Over a rolling thirty-day window in this consolidation market, the ten largest solver addresses on the major intent-based settlement venues filled between 62% and 78% of all matched orders. That range comes from public settlement logs. Anyone can reconstruct it. Almost no one has.
Intent architectures were sold as the end of MEV. The pitch was clean. Instead of broadcasting a transaction into a public mempool where a searcher sandwiches it, the user signs a message describing a desired outcome. A network of solvers competes to deliver that outcome. Extraction gets internalized. The user wins. The narrative held for roughly two years.
The on-chain record does not support it. Audit gap confirmed.
I have spent the last several months treating intent venues the way I treated the ICO contracts in 2017 and the emission schedules in 2020. I read the logs. I did not read the blog posts. What I found is a clean, verifiable settlement layer sitting underneath an auction layer that publishes nothing. That gap is the story. It is also a very old story in a new coat.
The distinction matters before the data does. A transaction is imperative. It says: execute this call, in this order, at this gas price, against this contract. An intent is declarative. It says: make this outcome true. The executor — the solver — assumes responsibility for how. That single shift moves the entire optimization problem from the user to a third party.
The venues are now familiar. UniswapX runs a Dutch auction, where the price the solver offers decays over time until someone fills it. CoW Protocol runs a batch auction and settles at a uniform clearing price. 1inch Fusion runs a sealed-bid competition among resolvers. Across and a growing set of cross-chain venues run relayers who front inventory on the destination chain. ERC-7683 is emerging as the cross-chain intent standard. Account abstraction, ERC-4337, is the plumbing that lets a signed intent be a consumer product rather than a developer toy.
The solver is the entity that matters. A solver holds inventory, signs a commitment, and takes settlement and market risk. It is not a validator. It is not a searcher in the classical sense. It is closer to a market maker with an execution obligation and no disclosure requirement.
The model earned traction for defensible reasons. There are no failed transactions and no gas burned on reverts. Gas can be abstracted, so a user pays in the output token. The order never enters the public mempool, so naive frontrunning is structurally removed. For the median retail order, the realized price is genuinely better than a naive AMM route, because the arbitrage a sandwich bot would have taken is instead internalized into the fill.
Then the discourse arrived. After MEV-Boost and proposer-builder separation normalized the block-building market, the argument became: MEV is a tax on users, and intents return that tax. That is the claim under audit.
Set the market context. Realized volatility is low. Spreads are compressed. Intent economics are volatility-dependent and inventory-dependent. In a chop, the arbitrage margin shrinks and the competition for orderflow intensifies. This is precisely when structural problems stop being theoretical and start showing up in the numbers.
The first thing to fix is the equation nobody publishes.
A user signs a limit price, p_L. There exists a market reference price, p_M. The solver fills at p_F. For a sell order, p_F is at least p_L. The user's surplus is U = p_F − p_L. The solver's capture is C = p_M − p_F, minus cost.
The user compares p_F to an AMM quote. The solver compares p_F to the best price available across every venue it can reach, including private orderflow it has already purchased. Those are two different reference points. The distance between them is where the value sits.
This is not fraud. It is an undisclosed variance. The ledger does not lie: the settlement log records p_F and the token transfers. It never records p_M. The reference price against which the surplus was computed is not on-chain, not signed, and not published. The user cannot reconstruct what they gave up, because the number they would need was never written down.
A limit order on a traditional venue has the same property, and that is the point. The intent model did not remove the hidden midpoint. It moved it behind a signature and called the result a better price.
The second thing to fix is who gets paid for the flow.
Wallets and frontends originate intents. Solvers need that flow, because an unfilled intent earns nothing. So solvers pay for it. Sometimes the payment is explicit — a per-order rebate. Sometimes it is a points program, a revenue share, or a token grant. The mechanism varies. The economic substance does not.
This is payment for orderflow. TradFi's version was disclosed, litigated, and regulated. The crypto version is none of those things. On a public AMM, the user could at least see their transaction in a mempool and watch a bot reorder it. With an intent, the user sees a filled order and a banner that says the price was better. The tax may be identical. The receipt is missing.
Audit gap confirmed. A market is only as competitive as its disclosure. When the only published number is the winner's price, the word competition is an assertion, not a measurement.
The third thing to fix is the boundary between what is provable and what is merely claimed.
Settlement is on-chain, final, and verifiable. That part is honest engineering. The auction is off-chain, private, and unattested.
The comparison is instructive. MEV-Boost normalized a builder market with real problems — relay trust, bid censorship, a small builder set. But relays published bid data. Researchers could audit block-building competition. A disclosure surface existed, however imperfect.
Intent venues have no equivalent standard. There is no signed log of losing bids. There is no attestation of how many solvers quoted. There is no published spread between the winning quote and the runner-up. The protocol can prove the trade happened. It cannot prove the trade was competitive. Those are different claims, and only the first one is currently verified.
In 2026 I reverse-engineered an AI-agent platform that advertised decentralized identity and found a centralized database under a blockchain overlay. The intent auction is a cleaner version of the same architecture: a verifiable settlement layer bolted onto a private control layer. The marketing is better. The structure is the same.
The fourth thing to fix is the assumption that solver competition stays competitive.
Solvers need working capital and hedging infrastructure. In a sideways market, the cost of hedging inventory falls, but so does the arbitrage spread available to capture. The auction becomes a scale game. Fixed costs — infrastructure, compliance, capital, integration — amortize across volume. Marginal solvers exit. Survivors gain pricing power, which they use to win more flow, which increases volume, which lowers their unit cost.
In 2020 I mapped the emission schedule of a farm promising four-digit APY and found the same shape: a fixed cost base, a shrinking marginal return, and terminal concentration of participants. The intents market reproduces that curve with better branding.
This is the dynamic that concentrated AMM liquidity into a handful of pools and MEV into a handful of builders. It is not a governance failure. It is an equilibrium. The ledger does not lie: solver leaderboards are public, and concentration is measurable in the settlement logs. What is missing is not the data. It is the willingness to publish the derived metric next to the marketing.
The fifth thing to fix is the compliance surface, because it is where censorship-resistance quietly went.
Solvers are legal entities in specific jurisdictions. They run screening. A signed intent is a request, and a solver has no obligation to answer it. There is no slashing condition for refusal, because no fill obligation attaches to an unfilled intent.
So the property degrades. A public mempool transaction can be censored by builders, but it remains visible and can be rebroadcast. An intent that every solver declines never becomes visible at all. It expires. There is no log entry, no on-chain artifact, and no proof that a refusal occurred. The sharpest structural loss in the model is also the least discussed: censorship moves from visible exclusion to silent non-response.
The sixth thing to fix is the claim that cross-chain intents are bridges. They are not. ERC-7683 and its contemporaries place trust in the solver's inventory on the destination chain. The user gets speed. The solver takes a credit exposure and a settlement-latency risk. If destination inventory is short, the fill fails or settles late after the source-chain commitment is already locked. Functionally, a fast cross-chain intent is a credit facility with an unrated counterparty, and the user's collateral is the asset they locked on the origin chain. That is a loan, not a bridge.
The seventh thing to fix is the assumption that current pricing is durable. Define the breaking condition cleanly. Let S be the surplus available on an order, c the solver's cost — gas, inventory, hedging, infrastructure — and k the kickback paid to the orderflow originator. Solver margin m = S − c − k.
If k rises toward S − c, margin approaches zero. Solvers either exit or widen spreads. Widening spreads pushes p_F back toward p_L. The user's realized price converges on their own limit, which they set against a public AMM quote. At that point the intent delivers no improvement over a direct swap, and the user has paid for the difference in lost optionality.
The efficiency gain is therefore a function of subsidy, not of structure. Token-funded solver incentives are the subsidy in question, and they are the reason the current pricing looks good. Yield trap detected: a return on deployed inventory that depends on emissions rather than on spread capture. Mathematical collapse verified. The equilibrium is reachable, and nothing in the current design prevents it.
The bull case deserves its strongest form, because part of it is correct.
First, for the median retail order, the realized price really is better. The arbitrage a sandwich bot would have extracted is internalized into the fill. That is a genuine improvement, and it is measurable in the settlement data.
Second, the failure modes are better. No reverted transactions, no gas burned on failed attempts, no slippage surprises at execution. A user who signs an intent gets an outcome or nothing.
Third, concentration is not unique to intents. AMM liquidity is concentrated. Builder share is concentrated. RPC endpoints are concentrated. Intents are a fourth instance, not a new category of sin. Singling them out for a structural property that defines the entire stack is not analysis. It is scapegoating.
Fourth, and most important: the model is auditable in principle. The data exists at the solver level. A signed per-epoch quote attestation is a design choice, not a technical impossibility. The bears' blind spot is that they treat MEV relocation as automatically worse than on-chain MEV. It is not. The question is disclosure. If solver auctions adopted the transparency standard that relays were pushed toward, the entire critique weakens within one release cycle.
The market has not demanded it because the current price looks good. That is exactly when standards get set, and exactly when they are hardest to impose later.
Three things to watch, all reconstructible from public data. First, whether any major venue publishes a signed per-epoch attestation of quotes, including losing bids. Second, whether the orderflow kickback becomes a disclosed field rather than a private transfer. Third, whether solver concentration crosses 80% on any single venue.
The question is not whether intents win. They probably do, because the user experience is better and the gas abstraction is real. The question is whether the auction becomes auditable before the orderflow capture becomes permanent. If the auction has no audit trail, what exactly is the user competing for?
