Order Book vs AMM (CLOB vs AMM): How the Two DEX Models Differ

On an order book (CLOB), you trade against other traders’ bids and asks at prices you can see before you click; on an AMM (automated market maker), you trade against a liquidity pool that prices every trade with a formula. That single difference, who your counterparty is, decides what a trade costs: the pool’s price moves against you roughly in proportion to your size ÷ the pool’s depth, while a book charges the spread plus a fee until you run through the visible orders.

Fees from each venue’s fee page (registry). The $1M figures are our simulations on public data read October 19, 2026, 06:06 UTC: DefiLlama for the Uniswap v2 USDC-WETH pool on Ethereum, the Hyperliquid l2Book and Lighter orderBookOrders endpoints for the ETH perps. GMX mechanics from docs.gmx.io. Depth changes by the second.

1. What is an order book (CLOB)?

An order book is the live list of every open buy order (bids) and sell order (asks) in one market, sorted by price. A central limit order book (CLOB) is the version almost every exchange runs: one shared book per market, limit orders rest in it at the price their owner chose, and a matching engine fills each incoming order against the best price first and, at the same price, the oldest order first. That is price-time priority. Binance, Coinbase and virtually every other centralized exchange work this way, and so do the order-book DEXs: Hyperliquid, Lighter, dYdX and Aster’s Pro mode.

How an order book works

• Makers post limit orders that rest on the book (“I will sell 1 ETH at $2,700”). They supply the liquidity.
• Takers send market orders, or limit orders that cross the spread, and fill immediately against the best resting orders. They consume the liquidity.
• The highest bid and the lowest ask form the top of the book; the gap between them is the spread, and a market order that is larger than the best level “walks” down the book to worse prices.
• Fees follow the same split: makers pay less than takers, and on some venues large makers are paid a rebate.

What an order book tells you

The book shows the price and size at which people are willing to trade right now: the spread is the cost of trading immediately, the size stacked near the price (the depth) tells you how large a market order can be before it moves the price, and the number of orders per level separates a crowd from one large player. It is a snapshot of intentions, which can be cancelled. The order flow, the orders that actually arrive and trade (the trades tape), shows what people are doing. Section 10 reads a real Hyperliquid BTC book level by level.

2. What is an AMM?

An automated market maker (AMM) replaces the book with a smart contract that holds two (or more) tokens, the liquidity pool, and a pricing formula. Anyone can deposit both tokens and become a liquidity provider (LP); traders swap against the pool; and after every trade the formula moves the price, because the pool now holds more of what you paid in and less of what you took out. In Uniswap’s classic constant-product pool the rule is x × y = k, and the price is simply the ratio of the two reserves (the math is in section 6). A swap fee on each trade is paid to the LPs. Uniswap, Curve, PancakeSwap, Aerodrome and Raydium are AMMs.

3. Order book vs AMM: key differences

If you are new to DEXs, start with What is a DEX?, which covers wallets, custody and first steps. The table puts the two models (CLOB vs AMM) side by side; the fee figures come from the venues’ fee pages and the cost figures from our dated snapshots below.

4. Which should you use?

Go down the list and stop at the first line that fits your trade:

1. Trading perpetual futures or using leverage? An order book: Hyperliquid or Lighter. Perps need stops, liquidations and funding, and on a deep book the spread is a fraction of a basis point.
2. Moving $100k or more in one go? An order book, or a TWAP on one. On October 19, 2026 a $1M ETH buy cost about 0.06% all-in on Hyperliquid against about 10.3% in a $19.9M Uniswap v2 pool. If you must use a pool, an aggregator splits the order across deeper pools.
3. Need limit, stop-loss, take-profit or TWAP orders? An order book. AMMs only offer a swap with a slippage limit.
4. Swapping a token you hold for another, once, at small size? An AMM or an aggregator, with slippage set by hand and MEV protection on (the sandwich checklist). At small size the pool fee is the whole cost.
5. Buying a brand-new or long-tail token? An AMM: no market maker quotes it, so it only exists in a pool. Check the contract address first; permissionless listing means fake tickers.
6. Swapping stablecoin to stablecoin? A stable-curve AMM (Curve) or a concentrated pool built for pegged pairs.
7. Earning passive yield? An AMM LP position (fees minus impermanent loss) or an order-book venue’s vault such as HLP (market-making PnL, with a lock-up). See LP economics.

For perpetual futures the choice is really between order books. Hyperliquid has the deepest book and a long market list; Lighter charges 0% on Standard accounts and proves its matching on Ethereum; dYdX is a fully validator-run chain. The live volume league table is on the perp DEX ranking, the buyer’s guide on best DEXs in 2026, and the fee comparison with Binance and Bybit in DEX vs CEX fees. For a first trade, follow how to use Hyperliquid; every fee tier is in Hyperliquid fees explained.

5. Why DeFi started with AMMs, and why perps moved to order books

On Ethereum mainnet every action is a transaction that costs gas and waits for a block. A market maker cancels and replaces quotes every time the price moves, so a fully on-chain order book would have charged makers for every requote and left them too slow to avoid being picked off by faster traders. The result was wide spreads and empty books. An AMM sidesteps the problem: there is no market maker to pay, one transaction per trade is enough, and the formula requotes automatically after each swap. That is why the first wave of DeFi trading went to pools such as Uniswap and Curve, and on-chain order books stayed a niche.

Two things changed. First, throughput: chains built for trading made placing and cancelling orders cheap or free. Hyperliquid runs its own layer 1 in which the book is chain state and orders cost no gas; dYdX runs its own Cosmos chain; Lighter matches on a sequencer and proves every batch on Ethereum (section 9). Second, perpetual futures: leverage multiplies trading costs, so a perp trader who turns over their account many times cares about the gap between a fraction of a basis point and a percent, and needs stops, liquidations and funding that a simple swap pool does not offer. Perps therefore moved to order books, while spot swaps of long-tail tokens stayed on AMMs.

Are AMMs replacing order books?

No, and order books are not replacing AMMs either. The two designs are converging. Uniswap v3 and v4 let LPs place liquidity in narrow ranges, which behave like limit orders. Hyperliquid’s HIP-2 runs a formula-driven market maker that places orders on its spot book. Jupiter, whose main perps use a pool, launched its newer Beta perp markets on an order-book engine. Each model is winning the trades it is built for.

6. How an AMM prices a trade: x × y = k

A Uniswap-style pool holds two tokens, say x ETH and y USDC. Uniswap’s docs describe the rule: the pool “prices tokens algorithmically using a constant product formula”, x × y = k, where k “must stay constant (or increase) after every trade”. The current price is simply the ratio of the reserves, y ÷ x.

When you buy ETH with Δy USDC, the pool must keep the product constant, so it pays out

ETH received = x × Δy ÷ (y + Δy)

That gives two rules of thumb you can apply to any constant-product pool (before the pool fee):

• Your average price is worse than the quoted price by Δy ÷ y, i.e. your trade size divided by the pool’s reserve on the side you pay in. Trade 1% of that reserve and you pay about 1% more than the spot price.
• The pool’s price after your trade moves by about twice that, (1 + Δy ÷ y)² − 1. Arbitrageurs then trade it back toward the market price and keep the difference, which is part of how LPs lose money (section 14).

The pool fee is added on top: 0.30% on every Uniswap v2 pool; 0.05%, 0.30% or 1% on standard v3 pools (plus a 0.01% tier); and any level the pool creator chooses on v4. The fee goes to liquidity providers, minus a protocol share on some pools.

7. Worked example: $10k, $100k and $1M ETH buys

Real pool, real books, same minute. At October 19, 2026, 06:06 UTC, DefiLlama listed the Uniswap v2 USDC-WETH pool on Ethereum with $19.9M of liquidity, so roughly half of that in USDC on the side you pay into. Hyperliquid’s ETH perp had a mid price of $2,432.85, a $0.10 spread and $9.5M of asks within 0.080% of mid on its 20 best levels; Lighter’s ETH perp showed $24.3M on its 250 best asks. We simulated market buys against all three.

October 19, 2026, 06:06 UTC. Pool: DefiLlama yields API, Uniswap v2 USDC-WETH (Ethereum), split 50/50, constant-product math with the fee on the input. Books: Hyperliquid l2Book ETH (20 levels) and Lighter orderBookOrders ETH (250 asks), walked level by level. Gas, funding and the book changing during execution are ignored. A snapshot, not a guarantee: depth changes by the second.

Earlier snapshots, same method: on September 25, 2026 the pool held $20.89M and a $1M buy cost about 9.9% against about 0.08% on Hyperliquid; on September 26 it was 9.8% against 0.05%. The ratio has stayed in the hundreds: on October 19, 2026 the pool was about 172 times more expensive than Hyperliquid’s book for the same $1M.

Three lessons follow. First, at small size both are cheap and the fee dominates: 0.30% on the pool versus 0.045% plus a sliver of spread on Hyperliquid, and almost nothing on Lighter Standard. Second, AMM cost grows linearly with size while a deep book stays nearly flat until you exhaust the visible levels. Third, nobody sensible routes $1M through a v2 pool: an aggregator would split it across deeper concentrated-liquidity pools and other venues. The point is structural. A pool’s depth is fixed by how much capital LPs have parked; a book’s depth is set by how many market makers compete to quote, and on a busy perp market that is far more.

8. Concentrated liquidity (Uniswap v3 and v4)

The flaw in plain x × y = k is that liquidity is spread across every price from zero to infinity, and most of it never trades. Uniswap’s docs give the example of the v2 DAI/USDC pair, which used only about 0.50% of its capital for trading between $0.99 and $1.01, where almost all the volume happens.

Uniswap v3 (and v4, which keeps the same implementation) lets each LP choose a price range. The price space is cut into ticks, each a 0.01% price step; a position is active only while the price is between its lower and upper tick. Within the active range the pool behaves like a much bigger constant-product pool, so traders get deeper liquidity near the mid and LPs earn more fees per dollar.

The numbers show the effect. On September 25, 2026 DefiLlama showed the v3 0.05% USDC-WETH pool on Ethereum turning over $56.7M a day on $99.1M of liquidity (0.57× its TVL), while the v2 pool turned over $0.83M on $20.9M (0.04×). Traders route to where the depth near the price is.

The costs are real, though:

• Out of range = no fees. The docs are explicit: when the price exits a position’s interval, “the position’s liquidity is no longer active and no longer earns fees”, and the position ends up holding only one of the two tokens.
• Active management. Narrow ranges earn more but need re-centering, which costs gas and realises the loss.
• Amplified impermanent loss. Concentrating capital concentrates the loss from price moves inside the range as well.

Concentrated liquidity also blurs the line with order books: a very narrow range just above the current price behaves like a sell limit order (Uniswap calls this a range order). An order book is, in a sense, the limit of that idea: every maker chooses exactly one price.

9. How order-book DEXs match orders: Hyperliquid, Lighter, dYdX

Every order-book DEX has to answer two questions a centralized exchange never faces: where does the book live, and who decides the order in which orders arrive. The designs below answer them differently.

Hyperliquid: the book is the chain state

Hyperliquid’s docs describe HyperCore state as including “an order book for each asset”, working “similarly to that of centralized exchanges”, with orders “matched in price-time priority”. Margin checks happen when an order is placed and again for the resting side at the moment of matching. The unusual part is that the mempool and consensus are aware of order-book transactions: within each block, actions that send no GTC or IOC order (post-only orders among them) go first, then cancels, then actions that send GTC or IOC orders. Processing cancels before new takers protects market makers from being picked off on stale quotes, which is one reason makers can quote tight spreads. As of October 20, 2026 the API showed 178 live perp markets, $4.03B of 24-hour volume and $11.05B of open interest. More: Hyperliquid review, order types.

Lighter: an off-chain matcher you can verify

Lighter takes the opposite route. A sequencer orders transactions first-in-first-out and executes them at low latency; a prover then generates succinct proofs of every batch, and smart contracts on Ethereum verify each proof before updating the state root. If the sequencer ignores exit requests submitted on Ethereum within the deadline, Escape Hatch mode freezes the contract and users withdraw with proofs built from the data Lighter posted to Ethereum. You trust the operator for liveness and ordering, not for correct execution. Standard accounts pay 0% maker and 0% taker. More: Lighter review, Hyperliquid vs Lighter.

dYdX: an in-memory book on every validator

dYdX Chain is a Cosmos SDK L1. Its docs say validators store orders “in an in-memory orderbook (i.e. off chain and not committed to consensus)” and gossip them to each other. The validator chosen as block proposer matches crossing orders and puts the fills in its proposed block, which is committed once validators with two-thirds of the stake approve. Only fills and settlement are on-chain; the resting book is not. Base fees are 0.050% taker and 0.010% maker. More: dYdX review, Hyperliquid vs dYdX.

10. How to read and trade an order book

If you trade perps on a DEX you will be looking at a book, so here is how to read one, using a real snapshot of Hyperliquid’s BTC perpetual pulled from the public l2Book endpoint at 17:19:49 UTC on September 25, 2026.

Source: Hyperliquid API l2Book for BTC, full precision, top 3 of 20 levels per side, 17:19:49 UTC September 25, 2026. The “orders” column is the API’s n field.

Bid, ask, spread and mid

Bids are resting buy orders, asks resting sell orders. The best bid (83,897) and best ask (83,898) form the top of book; their difference is the spread, here one $1 tick or about 0.12 basis points; their average is the mid, 83,897.5. A market buy pays the ask, a market sell receives the bid. Read spreads in basis points so you can compare markets: under 1 bp (BTC, ETH in normal conditions) the spread is noise next to the fee; when it grows wider than the 4.5 bps taker fee, liquidity has stepped away, usually around news or liquidation cascades, and a limit order is the better tool.

Depth: how much the market can absorb

The API returns at most 20 levels per side; at $1 ticks that covered only $83,878–$83,919 on BTC, with $3.67M of asks and $2.49M of bids. For a wider view, the nSigFigs parameter aggregates levels (2 to 5 significant figures). With 4 significant figures ($10 buckets, 17:20:19 UTC) the book held roughly $9.3M of asks and $8.2M of bids within 0.05% of mid, and $26.4M / $23.5M within 0.1%. Walking the raw ask side, a $10,000 or $100,000 market buy would have filled entirely at the best ask (0.06 bp above mid, the half-spread) and a $1M buy would have averaged about 0.67 bp above mid, roughly $67 of price impact against a $450 taker fee. On BTC, below seven figures, the fee is the cost; on a thin altcoin the order is reversed. A practical rule from this: keep a market order to a small fraction of the depth within 0.1% of mid, and use limit orders, a scale ladder or TWAP beyond that.

Maker vs taker

An order that rests on the book and is filled later is a maker order; one that crosses the spread and fills immediately is a taker. On Hyperliquid tier 0 that is 0.015% vs 0.045%, so on a one-tick book “buying the bid” means joining the bid and waiting, which costs a third of the fee but may not fill. Use the ALO (post-only) time-in-force when you want to be sure you never pay taker: the order is rejected instead of crossing. On Lighter’s Standard account both sides are 0%, so the spread is the only cost of crossing.

Queue priority

Matching is price-time priority: a better price always fills first, and at the same price the earlier order fills first. The 4.27 BTC at the best ask above is a queue of 30 orders; a new sell at 83,898 joins the back of it. Improving the price by one tick puts you at the front of a new level. Hyperliquid adds one twist: inside each block, actions without GTC/IOC orders (post-only orders among them) go first, then cancels, then GTC/IOC orders, so a maker can pull a stale quote before a taker in the same block hits it. The docs’ latency notes put cancel/ALO end-to-end latency at about 380 ms.

Reading the book without fooling yourself

• Look at the n field. 4.27 BTC from 30 orders is a crowd; the same size from one order is one decision that can be cancelled in one action.
• Walls are free to pull. There is no gas per order or cancel, and cancels are processed before takers, so a large resting order is only meaningful once it has absorbed trades in the tape without price moving through it.
• Imbalance is a filter, not a signal. Imbalance = (bid size − ask size) ÷ (bid size + ask size) over a fixed window, from −1 (only sellers) to +1 (only buyers). At 17:19:49 the 20 visible levels held $2.49M of bids against $3.67M of asks, an imbalance of about −0.19; that kind of reading changes within seconds as market makers requote, so only a sign that holds over many consecutive snapshots means anything.
• Pull it yourself. POST {"type": "l2Book", "coin": "BTC"} to https://api.hyperliquid.xyz/info (no API key). The response’s levels field holds two lists, bids then asks (with a time stamp alongside), each level with px (price), sz (size in coins) and n (number of orders). Multiply sz × px before adding levels up, and add "nSigFigs": 4 for the aggregated view used above.
• Stream it for anything timing-sensitive. The WebSocket l2Book subscription pushes the book on changes; polling REST snapshots is too slow for signals. See the WebSocket guide.

More on execution cost: price impact and slippage on Hyperliquid and DEX speed and fill quality.

11. Oracle-priced pools: the third model for perps

Perpetual DEXs that do not run a book usually do not use x × y = k either. They use a pool as the counterparty and an oracle for the price. Jupiter Perps is the largest example on Solana: traders borrow from the JLP pool, trades on SOL, ETH and wBTC execute at the oracle price “without orderbook slippage”, and a price impact fee replaces the slippage a book would have charged. The pool earns 75% of the fees and wins when traders lose. The trade-offs are an hourly borrow fee, a small set of markets and a dependence on oracle quality. We compare it directly with Hyperliquid in Hyperliquid vs Jupiter Perps, and the older GMX version in Hyperliquid vs GMX.

All three designs side by side:

Telling detail: Jupiter’s newer “Beta” perp markets (JUP, HYPE, ZEC and stock perps, open to any wallet since the JUP-stake gate was dropped) do not use the pool at all. The docs say they are “powered by GUM, an orderbook trading engine”, with maker and taker fees and hourly funding. Even pool-first venues reach for an order book when they add markets.

12. Hybrids: AMM + order book

• Velocity (formerly Drift, Solana). Drift’s relaunch runs at app.velocity.exchange, and Velocity’s docs call it a fork of Drift v2. A taker order first opens a short just-in-time auction for market makers; then the program walks crossing maker orders from its decentralized limit order book (an off-chain view of orders stored on-chain) and lets its own virtual AMM fill whatever is left, or step in wherever it prices better than a maker. Its docs say a PropAMM + CLOB redesign is in progress and not live yet.
• Hyperliquid HIP-2 (Hyperliquidity). A formula-driven strategy, built into the chain with no operator, that places orders on the spot order book of newly deployed tokens: levels 0.3% apart, updated no more than once every 3 seconds. It gives thin tokens AMM-like “always a quote” liquidity without leaving the book.
• Aster. A product-level hybrid: its Perpetuals (Pro) mode is an order book, while Shield Mode is a simplified AMM-style interface for private, high-leverage trades. See the Aster review.
• Aggregators. Jupiter swap, 1inch and CoW Swap route a spot swap across many AMMs and, where available, RFQ market makers. They are the practical way to get order-book-like prices on an AMM chain.

13. Sandwich attacks: the AMM-only risk

A sandwich attack is a bot buying just before your AMM swap and selling just after it, so you fill at a worse price and the bot keeps the difference. It needs two things: a pending transaction others can see, and a price curve that your own trade moves in a predictable way. An AMM swap on a public mempool offers both. An order book offers neither in that form: you trade against resting orders at stated prices, and a limit order cannot fill worse than its limit. That is why sandwiches are an AMM problem, and why the defences below are mostly about swaps.

How a sandwich works

When you submit a swap you also sign a minimum output, your slippage tolerance. If the transaction is visible before it is included in a block, a searcher bot can:

1. Front-run: buy the token you are buying, placing its transaction just before yours. The pool price rises.
2. Let your swap execute at the higher price, still inside your slippage tolerance, so it does not revert.
3. Back-run: sell right after you, into the price your trade pushed up.

The bot sizes its front-run so your swap lands just above your minimum output, so the tolerance you set is the budget you hand to the attacker. The same traits show up in every study of who gets sandwiched: a large swap relative to the pool, a loose slippage tolerance, a thin or low-fee pool, a transaction visible in a public mempool, and flow that repeats on a schedule. Memecoin buys tick most of those boxes at once.

A real case: jaredfromsubway.eth. The best-known sandwich bot on Ethereum was measured by the MEV analytics firm EigenPhi. Between February 27 and May 8, 2023 it made 238,000 sandwich attacks on more than 106,000 victims, earning $40.65M in revenue, paying about $34.35M in gas and keeping roughly $6.3M (EigenPhi, May 10, 2023). The gas bill shows why sandwiches concentrate on trades with room to take: the bot needs your slippage to pay its own costs.

Worked example: a $100,000 ETH buy

Same Uniswap v2 USDC-WETH pool as in section 7 ($19.9M, 0.30% fee, ETH at $2,432.85, October 19, 2026). Without an attacker, a $100,000 buy receives about 40.57 ETH, 1.30% worse than spot (price impact plus the fee). Now add a bot that front-runs exactly enough to push you to your minimum output:

Our arithmetic on the October 19, 2026 pool size: constant-product pool, 0.30% fee on each leg, before gas and builder tips. The gap between your loss and the bot’s profit goes to liquidity providers as fees.

• Your loss scales with your tolerance. Moving from 3% to 0.5% cuts the damage from about $2,961 to about $494 on the same trade.
• Small swaps in high-fee pools are poor targets. Repeating the simulation for a $10,000 swap, the bot loses money at every tolerance from 0.1% to 3%, because it pays the 0.30% pool fee on both legs. Low-fee pools (0.05%, 0.01%) and large swaps are where sandwiches pay.
• Gas and tips cut into the bot’s margin, which is why sandwiches cluster on large trades, thin pools and loose tolerances.

Sandwich and MEV protection checklist (anti-MEV settings)

1. Set slippage by hand, and tight. Around 0.5% or lower on majors, 0.1% or less on stablecoin pairs. If a swap fails, wait or split it instead of raising the tolerance: EigenPhi documented jaredfromsubway.eth catching traders within a few blocks after they retried a failed memecoin swap with a looser minimum (EigenPhi, July 2023).
2. Turn on MEV protection (a private RPC). MetaMask’s protection is automatic and free for Ethereum transactions on its default Infura RPC and for all Solana transactions (and BNB Smart Chain with Smart Transactions on); in other wallets add https://rpc.flashbots.net/fast (Flashbots Protect, hidden “from frontrunning and sandwich bots”) or https://rpc.mevblocker.io (MEV Blocker, which pays you 90% of the back-run profit), chain ID 1. On BNB Chain, PancakeSwap MEV Guard is https://bscrpc.pancakeswap.finance, chain ID 56. In MetaMask a custom RPC overrides Smart Transactions, so use one or the other.
3. Route larger swaps through an intent router. CoW Swap settles batches at one uniform clearing price, 1inch Fusion lets resolvers compete in a Dutch auction, and UniswapX has solvers fill your signed order (Uniswap’s API uses it above 300 USDC or when it quotes at least 0.2% better). You never broadcast a raw swap, but the worst price you sign is still your worst price.
4. On Solana, use a router with protected landing (Jupiter Ultra) and switch on your trading bot’s Anti-MEV setting. Developers can add a jitodontfront account so Jito’s block engine rejects bundles that put anything before the transaction.
5. Split size, prefer deep pools, and do not repeat the same trade on a timer. Predictable flow is easy to hunt.
6. For leverage, use an order-book venue. Limit and Post Only orders cap your price; lower Hyperliquid’s 8% default max slippage on market orders in thin markets (Hyperliquid slippage and price impact settings). Trade perps on Hyperliquid’s order book with code HOLYGRAIL for 4% off fees.

Should MEV protection stay on? Yes, for nearly every swap. Infura says its protection adds on average 1 extra slot and hands a transaction that nobody picks up within 2 slots back to the public mempool; MetaMask itself says protection “is not always guaranteed”. You also trust a relay and its builders with your transaction. Those costs are small next to the table above.

How protected is “protected”?

A September 2026 paper, “No Place to Hide: An Analysis on Protected Order Flow Sandwich Attacks” (arXiv 2609.28115), measured sandwiches against transactions that were supposed to be protected, from July 2023 to June 2026. On Ethereum it found 30,607 sandwiches on protected order flow, by 7 bots, with order-flow auctions the main leak; on Solana, 28.0 million sandwiches by 8,631 bots worth $383.4M gross, concentrated on a few trading front-ends and Telegram bots; on Base 1,889; on Arbitrum none. A few tens of thousands of attacks over three years on Ethereum’s protected flow is far less than the public mempool, so private RPCs remain worth using, but tight slippage stays your hard limit.

Why order-book DEXs are not sandwiched

• Order book: no sandwich in that form. A limit order cannot fill worse than its limit, so there is nothing to squeeze. A market order on a book still has a maximum-slippage setting, but moving a deep book costs the attacker real money. The remaining question is sequencing: who decides whether your order or someone else’s lands first. Hyperliquid answers with deterministic in-block ordering (cancels before new GTC and IOC orders), Lighter with a first-in-first-out sequencer whose execution is proven on Ethereum, dYdX by letting the proposer of each block do the matching. Speed is still for sale on Hyperliquid (IOC orders can pay a priority fee), so a paying taker can reach the same resting liquidity first; a limit order caps your price regardless.
• Oracle-priced pools: no book to front-run, but oracle risk. You trade at the oracle price, so a sandwich cannot move your fill. The attack surface moves to the oracle (stale or manipulated prices) and to the delay between your request and the keeper’s execution; GMX’s docs say its slippage setting only has to cover that oracle move.

Architecture from each venue’s official docs (re-read October 19, 2026); attack counts from arXiv 2609.28115.

Other MEV you may meet

MEV (maximal extractable value) is any profit a builder or bot can take by inserting, reordering or excluding transactions around yours. Sandwiches cost swappers most directly; the others matter less, or not at all, to a trader:

How secure are AMMs, and order-book DEXs?

The constant-product formula is simple and well understood; the risks sit around it. Contract risk is highest in new or forked pools and in Uniswap v4 hooks, which attach custom code to a pool. Token risk comes with permissionless listing: anyone can create a pool for a fake token with a familiar ticker, or for a token with transfer taxes or a blacklist. MEV is covered above, and LPs carry impermanent loss (next section). Order-book DEXs trade those risks for others: the bridge or deposit contract that holds your collateral, the validators or sequencer that order your transactions, and, for perps, liquidation and auto-deleveraging rules. On either model, use the official front-end, check contract addresses, keep slippage tight and never approve unlimited spending to a contract you do not know.

14. LP economics and impermanent loss

Someone has to provide the liquidity, and each design pays them differently.

AMM LPs: fees minus impermanent loss

An AMM LP earns the pool fee on every trade. A September 25, 2026 snapshot (earlier than section 7’s): the v2 pool had $0.83M of daily volume × 0.30% ≈ $2,480 a day on $20.9M, which DefiLlama showed as a 4.33% fee APY. The v3 0.05% pool: $56.7M × 0.05% ≈ $28,400 a day on $99.1M, a 10.45% fee APY. Against that stands impermanent loss (IL): the pool automatically sells the token that is rising and buys the one that is falling, so an LP ends up worse off than someone who just held both tokens. In a 50/50 constant-product pool, IL for a price ratio change r is 2√r ÷ (1 + r) − 1:

So at that 4.33% fee APY, a v2 LP who stays in for a year beats simply holding only if ETH ends the year less than about 80% higher or 45% lower against USDC than where it started (IL of 4.3% at those points). Fee APY moves with volume: on October 19 at 06:06 UTC the same pool did $438k of daily volume and DefiLlama showed a 2.41% fee APY, which narrows the break-even band to about 56% higher or 36% lower. In a concentrated position the same price move inside your range costs more, and outside your range you earn nothing. The loss is “impermanent” only if the price comes back.

Order-book liquidity: market makers and vaults

On an order book, liquidity comes from market makers who choose their own prices and pull quotes when they want. They are paid by the spread and, on many venues, by low maker fees or rebates (Hyperliquid’s largest makers earn a rebate of up to 0.003% of volume; Lighter Standard makers pay 0%). Retail users can join through a vault: Hyperliquid’s HLP vault runs market-making and liquidation strategies and held $179.1M at a 3.5% APR as of October 20, 2026 (vaultDetails API). HLP has a 4 days deposit lock-up, and it can lose money in violent markets. See Hyperliquid vaults and how perp DEX vaults (HLP, LLP, ALP) work.

Perp pool LPs: the house

In an oracle-priced pool such as JLP, LPs are the counterparty to every trader. They collect most of the fees (75% on Jupiter) plus traders’ losses, pay out traders’ profits, and hold the pool’s basket of SOL, ETH, wBTC and stablecoins. It is closer to owning a casino with a crypto balance sheet than to AMM market making.

15. FAQ

Disclosure: HOLYGRAIL (Hyperliquid) and 718610TD (Lighter) are our referral codes; we earn a share of referred accounts’ fees at no cost to you. We hold no codes for the swap and MEV-protection tools mentioned. Architecture descriptions come from each venue’s official documentation; fees come from the venues’ fee pages; pool and order-book snapshots were fetched from DefiLlama and the Hyperliquid and Lighter APIs on the dates stated, the sandwich table is our own simulation, and Hyperliquid market and vault figures are refreshed at every site build. Not investment advice.