Between about 21:16 and 21:27 UTC on 10 October 2025, Hyperliquid’s auto-deleveraging machinery fired again and again. According to a paper by Tarun Chitra of Gauntlet, about 2.1 billion dollars of positions were closed in those 12 minutes. Traders with profits on the books saw them cut at the previous mark price, with no margin call and nobody to phone. That is the moment a perpetual futures blockchain, where users swap tokens directly from their own wallets without handing custody of their funds to a company.">DEX shows what it really is: a machine with a rule for deciding who pays when a loss has no owner.
The product is the same on every venue. The answer to “who is on the other side of my trade?” is not. On-chain, it has four answers: other traders in an order book, liquidity providers in a smart contract holding reserves of two or more tokens, supplied by users, that a decentralized exchange uses to execute trades automatically at a price set by a formula.">pool, a virtual curve, or a matching engine nobody outside can inspect. Each answer moves the risk to a different person, and each has now been through at least one bad day.
The product itself gets one paragraph. A perpetual future is a leveraged contract with no expiry date. A periodic funding payment between longs and shorts keeps its price near the spot price: on Hyperliquid, whose documentation describes the mechanism in detail, funding is paid every hour. You post margin, and if your equity falls below the maintenance level the position is closed, the logic covered in our guide to liquidations. The lesson also follows on from how to provide liquidity, because anyone who deposits in a perp vault becomes a liquidity provider with a far less predictable counterparty than a swapper.
A perpetual futures DEX with a book on its own chain
The first answer copies the centralized exchange: traders post bids and offers, and a matching engine pairs them. The other side of your trade is another trader. Doing that on a general-purpose chain is expensive, because market makers cancel and replace orders constantly and every update would pay gas. So the two best-known book venues built their own chains.
Hyperliquid: the book inside the consensus
Hyperliquid runs its own layer 1, secured by a consensus called HyperBFT. Its execution splits in two: HyperCore, which holds the perp and spot order books and the margin state, and HyperEVM, a general-purpose smart contract layer. The documentation says every order, cancel, trade and liquidation is handled on-chain with one-block finality, and puts current capacity at about 200,000 orders a second. That number comes from the team’s own docs, so read it as a design claim.
Prices for margin do not come from the book alone. Validators publish an oracle price every 3 seconds, a weighted median of spot prices from several large exchanges and Hyperliquid’s own spot market, and the final figure is a stake-weighted median of what they report. If oracles are new to you, our advanced lesson on DeFi oracles explains why this input matters so much.
Hyperliquid also has a pool of sorts. HLP is a vault open to deposits, with a 4-day lock-up, which runs market-making strategies and contains the liquidator vault. When an account falls below two thirds of its maintenance margin and the book has not absorbed it, HLP’s liquidator can take the position over. That is a design choice with a cost: in normal times you trade against other people, but in a failed liquidation a vault of depositors inherits the trade.
Decentralization is the open question, and a later lesson treats it in depth. As of 30 October 2026, Hyperliquid’s validator API listed 27 active validators, and the five branded Hyper Foundation held about 47% of active stake by our addition. Quorum needs more than two thirds of stake.
The venue’s size is not in doubt, though the exact number depends on who measures it. DefiLlama’s Hyperliquid Perps entry showed about 7.71 billion dollars of open interest on 30 October 2026, and that is the figure used in the comparison table below, because the same source covers every venue there. Summing open interest times mark price over the 235 markets returned by Hyperliquid’s own API, 56 of them flagged as delisted, gave about 10.98 billion dollars on the same day. We have not reconciled the two, so neither should be read as a correction of the other. Neither figure counts tradeXYZ, listed separately on DefiLlama at about 3.88 billion dollars of open interest on the same chain. The same API showed about 16.5 billion dollars of notional volume over 24 hours, with a maximum of 40x on BTC, and DefiLlama put fees over 30 days at 80.9 million dollars.
Since the HIP-3 upgrade, third parties can also deploy their own perp markets on Hyperliquid, staking 500,000 HYPE that validators can slash and keeping up to 50% of the fees those markets earn.
dYdX v4: validators hold the book in memory
dYdX moved from an Ethereum layer 2 to a standalone chain built with the Cosmos SDK, whose alpha mainnet launched on 26 October 2023. Its design differs from Hyperliquid’s in one precise way. The dYdX documentation says validators store orders in an in-memory book that is off-chain and not committed to consensus. When orders match, the block proposer includes the match in its block, and the block is final once validators with two thirds of the stake approve it. So the trades are on-chain, while the resting orders are not. Hyperliquid’s docs take the opposite position and say HyperCore does not depend on off-chain books.
As for the state of dYdX in 2026, the numbers are modest. DefiLlama showed about 40.8 million dollars of open interest for dYdX V4 on 30 October 2026, about 70.3 million dollars of total value locked and about 0.24 million dollars of fees over 30 days. The chain is live, and on that date its open interest was about 0.5% of Hyperliquid’s on DefiLlama’s count. Maximum leverage on the dYdX indexer that day was 50x on BTC and ETH, with 20x on SOL.
The pool against the traders: GMX, Jupiter Perps and gTrade
The second answer removes the book. Traders open positions at an oracle price against a shared pool of assets, so the depth you see does not depend on a market maker’s quotes. The pool is the counterparty. If traders lose, it gains; if they win, it pays.
GMX: from GLP to GM and GLV
GMX began with GLP, one shared pool on Arbitrum and Avalanche. On Arbitrum its exposure was about 25% wrapped bitcoin, 25% ether and 50% stablecoins, going by the compensation plan described below. Its second version is built on GM and GLV pools. The GMX documentation says every order is routed against these pools, at an oracle index price and without an order book. GMX now runs on Arbitrum, Avalanche and MegaETH, offers up to 100x, and pays liquidity providers 63% of fees on Arbitrum and Avalanche.
Fees show how a pool defends itself against one-sided trading. The position fee is 0.04% of size when a trade narrows the gap between long and short open interest, and 0.06% when it widens the gap. Funding moves with that imbalance, adjusting gradually rather than at once, and only the larger side pays the borrow fee. If every trader is long, the pool is effectively short, and these charges are what pushes the book back toward balance. They do not guarantee balance.
As of 30 October 2026, DefiLlama’s “GMX EVM” entry showed about 202.7 million dollars of value locked, 54.2 million dollars of open interest and 1.2 million dollars of fees over 30 days. It also lists a separate GMX Solana entry at about 188.2 million dollars of open interest. GMX’s own pages that we opened name only Arbitrum, Avalanche and MegaETH, so that entry is left out of the table below.
Jupiter Perps and JLP
Jupiter‘s perps on Solana are the clearest statement of the model, because the docs say it outright: JLP holders are the counterparty to every trade on the SOL, ETH and wBTC markets, paying out when traders win and collecting when they lose. In return the pool earns 75% of the fees from those markets. Leverage runs from 1.1x to 250x. The base fee is 0.06% to open and to close, with a price impact fee of up to 0.44% on top, plus a borrow fee that accrues with utilization. Prices come from Edge by Chaos Labs, with Chainlink and Pyth used for verification and as fallback. Jupiter’s newer beta markets do not use the JLP pool and work on an hourly funding rate instead.
Compare the two columns that matter on 30 October 2026. DefiLlama showed about 703 million dollars of value locked in Jupiter’s perp exchange against only 41.2 million dollars of open interest, with 6.8 million dollars of fees over 30 days. On that date the pool was about 17 times the open interest DefiLlama recorded, so the size of the pool alone says little about how much of it is at work against traders.
gTrade: a small vault venue
Gains Network’s gTrade belongs to the same family, though its beginner documentation says little about the vault: you trade with USDC at a price from an oracle network, and pay an open and close fee plus an hourly holding fee. It is usually described as a synthetic venue with a vault on the other side. As of 30 October 2026, DefiLlama showed about 6.8 million dollars of value locked and 1.6 million dollars of open interest, so it is now marginal.
The virtual AMM: a curve with nothing in it
The third answer is history, and it is worth knowing because it shows how far the design has moved. Perpetual Protocol launched its first version on xDai in December 2020. It borrowed the constant product formula from swap pools, but with reserves that did not exist. A May 2021 technical write-up by the writer 0xkowloon gives the shape: a market might be set at 100 virtual ETH and 380,000 virtual USDC, so k is 38,000,000 and the price 3,800. Real collateral sat in a separate USDC vault.
Run the numbers with those example reserves. A trader who goes long with 38,000 dollars of notional adds to the virtual USDC side, which becomes 418,000. Virtual ETH must fall to 38,000,000 / 418,000, or 90.91, so the trader receives 9.09 virtual ETH at an average of about 4,180 dollars, and the curve’s price is now about 4,598. Nothing was bought anywhere, and a position of 10% of the virtual reserves moved the price by 21%. The only depth is the size chosen for k.
There were no liquidity providers in v1. One trader’s gain was another’s loss, so the vault could pay as long as the system netted out, and a share of fees went to an insurance fund that covered bad debt. If it ran short, the design allowed minting the PERP token. Trading was capped at 10x. Version 2, called Curie, arrived on Optimism in November 2021 and built its vAMM on Uniswap v3 concentrated liquidity, which brought in real liquidity providers.
Hyperliquid, dYdX, GMX and Jupiter, the venues whose designs we could check from their own documents, all use books or pools. We found no official announcement that Perpetual Protocol has shut down, so we do not say so; DefiLlama showed about 432,000 dollars of value locked on 30 October 2026, and the project is not covered further here.
New entrants and the volume problem
The fourth answer is the least comfortable: a venue where the matching happens somewhere you cannot inspect. DefiLlama lists Aster Perps’ chain simply as “Off Chain”. On 5 October 2025, The Block reported that DefiLlama’s pseudonymous founder, 0xngmi, was delisting Aster’s perp volumes after they began to track Binance’s closely, because Aster did not provide the order-level data needed to check for wash trading. That was not a finding of wrongdoing: the point was that nobody outside could test the figure. Wash trading, in which the same party sits on both sides of a trade, can inflate a leaderboard at the cost of fees alone, and that matters more when a token airdrop rewards activity.
So this lesson deliberately gives no volume ranking of perp DEXs. DefiLlama’s per-protocol perp volume endpoint is now behind a paid plan, and the venues’ self-reported numbers are the ones in dispute. Open interest and fees are harder to fake cheaply, because holding a position ties up margin and fees cost money, although neither is proof.
As of 30 October 2026, DefiLlama’s open-interest list showed Aster Perps at about 1.39 billion dollars, Variational (Arbitrum) at 876.5 million, Lighter Perps at 505.3 million and edgeX V2 at 131 million. Fees over 30 days were about 4.1 million dollars for Aster and 3.7 million for Lighter. Treat those names as a dated snapshot of a part of the market that changes fast. We have not verified their internal designs, so they are named here for scale only.
When you are the counterparty: the risks LPs carry
Depositing in a perp pool looks like the liquidity provision of the previous lesson, but the loss has a different source. A swap LP loses to price movement and arbitrage. A perp LP loses when traders win, when the protocol’s accounting breaks or when the keys are stolen.
Trader profits are your losses
This risk is in the design, not in an incident. If traders on a pool venue are right as a group, and funding has not balanced the book, the pool pays. We found no documented episode of GLP or a similar pool losing heavily to traders alone, and we are not citing one. The risk is still there: a pool’s history of gains only tells you what traders did during that period, and the pool’s main income, trader losses plus fees, can turn negative in a strong one-way market.
JELLY and the HLP vault, March 2025
On 26 March 2025, a trader held a large short on JELLYJELLY, a thinly traded Solana memecoin (about 6 million dollars at 20x, by Decrypt’s account), and the token’s spot price jumped. According to press accounts, which call it manipulation, the short was pushed into Hyperliquid’s HLP liquidator, leaving the vault with an unrealized loss. Figures vary: about 10 million dollars in The Defiant, 11 million in Decrypt, about 12 million in The Block, up to 13.5 million in CoinDesk.
Validators then voted to delist the market, citing suspicious activity, and positions were settled at 0.0095 dollars, the level at which one of the trader’s accounts had opened the short, while Decrypt reported a spot peak near 0.043 dollars. Hyperliquid said users other than flagged addresses would be made whole by the Hyper Foundation. Blockworks counted more than half of the stake with five Foundation validators and called it a validator put. Hyperliquid’s records, as reported at the time, showed HLP up about 700,000 USDC over the 24 hours. The vault avoided the loss because validators changed the terms of the market. That protected depositors, and it is a warning for anyone who assumes market rules cannot be rewritten mid-trade.
GMX v1, July 2025: 42 million dollars
On 9 July 2025 an attacker drained about 42 million dollars from GLP on Arbitrum. GMX’s post-mortem, as reported by The Block, blamed a re-entrancy bug in the OrderBook contract that let the attacker manipulate the average short price of BTC, inflate the GLP price and redeem at a profit. GMX offered a bounty, most of the funds came back and reports put the attacker’s bounty at 5 million dollars. In August 2025 GMX announced a distribution of about 44 million dollars, including 2 million from its treasury, paid in GLV tokens to affected users. The bug was in the accounting. The depositors who paid for it were hit by a flaw in the code, not by anyone’s trade.
Drift, 1 April 2026: admin powers, not a curve
The largest loss in this lesson had nothing to do with market design. Drift, on Solana, was drained on 1 April 2026. Its recovery plan puts user losses at 295 million dollars and attributes the attack to “a DPRK-affiliated threat actor”, as confirmed, it says, by the forensic firm Mandiant. Early estimates ran lower. Decrypt reported a first transfer of about 41 million JLP tokens, worth about 155 million dollars, out of the Drift Vault, and quoted PeckShield’s founder saying admin keys had been leaked or compromised. Later accounts were more precise. CoinDesk, drawing on Drift’s own timeline, described pre-signed transactions kept valid through Solana’s durable nonces, approvals from two of the five Security Council members, which met the threshold, and an admin transfer that handed over control of the protocol. A Jupiter pool token sat in a Drift vault, so one protocol’s stored value became a loss at another.
The plan offered recovery tokens representing 1 dollar of verified loss each, and a commitment from Tether of up to 127.5 million dollars for the relaunch and user recovery. On 1 July 2026 the exchange changed its name to Velocity DEX and announced a private beta for selected traders. On 30 October 2026 DefiLlama showed about 640,000 dollars of value locked under the old Drift Trade entry and about 110,000 dollars of open interest under Velocity. Sound margin rules did not help here: once the admin powers changed hands, the vaults were open.
Auto-deleveraging: 10 October 2025
Return to the opening scene. Chitra’s paper calls ADL a “last-resort loss socialization mechanism”, and that phrase is exact. Hyperliquid’s documentation gives the rule: when an account or isolated position goes negative, users on the opposite side are ranked by unrealized profit and leverage, using the index (mark price / entry price) × (notional / account value). Their positions are closed at the previous mark price against the underwater account, so the platform carries no bad debt. Anyone with no open position bears none of the loss.
An example with invented numbers shows how the index sorts traders. Trader A is long from 100, the mark is 150, with 150,000 dollars of notional on 50,000 dollars of account value: 1.5 × 3 = 4.5. Trader B is long from 120 with the same notional on 100,000 dollars: 1.25 × 1.5 = 1.875. A has the higher index. The documentation gives the formula but does not say in words which end of the ranking goes first; Chitra’s paper treats queue-based ADL as hitting the top-ranked winner, and we read the index the same way, so A would be cut before B. On that reading, high leverage on a winning trade moves you toward the front of the queue, which feels backwards to anyone who thought a profitable position was safe. ADL also breaks hedges: if your short here is one leg of a delta-neutral trade, it can vanish while the long on another venue stays open.
Chitra’s reconstruction, using public data, covers roughly 160 assets and puts the aggregate deficit of losing accounts at about 100.1 million dollars. He estimates that Hyperliquid’s algorithm overshot what was needed, costing profitable traders about 45.0 to 51.7 million dollars more than necessary. Some headlines reported a far larger excess loss. That misreads the paper: its figure of 653.6 million dollars is a stylized queue diagnostic in equity dollars, described in the abstract as positions closed, not profit lost. The paper also suggests Binance overused ADL far more than Hyperliquid did. The estimate rests on a model and partial data, so treat it as an estimate. What ADL is built to do is simpler: keep the platform free of bad debt, with the winners paying.
The dated record, side by side:
| Date | Event | Figure | Who bore it |
|---|---|---|---|
| 26 March 2025 | JELLY on Hyperliquid | Unrealized HLP loss of about 10 to 13.5 million dollars, by source | HLP at first; market delisted and settled at 0.0095 dollars |
| 9 July 2025 | GMX v1 GLP exploit | About 42 million dollars | GLP holders, compensated with about 44 million dollars |
| 5 October 2025 | DefiLlama delists Aster perp volumes | No funds involved | Not applicable |
| 10 October 2025 | Hyperliquid ADL | About 2.1 billion dollars closed, per Chitra; 45.0 to 51.7 million dollars estimated overshoot | Profitable traders |
| 1 April 2026 | Drift exploit | 295 million dollars, per Drift | Users with funds on Drift; recovery tokens planned |
Six perp venues compared, as of 30 October 2026
The table uses one source for each measure so the rows can be compared: open interest and 30-day fees from DefiLlama, retrieved on 30 October 2026. Hyperliquid’s own API gives a higher open-interest figure, as explained above. Fees and leverage come from each venue’s documentation or API and may change; “not verified” means we could not confirm the item from a primary page.
| Venue | Model | Chain | Counterparty | Base taker fee | Max leverage | Open interest | Fees, 30 days |
|---|---|---|---|---|---|---|---|
| Hyperliquid | Order book, on-chain | Hyperliquid L1 | Other traders; HLP on failed liquidations | 0.045% (lowest volume tier) | 40x on BTC, varies by market | 7.71bn | 80.9M |
| dYdX v4 | Order book in validators’ memory, matches on-chain | dYdX chain | Other traders | Not verified | 50x on BTC and ETH | 40.8M | 0.24M |
| GMX v2 | Pool at oracle price | Arbitrum, Avalanche, MegaETH | GM and GLV liquidity providers | 0.04% or 0.06% | 100x | 54.2M | 1.2M |
| Jupiter Perps | Pool at oracle price | Solana | JLP holders | 0.06% plus price impact fee | 250x | 41.2M | 6.8M |
| Gains gTrade | Vault, oracle price | Arbitrum, Base and others | Vault (not confirmed in its docs) | Not verified | Not verified | 1.6M | 0.24M |
| Aster Perps | Listed as “Off Chain” by DefiLlama | Not verified | Not verified | 0.04% (general crypto pairs) | Not verified | 1.39bn | 4.1M |
Open interest and fees are in US dollars. Fees here are DefiLlama’s fee measure for the whole venue, not your trading cost. What the table cannot show is the size of the vault behind a venue and who controls its contracts.
What the law says where you live
These products exist largely outside the rulebooks of the countries where many of their users live. Nothing here is legal advice, and rules aimed at firms do not automatically say what an individual using a self-custodial interface may do.
US rules moved this year. Perpetuals were until recently a largely offshore product, then on 29 May 2026 the CFTC approved KalshiEX to list BTCPERP, a Bitcoin perpetual, and CoinDesk reported a no-action letter the same day allowing Coinbase’s US affiliate to offer certain perpetual products routed through Coinbase Bermuda. A Proskauer client alert says the agency also issued a policy statement under which other perpetuals will be assessed case by case. On 5 October 2026 the CFTC issued a no-action letter letting designated contract markets turn perpetual-style broad-based security index futures into true perpetuals, with relief running to 20 October. None of the documents we read addresses DeFi venues. We make no claim here about which US users any given protocol blocks; check its terms. As for tax, the generic IRS guidance on digital assets we read does not spell out perpetuals, so treat funding payments and vault income as questions for a tax professional. Our guide to crypto taxes in the United States covers the basics.
British retail readers face an older rule. The FCA banned the sale, marketing and distribution to retail consumers of derivatives and exchange-traded notes that reference unregulated transferable cryptoassets. It announced the rule on 6 October 2020 and it took effect on 6 January 2021, applying to firms acting in or from the UK. Whether a self-custodial protocol counts as such a firm is unsettled.
EU law splits the subject in two. MiCA leaves out crypto-assets that qualify as financial instruments, so derivatives stay under MiFID II. On 24 February 2026 ESMA published a statement saying derivatives marketed as perpetual futures are likely to fall under the national restrictions on CFDs where they meet the CFD definition, and that the commercial name is irrelevant. Those measures include leverage limits, a mandatory risk warning, margin close-out, negative balance protection and a ban on benefits. A funding rate, a trading venue or an insurance fund does not change the analysis. The retail leverage cap ESMA agreed for crypto CFDs in March 2018 was 2:1, and its statement says all national regulators later adopted permanent measures, mostly mirroring the EU decision. France’s AMF, which had said on 22 February 2018 that offering cash-settled crypto derivatives needs authorization and falls under its ban on electronic advertising, took over the CFD restriction from 1 August 2019 with the same scope. Italy’s Consob did so with resolution 20976 of 20 June 2019, which requires a 50% initial margin on crypto CFDs, or 2:1 by our arithmetic. The statement is addressed to firms and regulators. Whether a given DEX falls within that framework is unsettled, and a user who sees 40x or 100x on one should not assume the protections listed above apply.
Questions to ask before you touch any perp DEX
Five questions cover most of what went wrong above, and all can be answered from public documents.
- Who is the counterparty in normal times, and who takes the position when a liquidation fails?
- Where does the mark price come from, how many sources feed it, and can a thin market move it?
- Who can upgrade the contracts, move funds or delist a market, and how many keys or validators does that take?
- What happens to a winner when a vault or insurance fund is empty? Read the ADL rule, including where you stand in the queue.
- What does your own regulator say about this product, and what does the venue’s terms page say about your country?
If you hold a vault token, add a sixth: what else is that token deposited in? Drift’s case shows that the answer can matter as much as the vault itself. The hands-on steps for one of these venues, including testnet use, come in a later lesson on Hyperliquid.
