tl;dr

• The nature of bidding by top builders in block auctions has evolved dramatically: they now start bidding later, bid much faster, and raise their bids by much smaller increments.
• Latency has become a crucial competitive factor, alongside order flow, transforming many auctions into statistical games.
• Timing games as a service has led to increased competition between relays and made auction end times more predictable, moving the ecosystem closer to a sealed-bid auction format.

Introduction

The current implementation of proposer—builder separation (PBS) assumes that honest proposers request the most valuable block from relays at the start of each slot. However, due to latency variations between proposers and relays, as well as consensus client runtimes, the requests arrive several hundred milliseconds later from honest proposers, and significantly later from proposers who intentionally delay their get_header() requests—known as timing games. As a result, MEV-Boost auctions have an uncertain end time and can be modelled as a digital adaptation of a candlestick auction with added cancellations.

The stemming uncertainty, coupled with public bidding, encourages continuous competitive block submissions from builders, where we denote a bid for a block as competitive if it exceeds the previous highest bid at the relay. Analyzing data from March to April 2023, Schwarz-Schilling et al. found that, although bidding rates accelerated late in the slot, median block values increased linearly. This linearity suggested that final block value could be approximated well before the end of the slot and shaped analyses of potential Ethereum protocol upgrades (e.g., the original MEV burn proposal) and conversations around shifting the timeline for allocating block proposal rights. However, over the past two years the behavior of block builders and the profile of block values have shifted, and any conclusions that rely on earlier assumptions should be revisited.

Beyond protocol-level modifications, the ability to anticipate terminal block values and likely winning builders has also shaped the order flow landscape. With this information, private mempool operators and order flow auctions are able to guarantee transaction inclusion, while keeping the cost to users low. By intelligently allocating order flow to the probable winner, order flow can be allocated exclusively or preferentially to builders, allowing for refunds of excess fees.

When examining the evolution of the MEV-Boost market structure, most discussions focus on macroscopic metrics, chief among them, as shown in Figure 1, is the consolidation of the builder market around two entities that now win approximately 95% of all auctions. Researchers who dig deeper often focus on exclusive order flow deals or builder profits and subsidization. However, behind the scenes, builders have also become faster and more sophisticated. This analysis shines a spotlight on shifting bidding patterns, the growing importance of latency minimization, and how some relays are already experimenting with alternative auction structures—often without wider ecosystem awareness.

figure-012400×1300 182 KB
Figure 1: the distribution of major block builders throughout Ethereum’s history. Although we saw a diverse set of builders in 2023, over the past year we’ve seen the market concentrate into two major entities: beaverbuild and Titan Builder who are now responsible for approximately 95% of all blocks built through PBS.

Evolving Bidding Strategies

Over the past two years, competitive bidding in MEV-Boost auctions has become significantly faster. The rise in bidding rates has been enabled by advances in optimistic relaying, which allows relays to skip block simulation for bonded builders, and by the introduction of top-bid websockets. These websockets let builders subscribe to real-time updates on the leading bid, eliminating the need to continually poll the relay. Meanwhile, the top builders, primarily beaverbuild and Titan, but also rsync-builder and Flashbots, have modernized their infrastructure and moved closer to relays, enabling faster block production and quicker responses to new information.

In this analysis, we focus on bids at the Ultrasound relay. Ultrasound holds the highest market share, receives bids from all the major builders, doesn’t handicap its block values by censoring OFAC non-compliant transactions, and is generally considered the fastest relay.

figure-021800×975 51.8 KB
Figure 2: the daily average count of competitive bids at Ultrasound relay since the beginning of 2023. We see that the number of competitive bids has tripled since January 2023—an increase of 73% per year. The data from January 2023 to June 2024 (inclusive) is spot-sampled from the Ultrasound relay bid archive. The data starting in July 2024 leverages the relayscan top bidsocket archive. The methodology was validated between the two sets and found to deviate by less than one percent.

Although MEV-Boost auctions have seen a marked increase in both total bids and competitive bids, this growth has not been uniform. In Figure 3, we note that despite the explosion in bids near the end of the auction, there has actually been a decrease in early bids. In fact, as a result of all four of the top builders choosing to delay their initial bid time until later in the slot, 2024H2 recorded the fewest competitive bids by the original MEV burn cutoff of any half-year since the launch of the Ultrasound relay. The early-bid rate has fallen by half since 2023H2, making it significantly harder for both individual actors and the protocol to predict final block values in advance. This shift in market structure has, in turn, forced a redesign of early MEV burn proposals.

It may appear that builders are manipulating the system by concealing their block values, however a more pragmatic perspective is that the system has not been designed to lead to this behaviour. In practice, builders now have a deeper understanding of the auction mechanism. Outside of the impact on MEV burn, the loss of this early information combined with the rise in private mempools creates challenges for intelligent order flow multiplexing and gas price estimation, issues that BuilderNet and others aim to address.

figure-031950×975 116 KB
Figure 3: the number of competitive bids arriving throughout the slot (bin size: 50 ms) at Ultrasound relay for each half-year starting in 2023. We see a general trend towards higher density bidding that occurs later in the auction. A breakdown by builder for 2024H2 is available in Appendix A1.

The most striking change is the large increase in late competitive bidding in 2024. Using a bin size of 50 ms, the bidding distribution for the half year peaked at an average of 65 competitive bids per second with roughly 500 ms left in the slot, coinciding with rsync-builder’s entry into the auction, seen in Appendix A2 (we note that on shorter timeframes, observable in Appendix A1, bidding rates per second reached more than double the half year peak). Taking the reciprocal, peak leader turnover at this stage hovers around 15 ms. Although auctions rarely conclude that early, bidding activity tapers off afterward, indicating that in many cases order flow is the dominant factor and that builders other than beaverbuild and Titan lack the required order flow to consistently compete—once the less competitive bidders are out of the auction, the net rate of bidding decreases.

Figure 3 reveals further nuances about MEV auctions. Noticeable bidding spikes at −4 seconds and −3 seconds mark when different builders begin submitting blocks. During 2023H2, rsync-builder triggered the −4 second spike, while Titan and Flashbots contributed to the −3 second spike. By 2024H2, these early spikes disappeared entirely, indicating that no top bidder was fully engaged prior to −2 seconds. By the end of 2024, rsync-builder shifted its initial bidding to −500 ms, resulting in the spike that aligns with the overall peak in the distribution.

An artifact of particular note is the two sawtooth patterns, one beginning at −8 seconds in 2023H1 and another at −3 seconds in 2023H2. These reflect periodic bidding strategies that were once prevalent in a less sophisticated market: Flashbots submitted its best bid every 500 ms, and Titan Builder did so every 250 ms—a magnified figure is available in Appendix A3.

A natural consequence of increased turnover among leading bids over a shorter period is that each bid holds the top spot for less time. Truncating the dataset to the timeframe where relays receive and respond to get_header() requests from honest proposers, the median duration of a best bid at Ultrasound relay dropped from over 80 ms in early 2023 to just 20 ms in late 2024. This trend reflects substantial infrastructure upgrades by both Ultrasound relay and dominant block builders. With the introduction of optimistic relaying and top-bid websockets, relays now process and serve bids faster. Meanwhile, block builders have colocated with relays and optimized their hardware and software to shave off thousandths of a second in their pipeline.

figure-041950×975 112 KB
Figure 4: cumulative distribution by half year of how many competitive bids are still leading after a given period of time has passed. We see significant quickening in bid turnover, with the median duration as the best bid dropping from 84 ms to 20 ms over the course of 18 months. The data used to generate this figure is filtered down to only include bids arriving between the start of the slot and one second later—the highest density period of winning bid arrival times over the past two years.

Rather than order flow dictating whether Titan or beaverbuild secures a block, in many cases bidding now more closely resembles a game of chance—one in which the winner is statistically favored by holding the top bid for longer. This transition away from pure order flow dominance has diminished the incentive for top builders to significantly outbid one another. Consequently, the network has seen a rise in pennying, where each successive bid increase is as small as possible. As illustrated in Figure 5, the median increase of a competitive bid has dropped by more than 80%, from 0.46% to 0.09%.

figure-051800×975 105 KB
Figure 5: cumulative distribution by half year of how many bids exceed the previous bid by a given percentage. We see a significant decrease with the median increase dropping from 0.47% to 0.09% over the course of 18 months. The data used to generate this figure is filtered down to only include bids arriving between the start of the slot and one second later—the highest density period of winning bid arrival times over the past two years.

The ongoing transition from pure order flow dominance to a model that also incorporates latency-driven pennying has amplified the performance gap between performant relays and the field. Evident in Figure 6, two distinct classes of relays have emerged. bloXroute, Titan, and Ultrasound see rapidly updating leading bids, while competitive bids at Aestus, Agnostic, and Flashbots tend to spend more time as the leader. Notably, the market shares of these less performant relays appear to be inflated by cooperative block propagation—the vast majority of blocks attributed to these relays are propagated by many relays, calling into question the magnitude of the role they play in the ecosystem. Aestus and Agnostic propagate only a handful of unique blocks per day, while Flashbots is the only relay on 50—100 blocks per day, a sharp decline from 1,000 blocks per day at the beginning of 2024.

figure-061800×900 130 KB
Figure 6: cumulative distribution of how many competitive bids are still leading after a given period of time has passed, separated by relay. We see two distinct groupings of relays, with bids at Aestus generally being the most stale despite their upgrades to the base Flashbots relay code—beaverbuild does not frequently submit bids to Aestus, therefore this figure shows the importance of relays accommodating both top builders in order to remain competitive. The data in this figure is from July 2024 and is filtered down to only include bids arriving between the start of the slot and one second later.

Despite being marketed as a ultra-fast Rust rewrite of relay functionality, the potential for builder–relay integration and the strict requirements for websocket access have kept beaverbuild from submitting the majority of its blocks to Titan Relay. As a result, price discovery primarily occurs at Ultrasound and bloXroute, while Titan Builder forwards all of its bids from other relays to Titan Relay. This approach keeps Titan Relay competitive; however, as shown in Figure 7, Titan Builder now wins over 90% of the blocks at Titan Relay, making it function increasingly like a direct connection to Titan Builder (see the comments for additional nuance).

figure-071950×975 70 KB
Figure 7: the distribution of winning block builders for blocks propagated by Titan Relay. Although rsync-builder once had a healthy market share of blocks at the relay, in recent times Titan Builder has become dominant, now winning over 90% of blocks at the relay.

Timing Games as a Service

Builders are not the only ones reshaping the MEV-Boost market; in fact, the most significant shift of 2024 came from relays. After P2P began intentionally delaying their get_header() requests in November 2023, researchers from the Ethereum Foundation proposed that relays could similarly introduce artificial latency into the system in order to reduce centralization pressure on the proposer set, an approach commonly known as timing games as a service.

However, encouraging relays to play these games has introduced several negative externalities. The increased sophistication has furthered the divide between the two classes of relays, with Ultrasound and bloXroute pushing the limits in order to attempt to monetize what has traditionally been viewed as a public good. These games, as well as the move towards second price auctions, have fuelled the emergence of relay PvP and advantaged relays that are closer to block proposers, leading to market share concentrating at just two entities and an increase in net censorship. Furthermore, having lost control over when blocks are released, proposers now experience a principal-agent problem: a relay’s cost of missing a slot is much lower than a block proposer’s, leading to misaligned incentives and lower trust in relays.

This artificial latency is visible in Figure 8, where we see that the profile of winning bids has shifted significantly later for all proposers, including honest proposers who continue to request block headers at the beginning of the slot. In addition, we’ve seen that the width of the distribution has shrunk dramatically, making the end time of the auction much more predictable for block builders.

Figure 8: the evolving distribution of winning bid times, colored by relay. From November 2023 to January 2025, the median winning bid arrival time has shifted 656 ms later, while the standard deviation has dropped from 581 ms to 494 ms—auctions are ending later and more predictably.

The sharpest peak of the distribution came from bloXroute in the summer of 2024. Analyzing the arrival times of winning bids at their max-profit and regulated relays, it appears that bloXroute implemented a soft end time for get_header() responses at 800 ms, delaying all earlier responses until this time.

The results in Figure 9 are striking: competitive bids at 790 ms were 15 times more likely to win than those arriving at 810 ms, since bids past the cutoff missed the response window. Essentially, bloXroute replaced the candlestick auction with an eBay-style auction that had a predictable end time. As highlighted by Quintus Kilbourn in 2022, removing the end-time uncertainty reduces incentives for early bidding, pushing the auction closer to a sealed-bid model.

figure-091950×975 62.5 KB
Figure 9: the comparative densities of winning bid block arrival times. The distribution for the whole network is in red, from a time period before any relays began engaging in timing games as a service. In blue, we see data isolated for bloXroute’s relays between June 25, 204 and July 24, 2024, during with they used their sharpest auction end time to date.

Timing games as a service was initially seen as an accelerationist strategy for mitigating proposer timing games, while some researchers favored alternate approaches. One such alternative was to conscript relays to act as gatekeepers, rejecting bids from builders that arrive after the slot boundary to encourage timely adherence to the PBS commit—reveal scheme. The most vocal opponent to this idea was Uri Klarman, the CEO of bloXroute Labs, who argued against burdening relays with additional policing responsibilities, especially in the absence of a clear funding model. Moreover, this proposal was not incentive-compatible: any single relay that declined to police timeliness could easily outcompete its peers. In response, bloXroute has since led the charge on timing games as a service, ultimately introducing a soft timeliness deadline highlighting how, in the limit, all these mechanisms converge on a fixed end time auction.

Concluding Thoughts

Driven by increasing sophistication, the MEV-Boost market has significantly centralized at both the builder and relay level. Top builders have acquired exclusive order flow and invested in infrastructure, while the top relays have eschewed collaboration and tried to set themselves apart. This heightened competition has the potential to push market structure towards vertical integration and sealed-bid auctions. Taken together, these trends pose new risks, as vertically integrated builder–relays enjoy privileged access to information. At the same time, they preview the solution: these trends accelerate market structure closer to removing the relay through either ePBS or TEE-Boost, both of which propose vertically-integrated sealed-bid auctions.

Appendix

Figure A1: the evolving distribution of competitive bid arrival times by builder in 2024H2. We note that Titan Builder shifted their initial bid timing from −2000 ms to −800 ms, while beaverbuild started bidding earlier, shifting from −1500 ms to −2000 ms. We also see the transition from Flashbots centralized builder into BuilderNet in November 2024.

figure-a021950×975 102 KB
Figure A2: rsync-builder’s competitive bid submission pattern by half-year. We see bid submissions starting at −4000 ms throughout 2023, but shifting to −1500 ms and then to −500 ms in 2024.

figure-a031800×975 60.5 KB
Figure A3: the distribution of bid submission times (all bids, not just competitive bids) by Titan Builder to Ultrasound Relay in October 2023. We note the periodicity, with bids arriving every 250 ms. The dataset used to generate the figure contains bid for slots 7,500,000 to 7,540,000 encompassing days from October 9 to October 14, 2023. The bin size in the figure is 2 ms.

Disclaimer:

1. This article is reprinted from [Flashbots]. All copyrights belong to the original author [Data Always]. If there are objections to this reprint, please contact the Gate Learn team, and they will handle it promptly.
2. Liability Disclaimer: The views and opinions expressed in this article are solely those of the author and do not constitute any investment advice.
3. The Gate Learn team does translations of the article into other languages. Copying, distributing, or plagiarizing the translated articles is prohibited unless mentioned.