Here is the framing you will find in almost every explainer on this topic: AMMs are simple but inefficient, order books are complex but precise. Pick your poison.
That framing is wrong. Not slightly wrong. Foundationally wrong. Both architectures are losing relative market share to a third model that most traders cannot name. The real question is not which system is better in absolute terms. It is which system fails you in which specific situation, and why the answer depends on math most retail traders have never seen.
This is that explanation.
What an AMM Actually Is (Beyond the Textbook Definition)
An Automated Market Maker replaces human counterparties with a mathematical formula. The most common variant uses the constant-product formula: x times y equals k, where x and y are the quantities of two tokens in a pool and k is a constant that must hold after every trade. The pool always quotes a price. It never runs dry. Anyone can provide liquidity. No counterparty needed.
That sounds elegant. The part nobody talks about is who is on the other side of every profitable trade you make against an AMM pool.
The answer is the liquidity providers, and they are losing.
A seminal academic paper published in 2022 by Milionis, Moallemi, Roughgarden, and Zhang introduced a concept called Loss-Versus-Rebalancing, or LVR. It is the closest thing to a Black-Scholes formula that DeFi has produced. The math shows that AMM liquidity providers do not just suffer impermanent loss, the familiar concept that most LP guides explain. They suffer a structurally different and far more serious cost: the gap between the stale price the AMM quotes and the real market price that arbitrageurs observe on centralized venues and exploit every single block.
The difference between LVR and impermanent loss matters enormously. Impermanent loss reverses when prices retrace. If you provide liquidity at $2,000 ETH and price goes to $2,500 and then back to $2,000, your IL disappears. LVR does not reverse. Every time an arbitrageur corrects the AMM price to match the market, that correction extracts real value from the pool permanently, even if price returns to its starting point.
For a standard constant-product pool, LVR scales roughly with the square of the asset's volatility multiplied by pool value divided by eight, per unit time. On volatile assets like ETH, the empirical data from major pools shows LVR frequently exceeding total fee income. Liquidity providers in those pools are, on net, subsidizing arbitrageurs while receiving fee income that does not compensate for what they give up.
A follow-up paper published in February 2025 by Alexander, Lambert, and Fritz confirmed that IL and LVR only converge at short time horizons. At intermediate and long horizons, they diverge significantly. The 2024 Rebalancing-Versus-Rebalancing refinement added a useful counter-point: when measured against the realistic costs of rebalancing on a centralized exchange, AMMs can outperform centralized rebalancing for certain portfolio managers, particularly when CEX fees and slippage are accounted for. So the picture is not simply "AMMs are bad for LPs." It is that AMMs are systematically bad for LPs on high-volatility pairs where arbitrage flow dominates, and potentially acceptable on stable pairs where it does not.
The Casino Problem AMMs Cannot Solve
There is an analogy from traditional finance that maps cleanly onto AMM economics. A casino profits from uninformed gamblers but loses consistently to card counters. The solution in every casino is to identify the counters and refuse service. Foreign exchange market makers do the same thing: they tag client flow, measure how much of it is informed versus uninformed, and widen spreads or reduce size for clients whose trades persistently predict price movement.
AMMs cannot do this. The formula is public. The pool prices algorithmically. There is no mechanism to distinguish a retail trader buying ETH for long-term holding from a high-frequency arbitrage bot that has already seen the price move on three other venues. Both pay the same fee. Both get filled from the same liquidity. The bot extracts value. The retail trader contributes it.
This is why token launchpad pools are LP-profitable while major ETH pools on mainnet are often LP-toxic. A newly launched memecoin pool has no arbitrage market. Every swap is uninformed flow. The LP captures fees without giving up much LVR because there is no external reference price for bots to arb against. The moment a token develops meaningful price discovery on liquid venues, the arbitrage flow arrives, and the LP economics deteriorate.
The practical implication: AMM liquidity provision is not passive income. It is a bet that the fee income from uninformed retail flow exceeds the losses from arbitrage. That bet resolves differently on every pool, and most guides do not tell you how to evaluate it.
What a Concentrated Liquidity AMM Gets You (and Costs You)
Concentrated liquidity, introduced in the third major iteration of the leading AMM protocol, was presented as a solution to capital inefficiency. The standard constant-product formula spreads liquidity across every price from zero to infinity, meaning most capital earns no fees because most trading happens within a narrow price range. Concentrated liquidity lets providers specify a range and put 100% of their capital to work within it, potentially achieving 200 times or more the capital efficiency of the original formula for stable pairs trading near peg.
The catch is that the fee amplification and the impermanent loss amplification are the same multiplier.
A position concentrated within plus or minus 5% of current price behaves like a leveraged constant-product pool with roughly 20 times the IL exposure of a full-range position for the same capital. When price exits your range, the position converts entirely to the underperforming asset and earns zero fees. Concentrated LPs who do not actively manage their ranges do not earn 200 times more fees. They earn fees while price is in range, then sit fully allocated to a declining asset while price is out of range, then pay gas to rebalance, and start over.
The further complication is Just-In-Time liquidity. Sophisticated market-making bots have developed a strategy where they deposit a large concentrated liquidity position directly into a pool in the same transaction block as a large incoming swap, capture the fees from that swap, and withdraw in the next block. Current data shows JIT liquidity captures less than 2% of total fees. But modeling of optimal JIT strategies suggests they could increase JIT earnings by up to 69% and reduce passive LP fee income by up to 44% per trade as adoption grows. The fourth-generation version of the leading AMM protocol introduced a hooks architecture explicitly designed to make programmatic liquidity management easier, which institutionalizes this strategy. The beneficiaries are Wintermute-class quantitative firms. The cost is borne by passive retail LPs who cannot execute microsecond block-timed transactions.
What an On-Chain Order Book Actually Requires
An order book sounds simpler conceptually: buyers post bids, sellers post asks, trades happen when they match. The complexity is in the execution environment.
A fully functional on-chain order book requires that placing, canceling, and modifying orders all occur on-chain, at the speed and cost that makes active market-making viable. On Ethereum mainnet, this is not achievable. A passive market maker may post a thousand orders per day to generate fifty fills. At mainnet gas prices, the cancel transactions alone would consume hundreds of dollars. And the 12-second average block time creates a structural problem: every resting limit order is a free option for any informed taker from the moment it goes stale.
The math on the free-option cost: 12-second block time multiplied by 2% hourly volatility on ETH produces roughly 17 basis points of at-the-money optionality on every resting order. Arbitrageurs with sub-second access to CEX prices will pick off stale quotes before the maker can cancel. This is not a solvable problem on a general-purpose L1 with 12-second blocks. It is a property of the block time.
The on-chain CLOBs that have achieved real market share in 2025 all solved this with the same approach: purpose-built chains optimized for trading, with zero gas on order placement and cancellation, with fees charged only on fills as a percentage of notional, and with block times under one second. One leading on-chain CLOB operates with sub-second order finality, claimed throughput of up to 200,000 orders per second, and a consensus mechanism designed to prioritize order cancellations before new fill executions within each block window. This is not an Ethereum application. It is a ground-up redesign of the execution environment with market structure as the primary design constraint.
That platform processed approximately 2.95 trillion dollars in cumulative volume during 2025, generated approximately 844 million dollars in annual revenue, and at peak held over 80% of the decentralized derivatives market. For context, total DEX market share of global spot trading rose from roughly 4% to roughly 20% over the same period.
The Symmetric Problem Nobody Discusses
The debate around AMMs versus order books usually attacks AMMs and defends CLOBs, or vice versa. The honest version acknowledges that on-chain CLOBs trade one set of problems for a different set.
Sequencer and validator trust. An off-chain-matching, on-chain-settling CLOB keeps the order book in validator memory and gossips state across the network before block finalization. Any validator with a marginally faster network connection than consensus sees order flow before the market settles. This is the on-chain version of co-location advantage in traditional high-frequency trading. Several monitoring tools have been built specifically to detect validator-level ordering discrepancies on CLOB-based DEXs because the problem is real and observable.
Sandwich attacks disappear; other forms of MEV do not. On an AMM with a public mempool, a bot can see your pending swap and insert buys and sells around it. On a CLOB without a public mempool, this specific attack vector closes. But priority-fee auctions, inter-validator latency arbitrage, and sequencer-level reordering remain. MEV changes form rather than disappearing.
Centralization at the infrastructure layer. A purpose-built trading chain with proprietary consensus and closed-source matching components achieves performance by removing the constraints that make general-purpose L1s trustless. The validator set may be small by design. The matching engine may not be independently auditable. The tradeoff between performance and decentralization is real, and the 2025 data does not resolve it. Performance wins in markets; the long-run security model is unproven at scale.
Market-making professionalization. AMMs allow anyone to provide liquidity permissionlessly. On-chain CLOBs with sub-second blocks technically also allow this, but topping the order book requires infrastructure comparable to institutional HFT: co-location-equivalent latency to validators, automated quoting and hedging software, and risk management systems. Retail market makers on CLOBs are systematically disadvantaged relative to professional firms the same way they are on centralized exchanges. The permissionless label does not change the competitive reality.
The Architecture That Is Actually Winning
The most interesting development in DEX market structure during 2025 was not on AMMs or on order books. It was on intent-based execution systems, which grew faster than either.
The intent model works like this: instead of specifying how a trade executes, the trader signs a message expressing what outcome they want. A network of competing solvers then races to find the best path to that outcome, routing through AMMs, order books, peer-to-peer matching, private inventory, or any combination. The solver that delivers the best price wins the order and captures a spread. Gas is paid on fill, not on signing. The trader never touches a mempool.
One intent-based system grew from approximately 3 million dollars in cumulative volume at the start of 2025 to over 6 billion dollars by year end. That is a growth rate exceeding 200,000%. A competing intent protocol reached approximately 10 billion dollars in monthly volume in 2025, representing roughly a fivefold increase versus late 2024 and over 35 billion dollars in lifetime settled volume.
The structural advantage of intent-based systems is that they convert the venue selection problem into a solved optimization the user never sees. An AMM user must choose which pool, at which slippage tolerance, on which chain. A CLOB user must choose at which price to enter, how to manage partial fills, and whether their order is safe from adversarial takers. An intent user signs once and receives a fill at the best available price across all venues simultaneously.
The mechanism that eliminates intra-batch sandwich attacks in the leading batch-auction system is worth understanding explicitly. Orders are collected over roughly 30-second windows and auctioned to solver competitors who must submit a uniform clearing price per token pair for the entire batch. Because no order has priority over another within the batch, and because the clearing price is uniform, there is no profitable position to insert before or after a specific trade. The attack surface that costs Ethereum users tens of millions of dollars per year on standard AMM swaps simply does not exist in this execution model. The residual risk is solver concentration: one solver captures over 50% of volume in the leading system by market share, which introduces counterparty dependency the decentralization framing tends to understate.
Where Each Architecture Actually Wins
The take that frames this as a binary choice misses the empirical reality of 2025, which is that all three architectures are growing in absolute terms while competing for different use cases.
AMMs win on permissionless listing and the long tail. Any token can launch a pool. No approval, no market-maker relationship, no minimum liquidity threshold. This is structurally irreplaceable. A memecoin launching with 10,000 dollars of liquidity cannot list on a CLOB. It can list on an AMM in one transaction. The bonding curve variant of AMMs, used by the dominant Solana token launchpad, bootstraps price discovery from zero by selling tokens against a deterministic curve with no LP required at all. Over 800 million tokens were created on that platform in 2025. None of them could have bootstrapped on an order book.
CLOBs win on derivatives and blue-chip spot. Perpetual futures require precise margin accounting, partial fills, reduce-only orders, and sub-second liquidations. These are order book primitives. An AMM cannot run a perpetual contract. Every major on-chain derivatives venue uses a CLOB architecture. For large spot trades on liquid assets, a deep order book also produces better execution than any AMM: no price impact from pool depth constraints, predictable fill prices, and full position control.
Intent systems win on protected execution for mid-size spot trades. For a trader executing between 1,000 and 1,000,000 dollars in spot on a liquid pair, intent-based systems consistently produce better net prices than either raw AMM swaps or manually-placed CLOB orders. The solver competition internalizes MEV that would otherwise be extracted, and peer-to-peer matching bypasses LP fees entirely on a portion of volume. This is not theoretical. The execution quality data supports it.
The graduation pipeline is the underappreciated synthesis. Long-tail tokens launch on bonding curves or AMMs, where permissionlessness matters more than execution quality. If they survive and develop sufficient liquidity, they migrate to AMM pools with real market depth. If they achieve blue-chip status, they list on on-chain CLOBs where professional market-makers provide tight spreads. This is the on-chain equivalent of the OTC-to-exchange progression in traditional markets, and it means AMMs and CLOBs are increasingly complements operating at different phases of an asset's lifecycle rather than competitors solving the same problem.
What This Means If You Are Trading
The architecture underneath your trade determines your costs in ways the interface never shows you. Three practical implications follow directly from everything above.
First, if you are providing liquidity on a volatile pair, your realized return is not the advertised APY. The advertised figure is gross fees. The actual return is gross fees minus LVR minus gas minus the opportunity cost of capital concentrated out of range. The only way to evaluate an LP position honestly is to compare it against simply holding the underlying assets, not against a zero-risk benchmark. On high-volatility pairs, the comparison often favors holding.
Second, if you are executing a mid-size swap on a liquid token, an intent-based aggregator will almost always produce better net execution than a direct AMM swap, at the same or lower effective cost. The MEV that would be extracted from your public mempool transaction is instead competed away by solver networks and returned to you as price improvement. There is no speed or cost argument for routing around this.
Third, if you are trading derivatives or need resting limit orders with precise fill guarantees, an on-chain CLOB on a purpose-built chain is the only architecture that delivers CEX-equivalent order types and execution. Concentrated liquidity range orders, keeper-monitored off-chain orders, and intent-based systems do not substitute for a true price-time priority order book when the order type requirements are strict.
The architecture is not the product. The architecture is the set of tradeoffs you accept when you sign the transaction. Knowing which tradeoffs you are accepting is the most practical thing a DEX trader can learn in 2026.
