitnodes finds nodes by crawling peers it can connect to, so nodes that do not accept incoming connections are not in this count; it is a floor, not a census.

The user-agent field, which each node sets for itself, began with "Satoshi" for 24,620 nodes (99.7%), and 3,893 (15.8%) also carried a "Knots" suffix. The field is self-declared, so it indicates code lineage rather than proving it, but it shows why client software counts as its own subsystem.

Ethereum staking

Lido's H1 2026 report lists the top staking entities by share of all staked ETH as of June 30, 2026. Lido is one of the entities measured, so treat its table as one attribution among several. Running totals are computed from the listed shares:

Rank	Entity	Share of staked ETH	Cumulative
1	Lido	21.2%	21.2%
2	Bitmine	11.5%	32.7%
3	Coinbase	10.9%	43.6%
4	Binance	7.9%	51.5%
5	Ether.fi	4.1%	55.6%

On this grouping the cumulative share first exceeds one-third at the third entity and 50% at the fourth, so the coefficients are 3 and 4. The report lists "Grayscale (via Coinbase)" separately at 3.1%. Counted with Coinbase (14.0%), the top two would hold 35.2% and the one-third coefficient would drop to 2, while the 50% coefficient stays at 4 (54.6%).

The hildobby staking dashboard on Dune, viewed on October 20, 2026 (no as-of date shown; latest data points October 16), lists Coinbase at 3.25%, has no Bitmine entry in its top ten and leaves about 39% of stake unidentified. On its shares, one-third is passed at the fourth entity (36.59%) and 50% only at the tenth (50.45%). The two datasets differ in date as well as in how they attribute stake, so the coefficient you get depends on which source you pick and how it groups entities.

Counting Lido as one entity also understates the coordination an attack through it would take: GSR's 2023 analysis said it would require collusion among Lido's 37 node operators at the time, and called its token-holder governance the larger concern.

Layers beyond consensus

Consensus is only one place where control can sit.

Layer	Control question	Example evidence
Consensus	How many entities produce blocks or hold stake?	Pool shares; staking entity shares
Nodes	Can independent parties verify the chain?	Bitnodes counts and network mix
Client software	Does one codebase run most of the network?	User-agent shares; Ethereum client-diversity pages
Development and governance	Who can merge changes or push upgrades?	Commit concentration; multisig signers; voting power
Token ownership	Who holds the voting or staking weight?	Insider allocations; holder distribution
Front ends and data	Does the user path depend on one hosted service?	Default RPC endpoint; hosted indexer
Stablecoins and oracles	Can an issuer or data feed operator override the contract?	Issuer terms; oracle operator set

Governance and token ownership

A December 2021 BIS Quarterly Review article by Sirio Aramonte, Wenqian Huang and Andreas Schrimpf argues that full decentralization in DeFi is illusory. Governance revolves around governance-token holders, often platform developers, who vote much as shareholders do. In proof of stake, bigger stakers win rewards more often while costs are mostly fixed, which "naturally leads to concentration", and many chains allocate a substantial part of initial coins to insiders.

Front ends and infrastructure

Wallets and apps usually reach a chain through a node service. Ethereum.org's page on node services states that using one means "centralizing the infrastructure aspect of your product" and lists 30 providers. A ledger can keep producing blocks while an app depends on one provider's uptime and filtering.

Stablecoins and oracles

Circle's USDC terms (updated October 15, 2026) reserve the right to block transfers of USDC to and from an address on chain, and say Circle may be required to freeze USDC when a valid government authority orders it. A chain can be decentralized while a token on it is controlled by its issuer.

Oracles feed outside data, such as prices, into contracts. Ethereum.org's oracle documentation says decentralized oracles aim to remove a centralized oracle's single point of failure, which needs decentralization of both data sources and node operators, since nodes pulling from the same source fail the same way.

What regulators mean by the word

United States. The SEC's interpretation, published in the Federal Register on March 23, 2026 (91 FR 13714, with CFTC guidance, effective that day), says that for purposes of the release a crypto system is "decentralized" if it functions and operates autonomously with no person, entity or group having operational, economic or voting control. A digital commodity "may be" native to such a system, but decentralization is not part of the definition; the listed examples include Bitcoin, Ether, Solana and XRP. Where an issuer promises to achieve decentralization, whether it did is judged by how the issuer defined the term, not by a general market conception.

European Union. Recital 22 of MiCA (Regulation (EU) 2023/1114) says the regulation should apply to persons and their crypto-asset services even when part of the activity is performed in a decentralized manner, and that services provided in a fully decentralized manner without any intermediary should fall outside its scope. A May 2023 CMS analysis lists software that individuals can modify, a DAO acting as a legal entity and voting power concentrated in one person as signs a project is not fully decentralized, and noted that authorities had not yet issued guidance on the term.

Frequently asked questions

Is open-source code the same as decentralization?

No. Open code lets anyone inspect and copy the software, but control depends on who runs nodes, who can merge changes and who holds keys. The Nakamoto coefficient method treats developers as a separate subsystem for this reason.

Does a token have to be decentralized to be a digital commodity under the SEC interpretation?

No. The release defines a digital commodity by value derived from the programmatic operation of a "functional" crypto system and supply and demand, rather than from expected profits from others' essential managerial efforts. It says the examples were chosen because each underlies a CFTC-regulated futures contract, while stating that such a contract is not required.

Why does Ethereum client diversity matter?

Clientdiversity.org sets a goal of under 33% for every client, execution and consensus alike. A client above 33% that forks can stop the chain finalizing, and one above 66% could finalize its own chain, after which its validators could not return to the real chain without being slashed. Its data sources disagree widely, so no single share figure is settled.

Does running a validator remove the node-service dependency?

For solo staking, yes. Per ethereum.org, solo stakers must run their own execution and consensus clients rather than rely on third-party providers, which costs storage, bandwidth and time.

Why do the staking and mining coefficients differ?

They count different things at different thresholds. Mining pools are measured against 50% of blocks, staking entities here against one-third of stake, and each source groups entities differently. Each is most useful for tracking one subsystem over time with the same source.

A checklist for testing a claim

Name the layer the claim covers: consensus, nodes, clients, governance, ownership, front end, or assets.
Find a dated concentration figure for that layer, note how the source groups entities, and compute the smallest group that passes the relevant threshold (50% or one-third).
List who can change the rules or upgrade contracts, and whether participants can refuse an upgrade and keep running.
Trace the user path from wallet to RPC provider to indexer to front end, and mark any step with a single operator.
Check whether the asset itself has an issuer or admin who can freeze or block transfers, and whether the claim defines "decentralized" in terms a rule or document actually uses.