Who Really Controls Numerai's Meta-Model Stake

Numerai's 3,373 staked models collapse into 948 accounts. Ten of them hold 46% of all staked NMR, and the pool has been ~23 accounts deep since 2023.

The Numerai leaderboard shows 3,373 staked Classic models, which sounds like a crowd. Follow the stake to the accounts behind the models and the crowd shrinks to 948 names, an effective depth of about 23 once you weight by NMR, and a single account holding 13% of the entire pool. The meta-model's capital base is not a swarm. It is a couple dozen books, and that number has barely moved in five years.

This matters for the same reason bank concentration matters. The payout factor and the stability of the staked pool are set by aggregate NMR at risk, and when the pool is effectively 23 accounts, "aggregate" behavior is really a handful of large private decisions. Signals demonstrated the failure mode in September 2025, when its at-risk stake collapsed 80% inside a month and the payout regime snapped with it.

Ten accounts are half the pool

Grouping July 2026's 852K staked NMR by account rather than model makes the pool visibly thinner: the top 10 accounts hold 46% of everything at stake, and the top 100 hold 92%.

Grouped bars comparing stake share of the largest N accounts versus largest N models: top 10 accounts hold 46 percent versus 37 percent for models, top 100 accounts 92 percent
Grouped bars comparing stake share of the largest N accounts versus largest N models: top 10 accounts hold 46 percent versus 37 percent for models, top 100 accounts 92 percent

The account view is the honest one for systemic questions. Model-level concentration (covered in the meta-model deep dive) understates who can actually move NMR, because a withdrawal decision happens at the wallet, not the model slot. The gap between the two views is widest exactly where it matters, in the top 10–50 range where multi-model books aggregate.

Effectively ~23 accounts, for five years running

Squeezing each round's stake shares into an effective N says the meta-model has been roughly as deep as a small prop desk for years: a median of 22 effective accounts in 2022, 24 in 2024, 21 in 2026.

Effective number of models and accounts per round since 2021, with models collapsing from about 180 to 27 and accounts stable near 23, the two lines converging after 2024
Effective number of models and accounts per round since 2021, with models collapsing from about 180 to 27 and accounts stable near 23, the two lines converging after 2024

The model line is the more dramatic one: effective models collapsed from ~180 in 2022 to ~27 today, converging onto the account line. In the weekly era, stake was spread across many mid-sized models and diversification across model slots was real; since 2024 the two series are nearly the same number, which means model count no longer adds effective depth. The brief early-2023 dip to single digits, when the shrunken pool was briefly dominated by its largest wallet, shows how fast this structure can pinch when total stake drains.

Most accounts run one model; the shops run fifty

The composition behind the concentration is a barbell. 60% of staking accounts run a single model, and those 572 accounts control 39% of the stake, because the two biggest books in the tournament are one-model accounts.

Distribution of staked models per account: 572 accounts run one model holding 39 percent of stake, while 52 accounts running 11 or more models hold just 18 percent
Distribution of staked models per account: 572 accounts run one model holding 39 percent of stake, while 52 accounts running 11 or more models hold just 18 percent

At the other end, 52 accounts run 11+ models each and still sum to just 18% of the stake. Model-farming buys breadth, not weight, the same asymmetry the diversification paradox found in returns. The top of the book makes it concrete:

Top ten Numerai accounts by staked NMR in July 2026, led by shatteredx with 110.7K NMR and 13 percent of the pool from a single model
Top ten Numerai accounts by staked NMR in July 2026, led by shatteredx with 110.7K NMR and 13 percent of the pool from a single model

shatteredx stakes 110.7K NMR — 13% of the pool — on one model; the second-largest book is also a single model. Meanwhile the largest multi-model operators (83 and 57 staked models) each control just over 2%. Whatever the meta-model's ensembling gains from thousands of submissions, its capital weighting is set by a few conviction bets, which is presumably why the stake-weighted performance gap the alpha-decay analysis found is so persistent: a handful of large, apparently well-informed books are the stake weighting.

What I'm watching

Concentration this stable is not obviously fragile — five years at ~23 effective accounts through two drawdown cycles is a decent stress test. The risk is state-dependent: the structure holds until a top-5 book exits, and at 13% for the largest, a single departure is a Signals-September event for Classic. Two things would update me. First, the effective-account series breaking sustainably out of its 20–30 band in either direction. Second, the v3 atomic staking migration: per-round stake management makes large books more mobile, and the Crypto pilot showed migration itself can empty a pool's at-risk stake within weeks. I'll re-cut this by account when Classic completes the move.

Method notes: current stakes from model profiles as of 2026-07-20, threshold ≥1 NMR (excludes the 2023–24 onboarding dust credits). Historical series joins each round's staked performances to today's model→account mapping; coverage is 96–99.9% of staked rows in every year since 2021, with unmatched (deleted) models treated as their own accounts. "Account" is a leaderboard identity: one person or team can operate several accounts, so true control is at least this concentrated. Benchmark models excluded throughout.