Holder Concentration: Measuring Token Distribution Honestly

Arkham's Bitcoin holder data, timestamped October 15, 2026, shows Grayscale's trust spread across more than 1,750 addresses, none holding more than 1,000 BTC. One beneficial structure, deliberately fragmented. Every address-level concentration metric computed on that chain reads the trust as seventeen hundred small holders, and not one of those metrics is miscalculated. The arithmetic is right and the inference is backwards, which is this measure's ordinary condition rather than an edge case.

What the Metric Counts

Top-N share is the summed balance of the N largest addresses over a supply figure, at a stated block height. Holder count is addresses with a nonzero balance. Gini is computed over the distribution of address balances. The Nakamoto coefficient is the smallest number of entities needed to cross a control threshold.

The unit of observation in all four is an address balance at a block height. Not a holder, not an owner, not a claim. Everything useful and everything misleading follows from that substitution, because the snapshot is exact and reproducible from any full node while the label set that turns addresses into entities is neither, and the two are published as one number.

Denominators compound it. Roughly 3.7 million BTC sit permanently in inaccessible wallets, and Satoshi Nakamoto's 1.096 million, 5.5 percent of supply, has not moved in a decade, so a share of circulating supply counts coins that cannot participate in anything.

Two Errors That Point in Opposite Directions

Pooling overstates concentration: one address, many owners. Binance cold wallets hold 249,000 and 177,000 BTC, Robinhood 141,000, and those custody client deposits rather than the exchange's own position. Staking contracts, bridges, pooled liquidity and ETF custody have the same shape.

Splitting understates it: one owner, many addresses. Grayscale's 1,750 is the clean case. Fidelity Custody's 444,000 BTC includes roughly 184,000 held for Strategy, and Strategy's entity figure of 848,000 counts those same coins, so one holding attributes to two named entities depending on whether custody or beneficial ownership is measured.

Because the errors run opposite ways, the sign of the aggregate error is unknown without labels. A top-10 share of 40 percent could be eight custodians and two whales or the reverse, and those are different assets. The top unattributed addresses in that dataset hold 92,000, 78,000, 54,000, 52,000 and 44,000 BTC, so the question is live at the top of the table.

Gini is worse again. The dust tail dominates it and addresses cost nothing to create, so the denominator belongs to whoever wants the number to move.

The Reading Improves as the Risk Arrives

Suppose a concentrated holder starts selling. Coins move from few addresses to many. Top-N share falls, holder count rises, Gini falls, the Nakamoto coefficient rises. Every published distribution measure improves, monotonically, for as long as the selling continues.

So the metric is not merely noisy. It is counter-indicating in the one regime where anyone consults it. A dashboard turning green during distribution is not malfunctioning. It reports what it was built to report, which is dispersion of address balances, and dispersion is what distribution produces.

No level detects this and no smoothing of a level will. This is also distinct from scheduled vesting, where quantity and date are known in advance. Here nothing is scheduled, and the signal sits entirely in flow and provenance.

What Breaks First Is the Label

The failure arrives through the label layer, and it breaks quietly because there is almost nothing to break against.

Bitcoin clustering rests on the co-spend heuristic. A 2026 preprint evaluating it across full clusters reported precision of 0.36, recall of 0.44 and an average error rate of 0.51, with an F1 of 0.000 on the second-largest service studied, and called it an investigative lead rather than a reliable attribution mechanism. Account-based chains have no direct equivalent: the best substitute, deposit address reuse, clustered 10.6 percent of active externally owned accounts, all heuristics together reached 17.9 percent, and no error rate was published because there is no ground truth to measure one against. NYDFS warned in April 2022 that these tools may miss beneficial owners absent off-chain verification.

The first operational failure is therefore a relabeling event read as an economic one. An exchange migrates a hot wallet or a staking contract is redeployed, and a fresh entry appears atop the holder table overnight. The alert fires, the figure steps, and nothing happened.

One cheap test separates them. A real economic change has a counterparty: coins left somewhere and arrived somewhere, with a path. A relabeling has none. If a top-holder balance steps and no other balance fell, the label moved rather than the asset.

Automation belongs here, since label maintenance is classification over a shifting address space. Make the control evidentiary rather than statistical: require every label change to carry the path or attestation justifying it, and version the label set so a historical figure can be recomputed under the labels then in force.

Other Numbers That Matter, and What Is Improving

Four substitutes, all computable. Change in the balance of named cohorts rather than the top-N level, since a level mixes custodians with owners. The funding ancestry of new holders, because a rising holder count whose new addresses were funded by one parent is distribution rather than decentralization. The largest labeled non-custodial balance in days of average traded volume, since twenty days in one hand is a different asset from two. And supply sitting in contracts rather than externally owned accounts, which is pledged or pooled and not discretionary selling capacity.

The honest limit on all four is that balances cannot show encumbrance. Loans, rehypothecation and side letters leave no on-chain trace, so no distribution metric speaks to solvency.

What is improving is that attribution is increasingly adjudicated rather than asserted. The same October 2026 dataset sizes holdings specifically, the United States at 307,000 BTC and BlackRock at 805,000, and publishes its own coverage, conceding roughly 83 percent verification of Strategy's on-chain holdings rather than implying completeness. In the Bitcoin Fog case on February 29, 2024, clustering testimony was admitted on corroboration and disclosure rather than a compiled error rate. Acknowledged gaps are what make a number auditable.

For an allocator the instruction is narrow. Stop reading the level, read labeled cohort flow, and require that any concentration figure in a manager's materials state its block height, denominator and label vintage. Without those three it cannot be reproduced, and a number that cannot be reproduced is a claim rather than a measurement.

For informational purposes only. Not an offer to buy or sell any security. Available only to accredited investors who meet regulatory requirements.

‍