The MEV Invisible Tax

• MEV didn’t disappear, it migrated. Every public-mempool cleanup pushed extraction into private order flow, builder markets, or spam, so the rent just changed hands.
• Private routing helps and isn’t a fix. A Dec 2025 study found 3,126 privately-routed victims sandwiched in two months for $409k, and ~40-54% of victims then migrate to private routing anyway.
• On L2s the tax is congestion. Base kills classic sandwiches with private mempools but burns ~132M gas and ~350 failed attempts per successful arb, raising base fees for everyone.
• No public mempool is no defense. Solana runs 90M+ arbitrages a year and real memecoin sandwiching; BSC concentrates 90%+ of MEV profit inside two builders.
• Every measured number is a lower bound, and almost nobody separates neutral arbitrage from surplus taken straight out of a user’s slippage, which is the number that should drive policy.

Every serious attempt to kill MEV over the past four years has succeeded at exactly one thing, which is moving it somewhere the old measurement tools can no longer see. Public mempools got cleaned up and the extraction moved into private order flow. Rollups adopted private sequencer mempools and the extraction turned into spam. Solana never had a global mempool at all and still runs industrial-scale arbitrage and sandwiching through validator-adjacent channels. The value did not evaporate anywhere in that story, because maximal extractable value behaves like a rent that follows whoever controls transaction ordering, and cleaning up one venue only hands that rent to whoever controls the next one. Ordering power has simply been changing hands, and the tax has followed it.

That is why the “invisible tax” framing is worth more than the usual leaderboard coverage of MEV. Most users never see a line item. They see a swap that filled at the worst price their slippage tolerance allowed, a base fee that sits higher than the chain’s real demand justifies, a transaction that failed twice before it landed. The surplus that leaks out of those moments does not show up as a fee, and it gets collected by a small set of searchers, builders, and relays who sit closer to the ordering decision than anyone routing an ordinary trade. The argument over whether MEV exists ended years ago. The live question in 2026 is how much of the tax users are quietly paying, where it moved to, and which “solutions” genuinely removed the rent versus just handing it to a different intermediary who is harder to see.

The mechanics haven’t changed, the plumbing has

Frontrunning, backrunning, and sandwiching are still the same three moves they always were. A searcher who sees your pending trade can jump in front of it to profit from the price move you are about to cause, slot in immediately behind it to capture the dislocation you left, or do both at once and squeeze your entire slippage budget out as profit. Liquidation extraction sits in a more ambiguous place, because closing an undercollateralized loan quickly keeps a lending protocol solvent and that is genuinely useful work, right up until someone censors a borrower’s attempt to top up their position or nudges a price specifically to trigger the harvest.

In crypto, one bad click is final. Learn to verify before you lose.

What changed after the Merge is the supply chain wrapped around those moves. Early Ethereum turned MEV competition into priority gas auctions fought in the open, where searchers watched the same opportunity, copied each other’s transactions, and bid up fees until one of them won inclusion. The 2024 remeasurement of Ethereum MEV, which parsed billions of transactions and traces with a profitability-first method rather than shape-matching heuristics, identified nearly 9.4 million MEV activities through August 2023, including 6.3 million arbitrages and just over 3 million sandwich attacks. That is the visible, public-mempool era, and it is the part of the picture we can measure most confidently precisely because it happened in the open.

Bundle relays and private order flow broke that openness on purpose. A searcher on Ethereum today can send a bundle straight to a builder through mev-boost, and a user can route through a private RPC to avoid the public mempool entirely. Proposer-builder separation splits the validator who proposes a block from the specialized builders who construct it, and it did reduce some of the on-chain waste from failed auction spam. But it also concentrated the ordering decision into a handful of builders and the relays that route to them, which is a different distribution of power.

Private routing helped users and did not solve the problem

The comforting story about private RPCs is that they end sandwiching, because an attacker cannot frontrun a trade it never saw. The execution data mostly supports the first half of that. The most careful public study of DEX trading costs, the “Don’t Let MEV Slip” analysis of Uniswap swaps, found that mean transaction cost ran around 22 basis points on the deep WETH-USDC pool and roughly 140 basis points on the thin WETH-PEPE pool, and that most of the size-dependent slippage was adversarial reordering rather than benign price movement. Trades in the illiquid PEPE pool carried about an 80% higher probability of adversarial slippage than the blue-chip pair, which tells you where the tax falls hardest. It concentrates on memecoins, thin books, and the retail flow that trades them, so the people least able to model their own execution costs are the ones paying the most, and private routing does take a real bite out of that adversarial component.

The trouble is that “private” is a trust arrangement and a December 2025 study made the cost of that assumption legible. Looking only at privately routed transactions on Ethereum across November and December 2024, the authors confirmed 2,932 private sandwich attacks against 3,126 victim transactions, for $409,236 in user losses and $293,786 in attacker profit. The losses are heavy-tailed, with a mean near $137 but a median of only about $27, so most victims lost lunch money while a handful got taken for thousands, and one address ate 28 separate private sandwiches. A single bot ran nearly 65% of the attacks, which is the same lopsided concentration that lets one dominant sandwich operator stay recognizable for years, to the point that in May 2026 a bot caught Vitalik Buterin himself in a sandwich on a small token swap. When the person who co-designed the chain gets taxed on a routine trade, the idea that ordinary users are opting into this risk with full information stops being credible. The same paper tracked how users respond, and around 40% of sandwiched traders migrated to private routing within 60 days, rising to 54% after repeated hits, which is the invisible tax changing behavior even where the per-trade loss is small. Private channels help, and they are being exploited by the privileged actors who operate them, and both of those things are true at once.

There is a second Ethereum tax that has nothing to do with getting sandwiched directly. Competition for ordering is itself a system-wide rent, because bots read every mempool transaction and bid fees high enough to win priority, and that bidding consumes blockspace and pushes base fees up for everyone regardless of whether any individual user is targeted. Recent work on auctioning time to defuse these latency races estimates a shadow price for that priority in the millions of dollars a month, a congestion cost paid collectively by people who never appear in a single sandwich dataset.

On the L2s, the tax is mostly congestion

The rollups tell the opposite story from Ethereum, and it is the more counterintuitive one. Private sequencer mempools were supposed to make sandwiching impossible, and by the numbers they largely did. The CCS 2024 study Rolling in the Shadows, which reconstructed MEV across Arbitrum, Optimism, and zkSync over nearly three years, detected no same-layer sandwiching on those chains at all, though it did show that simulated cross-layer attacks bridging L2 and Ethereum could have cleared around $2 million. A January 2026 follow-up on private L2 mempools went further and argued that over 95% of flagged sandwich patterns fail basic economic-consistency checks, with median profitability negative, so classic sandwiching on private-mempool rollups is presently rare and usually not worth doing.

That win came with a bill attached, because suppressing sandwiching did not suppress the underlying competition, it redirected it into brute-force probing. Flashbots’ 2025 analysis of what it calls optimistic MEV found that on Base, a single successful arbitrage could take around 350 failed attempts and roughly 132 million gas, which is close to four full Ethereum mainnet blocks burned to capture one small profit. Because native rollup mempools are private, searchers cannot see the opportunity a user’s trade creates until the block is published, so they spray speculative transactions into the same block and hope one lands behind the right trade. Nearly all of the additional throughput Base added between late 2024 and early 2025 got eaten by these spam bots, two of which accounted for more than 80% of the spam, and across leading OP-Stack rollups this behavior routinely consumed more than half of gas while paying under 10% of fees. The user never gets sandwiched. The user just pays a higher base fee forever, subsidizing a race they are not part of, which is a tax by any honest accounting even if no one calls it one.

No public mempool is not the same as no MEV

Solana is the cleanest refutation of the idea that removing the global mempool removes the extraction. There is no Ethereum-style public waiting room for transactions, and the chain still runs one of the largest arbitrage markets in crypto. Umbra Research documented at least $3 million in purely atomic arbitrage profit with a failure rate around 96%, and Jito’s own detection identified more than 90 million successful arbitrage transactions in 2024 averaging about $1.58 each. Sandwiching, which Solana was long assumed to be structurally resistant to, has become a real harm in memecoin order flow, with one on-chain program running over a million sandwiches in a single month. Jito bundles and MEV-protect routing move the competition off the public path and toward validators, which reduces wasted spam and recaptures some value, and they also concentrate order flow into a particular set of block engines, which is the familiar trade of one dependency for another.

BSC shows what happens when this concentration is designed in from the start. The chain runs a PBS-like architecture where only whitelisted builders participate, blocks come every 750 milliseconds, and private order flow bypasses the public mempool entirely. A 2026 study of MEV on the Binance builder found that within months, 48Club and Blockrazor produced over 87% of blocks and captured north of 90% of MEV profit, concentrated in short arbitrage routes over wrapped tokens and stablecoins. The block time is so tight that external searchers have to land bundles inside a couple hundred milliseconds to be considered, which they mostly cannot, so the builders who already control the block simply attach the last-moment arbitrage themselves. The extraction is not chaotic here. It is orderly, internalized, and owned by two firms, which is arguably a worse outcome for anyone who cares about who captures the surplus.

The future of finance is here. Trade crypto and more

The tax is going multi-chain

The newest frontier makes the single-chain framing obsolete. A year-long study of cross-chain arbitrage covering September 2023 through August 2024 found 242,535 executed arbitrages worth $868.64 million in volume across nine blockchains, with activity growing 5.5x over the period and surging after the Dencun upgrade cut L2 data costs in March 2024. Most of it clustered on Ethereum-centric L1 and L2 pairs, the five largest addresses ran more than half the trades, and one address alone captured nearly 40% of daily volume after Dencun. Some of what looks like benign price alignment between two chains is already part of a broader competition for sequencing power spanning both of them, and it is centralizing the infrastructure that captures it, because a firm that controls inventory on both sides and its own fast path between them wins by default.

Put the chains side by side and two trend lines dominate everything since 2022. Toxic MEV did not disappear, it dispersed into private channels, builder markets, and high-throughput spam. And user harm now arrives through two doors at once, the direct extraction of a sandwich and the generalized fee externality of everyone else’s arms race, with different chains weighted differently between them. Ethereum carries the richest direct-sandwich evidence, Base and Optimism illustrate the congestion tax, Solana proves that no public mempool is no defense, and BSC shows what builder capture looks like when it is fully concentrated.

What the fixes relocate

The honest way to evaluate any MEV mitigation is to ask which rent it removes, which rent it merely relocates, and what new trust assumption it smuggles in. Almost none of them clear all three cleanly.

Proposer-builder separation improved validator revenue and cut some failed-transaction waste, and the low revert rates on PBS Ethereum against the much higher revert rates on non-PBS rollups are a real, measurable benefit. It also parked enormous influence with a few builders and the relays that route to them, which remain trusted for validity and data availability, and builder concentration on Ethereum has stayed high enough that a couple of firms produce most blocks. Private RPCs and relays give users faster relief and, in the Uniswap execution study, effectively eliminated adversarial slippage in-sample, but the private-sandwich data proves the protection is conditional on the operator’s behavior. Private sequencer mempools on rollups suppress classic sandwiching today and push the extraction into spam and sequencer discretion instead, and their protection lasts only as long as the current architecture does, because the moment a rollup moves toward a public mempool or a richer builder market, the feasibility of sandwiching moves with it.

A newer class of design tries to replace the latency race with an explicit market. Arbitrum’s Timeboost auctions access to a faster inclusion lane, and Optimism has tested stake-based priority ordering to change how heavy blockspace consumers behave. These can convert a chaotic hidden tax into a legible auction, which is genuinely better than a spam war, and they also risk formalizing and monetizing privileged access to sequencing, so they reduce waste while writing the rent into the protocol. The intent and solver systems, CoW Swap, 1inch Fusion, FastLane’s Atlas on Polygon, and the RFQ desks that fill orders off-chain on Solana, take a different route by restructuring execution so professional solvers compete on the user’s order before it ever hits a mempool. That can sharply cut sandwich exposure, and it substitutes solver concentration and a new set of governance questions about who gets to see and fill order flow for the old public-mempool risk. Every row in this table trades one exposure for another, and pretending otherwise is how the space keeps announcing that MEV has been solved.

What we still cannot measure

Every number in this piece is a lower bound, and the researchers producing them say so plainly. ZeroMEV warns that its totals materially understate reality, especially for cross-domain activity and poorly classified categories. The remeasurement work still leans on inferred private-transaction data. The rollup studies concede that private sequencer mempools hide behavior by construction. Solana figures depend on what Jito and affiliated analytics can observe. What none of these capture is the full social cost, because they miss builder-internalized backruns, off-chain hedging against centralized exchanges, and the induced behavior of users who route privately or stop trading on-chain altogether. The invisible tax is almost certainly larger than anything public measurement can currently prove, which is an uncomfortable place to be when people keep citing these totals as if they were complete.

The deeper problem is conceptual rather than instrumental. Not all MEV is a transfer away from users, because some arbitrage tightens prices and some liquidation activity keeps protocols solvent, and a fair amount of backrunning can be structured to hand value back. But most public dashboards still do not cleanly separate neutral arbitrage from the toxic extraction that comes straight out of a user’s slippage, which makes headline “MEV revenue” figures close to useless for policy. The question worth asking is not how much MEV exists but how much of it is surplus a user would otherwise have kept, and almost no one is measuring that number directly.

Disclosure will beat prohibition

For anyone thinking about rules, the hard part is that sandwiching looks a lot like the order anticipation and manipulation that traditional markets already treat as abuse, while the venue and the actors map badly onto existing legal categories. A validator is not a broker, a relay is not an exchange, and a builder pre-filtering market access is doing something with no clean regulatory analog. Legal analysis since 2022 keeps landing on the same caution, which is that transplanting broker-dealer or licensing frameworks onto this stack invites regulatory arbitrage, because the operators can move jurisdictions while the market stays globally accessible.

That argues for a structural agenda rather than an actor-by-actor licensing regime. The choke points already exist and are named, being relays, builders, private transaction routers, centralized sequencers, and validator-side block engines, and every one of them can shape execution quality, censorship, and who captures value long before an ordinary user touches a block. The credible policy ask is standardized execution-quality reporting, ordering-policy disclosure, relay and builder transparency, anti-censorship monitoring, and competition audits where sequencing power concentrates. Those survive cross-border market structure in a way that jurisdiction-specific classification does not, and they put pressure exactly where the concentration is already visible in the BSC and Ethereum builder data.

New Scam Prevention Course. Never make a rushed crypto decision again.

The practical takeaway for builders is that the tax is cheapest to kill at the point of order creation, before a toxic opportunity is ever broadcast, which means defaulting vulnerable retail swaps to private or intent-based routing, shipping strict slippage presets, and publishing execution-quality dashboards instead of explicit-fee comparisons that hide the real cost. Wallets should turn protection on by default where it is reliable and warn users that a “successful” fill on an illiquid token can still be a predatory one. Exchanges and aggregators moving large uninformed flow, which is the clearest target of all, should be routing through RFQ or batch mechanisms and reporting effective execution rather than headline spread.

MEV is best treated as a hidden execution-cost layer sitting under every on-chain trade, rather than a searcher-profit leaderboard to gawk at. The number that decides whether a chain is healthy is how much of the captured value came straight out of users who would otherwise have kept it, and until the ecosystem measures that figure directly and discloses it by default, the tax stays invisible by design and gets paid by the people least equipped to see it coming.