Paper: Order Flow Exclusivity and Value Extraction Mechanisms — Ethereum Builder Centralization

Citation

Zhang, A., Liu, Y., Shan, Y., Zhang, R., Wu, Y. “Order Flow Exclusivity and Value Extraction Mechanisms: An Analysis of Ethereum Builder Centralization.” arXiv:2605.04471v1 [cs.CR] (May 6, 2026). Tsinghua / KU Leuven / Cryptape & Nervos.

Setup

Ethereum builder market collapsed from a distributed landscape (no builder >20% in late 2022) to an oligopoly: by Sep 2024, Titan + Beaverbuild = 87.7% combined; by Dec 2025, BuilderNet absorbed Beaverbuild but concentration remained.

Existing research focused on influential (high-revenue) order flows. This paper closes two gaps: (1) EOFs (Exclusive Order Flows) — which order flows are exclusive vs which are public, systematically, including for small/emerging builders; (2) non-atomic MEV — only 13 instances had been documented before; the systematic landscape was unknown.

Two Methodological Contributions

1. KL-divergence-based exclusivity metric

Identifies all EOFs (164,249 order flows analyzed Sep 2023–Aug 2025) without two prior biases:

Endogeneity bias — high market share trivially leads to high captured order flow.
Sample-size confound — low-frequency flows produce stochastic errors.

Solution: KL divergence between an order flow’s bribe distribution across builders vs the aggregate builder market-construction distribution, weighted by the flow’s total bribe.

Result: identifies 75 EOFs (68 previously unreported) that account for 70.53% of trading-related builder revenue. Newly identified flows contribute 34.96% of total EOF bribe.

2. Supervised-learning identifier for non-atomic MEV

Manually labeled top 210 revenue-contributing order flows → ground truth. Decision-tree classifier with augmented feature set (entry-point contract aggregate stats: unique interacting addresses, average tx-frequency per sender). 92.06% accuracy.

Result: 322 non-atomic MEV order flows (316 previously undocumented) → 22.99% of total builder revenue. Newly identified contribute 39.84% of that subtotal.

Notable: top-2 non-atomic flows capture 56% of all non-atomic bribes. Power-law distribution (α = 1.47) means a persistent long tail of niche players remains active.

Four Eras of Builder Market Evolution

Weekly Herfindahl-Hirschman Index (HHI) traces a clear centralization trajectory. The four eras:

Era	Defining feature
Genesis	Distributed; no dominant builders
Algorithm Wars	Builders compete on technical sophistication; algorithm quality is the moat
EOF Moats	EOFs become the primary differentiator; “chicken-and-egg” most acute
Oligopoly (current)	Titan + Beaverbuild + (BuilderNet) entrenched; EOF-bribe correlation with market share decoupled

Pearson correlation analysis: dominant builders had strong EOF-bribe ↔ market-share correlation during EOF Moats — but the relationship decoupled in the Oligopoly era. EOFs were instrumental in establishing dominance; incumbents now sustain market share via entrenched network effects that no longer require steady EOF income.

Central Thesis

Builder centralization is an emergent property of the PBS architecture itself, not just of EOFs or MEV per se. EOFs and non-atomic MEV are symptoms and catalysts, not the root cause. PBS systematically violates three prerequisites of a competitive market:

Diminishing returns to scale — instead, scale → more EOFs → more revenue → more scale (positive feedback).
Information symmetry — searchers selectively share private flow with preferred builders.
Low entry barriers — chicken-and-egg dilemma blocks new entrants.

⟹ PBS framework facilitates an inevitable trajectory toward centralization.

Notable Quantitative Results

Metric	Value
Order flows analyzed	164,249
EOFs identified	75 (68 newly unreported)
EOF share of trading-related revenue	70.53%
Non-atomic MEV flows identified	322 (316 previously undocumented)
Non-atomic MEV share of revenue	22.99%
Top-2 non-atomic concentration	56% of all non-atomic bribes
Power-law exponent (non-atomic bribes)	α = 1.47
Builder market share (Sep 2024)	Titan + Beaverbuild = 87.7%
Classifier accuracy (non-atomic MEV)	92.06%
Data window	Sep 2023 – Aug 2025; 5.23M blocks; 889M tx; 503,853 ETH in tips + bribes

Connection to Existing Wiki

Quantifies and empirically dates the Exclusive Order Flow and the Builder Flywheel flywheel argument — 70.53% of trading-related revenue from EOFs is a very strong validation.
The era-decomposition gives historical context for the MEV Supply Chain: Searchers, Builders, Relays, and Validators market-share snapshots and Mensah’s first-person account (PBS and MEV-Boost § “Why Only 3 Builders Control”) of refund-mechanic flywheels.
The PBS-as-emergent-centralization conclusion supports the case for the BuilderNet and Decentralized Block Building / SUAVE / Phase-2 distribution-of-block-building research direction.
The non-atomic MEV measurement work pairs with Paper: Mapping Public and Private JIT Liquidity in Ethereum’s Builder Ecosystem and Paper: Just-in-Time Resale in an Ahead-of-Time Auction — Timeboost/Kairos Event Study in the empirical-MEV-attribution genre.

Open Questions

❓ Does ePBS (EIP-7732) actually break the EOF flywheel, or just shift it onto on-protocol auction structures? The paper’s “PBS-as-architecture” thesis would predict no — the architecture is the cause.

❓ The Oligopoly-era decoupling of EOF-bribe vs market-share — is this causally network effects, or does it reflect that EOFs have reached a saturation point (no more incremental EOF revenue available beyond the entrenched ones)?

❓ How sensitive is the KL-divergence metric to the choice of aggregate baseline distribution? Robustness to baselines that themselves shift with market structure.