←Back to all guides

Who Keeps the Liquidation Bonus

Solver Guides | By Daniel Peter

A week of Arbitrum liquidations, measured from our own node: the winners bid away 97% of it, and in a crash they bid away more.

We run a dedicated Arbitrum node for Kaisersolver. In November we asked whether that node could earn a second line of revenue from liquidations, the oldest dependable income in DeFi. So we built a watcher, ran it for a week that included two sharp market moves, and measured who actually keeps the money.

The answer is not us, and not anyone else bidding for it either. On Aave, which carries most of Arbitrum's liquidation volume, the bonus now goes through an auction. Winners bid a median of 97% of it away, and keep under two dollars per auction. In a crash they bid away more than the bonus. We are publishing the measurement because we could not find it anywhere else, and because a clear "no" is worth more to someone deciding whether to build than another explainer.

What a liquidation pays, and where it used to go

A lending protocol lets you borrow against collateral up to a limit. When the collateral falls far enough that a loan crosses the limit, anyone may repay part of the debt and receive the borrower's collateral at a discount. Each collateral asset has its own discount, set by Aave governance, and the protocol keeps a share of it. In our data, liquidators of WETH collateral on Aave's Arbitrum market netted 4.5% of the debt they repaid, and of WBTC collateral about 6.3%. That discount is the liquidation bonus. It is the protocol paying for prompt cleanup.

For years the bonus went to whoever got there first. That meant a race for transaction ordering: priority fees, private order flow, and latency. The value stayed with the fastest bots.

On Arbitrum, Aave now routes most liquidations through Chainlink's Smart Value Recapture (SVR). When a price update is due, it is offered in a short sealed-bid auction run on FastLane's Atlas. The winning searcher's liquidation is placed in the same transaction, immediately after the update, and the winning bid is paid to the protocol side rather than kept by the liquidator. Chainlink says the proceeds are split between the integrating protocol and Chainlink at a standard rate. For Ethereum, Aave's governance forum gives that split as 65% to the Aave ecosystem, 31.5% to Chainlink and 3.5% to block builders. We did not find the Arbitrum split published. So the race for ordering becomes an auction for the right to liquidate, and the question becomes how much of the bonus the auction takes.

How we measured it

The watcher runs on the same node as Kaisersolver, reads only, and holds no keys. Over the week it tracked about 11,000 borrowers across eleven lending markets on Arbitrum: Aave, Compound, Morpho, Silo, Dolomite and Venus. For every liquidation it recorded:

  • The winning bid, read from the auction's own on-chain result events, and priced in ETH at that moment. Our node is pruned and cannot read old state, so historical prices come from the Chainlink feeds' own round history, which stays readable.
  • The bonus the liquidation actually paid: collateral seized minus debt repaid, from the protocol's own event, at prices of the same moment.
  • A full breakdown of the winner's transaction: every token movement in and out of the winning solver, who it traded with, what price it sold the seized collateral at, and what it kept.

Between 8 and 15 November it recorded 428 liquidations. 264 of them, about six in ten, arrived inside one of 136 auctions. Every auction and liquidation in this post is in a public dataset, with a script to check any row against the chain: github.com/fiftysixtech/arbitrum-liquidations-2026-11.

On an ordinary day

On ordinary days the auction is close to perfectly efficient. On 10 November, a calm day, every one of the 13 auctions we could fully price bid between 95% and 102% of the bonus. Across the whole week, the 121 auctions we could fully price had a median winning bid of 97% of the bonus, and most fell between 92% and 101%. The median winner kept $1.64 before gas and before the cost of selling the collateral it had just received. Added together, the gap between bonus and bid across all 121 came to $3,365, and more than half of it was a single auction, the largest of the week, where most of it went on selling the collateral.

Thirteen solvers won at least one auction, and one won 64% of them. In the week, about $73,000 in winning bids went to the protocol side of the auction rather than to liquidators.

Competition depth is the one thing the chain does not show. Losing bids usually never land on-chain, so we can see what the winner paid but rarely what the runner-up offered.

The crash: bidding more than the bonus

On the evening of 8 November, ETH fell from about $2,190 to $2,110 in fifteen minutes. Several auctions in that window bid more than the bonus, up to 128% of it. That looks irrational until you break the transactions down.

The largest was a single liquidation of $99,570 of debt:

Collateral seized49.33 WETH, worth $104,022 at $2,108.59, the price in the Chainlink update Aave used, inside the same transaction
Bonus at that price$4,451.99
WETH the winner then sold on Uniswap46.60, at an effective $2,136.58, 1.33% above it
Total value to the winner$5,756.33
Winning bid$5,711.65
Kept, before gas$44.68

In a fast fall, the price the oracle has just written and the price in on-chain pools are not the same for a few seconds. On 8 November the pools had not yet caught down to the new oracle price. The winner repaid the debt, took the collateral and sold it into those pools in the same transaction, at better than the oracle price. Then it bid away almost exactly that extra too.

Transaction

The bookkeeping is exact. Each winner converts its entire surplus to ETH and pays its bid out of it. In every crash auction we broke down, the ETH it converted equals the bonus plus the gain on the sale, to the cent. So what the winner keeps is that ETH minus the bid, and in the crash it came to between about −$20 and +$45.

“Every mistake we made pointed toward money that was not there.”

The sell-off: the same efficiency, in the other direction

From 12 to 15 November ETH fell from about $2,010 to about $1,710, more slowly but further. There were 106 auctions, and more than $1.2 million of debt was repaid, $327,000 of it in a single hour. This time the pools sat below the oracle price, and winning bids fell below the bonus by almost exactly the loss on the sale.

AuctionBonusCollateral sold vs Aave's priceWinning bidBid / bonusKept, before gas
0x1c76…$10,205−0.51%$8,21681%$125, plus $686 paid to a third address*
0x9e82…$2,310−0.72%$2,02388%−$85
0x79dc…$2,018−0.51%$1,78689%+$69
0x7e42…$2,915−0.23%$2,72694%+$39
0x2a69…$2,016−0.15%$1,95497%−$7

"Aave's price" is the price in the Chainlink update inside each transaction, which is the one Aave used for the liquidation.

* The largest auction is the one place the winner plainly kept something. Besides the bid, its contract kept $125 of WETH and paid $686 of USDC to an address that passed nothing on. That is either a fee or the winner's own wallet, so the winner kept between $125 and $811, under 8% of the bonus. The arithmetic closes exactly: the $10,205 bonus, less the $1,177 lost selling the collateral, equals the $8,216 bid plus the $811.

Read together, the two episodes say one thing. The winning bid equals the bonus, plus or minus whatever the collateral fetches against the oracle at that moment. The winner keeps roughly nothing, and in the largest case a few percent. Whichever way the gap runs, the auction prices it in.

When the highest bid fails

One exception is worth describing, because it shows where value can survive.

On 14 November at 22:09 UTC, the highest bid in an auction, 0.91 ETH, landed on-chain but its execution failed. Atlas then ran the next solver in line, which paid only its own bid of 0.47 ETH. That solver kept $723 over its bid, even after selling the WBTC it received 0.52% below Aave's price.

Transaction

A higher bid failed in only one other auction that week, and there it was worth a few dollars. But it is the only case in the week where winning came down to whose transaction worked rather than who bid most, and failed bids are recorded on-chain, so it can be counted. We will keep counting it.

Three mistakes we caught in our own numbers

We got three things wrong on the way. All of them made the opportunity look bigger than it was.

The first was timing. We were valuing older bids at the current ETH price instead of the price when the bid was made. That inflated some bids to well over 100% of the bonus, and for a day it looked like winners were overpaying. Pricing every bid from the feed's history at its own moment removed most of it. The overpayments that survived are the real ones described above.

The second we very nearly took as a finding. In transactions that liquidated several positions at once, our code gave every liquidation the first one's dollar figures. A bundle of eight liquidations had its bonus counted eight times. For one morning, the sell-off looked like auctions clearing at 14–50% of the bonus, with winners keeping about $85,000. It came apart when we checked the recorded totals against the token flows in the transactions themselves, which did not agree. Once fixed, those auctions sat in the 80s and 90s, and the $85,000 became roughly nothing.

The third surfaced when we checked this post's own transactions against our node. We had valued collateral at the public Chainlink feed. Aave used the Chainlink update bundled inside each auction transaction, and in the sell-off the public feed ran up to 1.1% above it. That made the bonus look bigger and the discount on selling the collateral look wider. The same check showed a payment to a third address that we had counted as a trading loss. Neither changes what winners kept, which we measure in ETH, but both changed the table above.

We mention all three because the direction of the errors is the lesson. Every mistake we made pointed toward money that was not there. If you are measuring an opportunity you would like to be real, check your totals against the chain, transaction by transaction, before you believe a number.

What is left outside the auction

Not every market uses it. The week turned up two small places where a bonus was still up for grabs:

  • Morpho Blue: about $3,600 of bonus over three days, taken in same-block races among a handful of liquidators.
  • Dolomite: a few positions stayed liquidatable for seconds, and one for more than two hours. The ones our deliberately simple router could price came to about $2,100 over three days, against two or three competitors. Those liquidators use their own capital, and some of the collateral cannot be sold instantly.

Both are real. Both are episodic: those figures come from a volatile week, and calm weeks are close to zero. And both would require us to run a signing key on a hot transaction path, and in Dolomite's case to hold inventory. Fiftysix operates without either by design, for reasons we have written about separately. A few thousand dollars in a good month does not change that.

If you are thinking about liquidations on Arbitrum

Three things, offered as observation rather than advice.

The protocol is getting what it designed for. On Aave, the auction recovers almost the entire bonus, and in a crash some of the value that used to leak to arbitrageurs. For a protocol, that is the system working.

Speed is no longer the edge, and price mostly is not either. The auction prices the bonus and the DEX–oracle gap to within about one percent. What is left goes to whoever sells the collateral best, and occasionally to whoever is still standing when a higher bidder fails.

Measure before you build, and measure the transaction, not the event. A protocol's event tells you what the bonus was. Only the winner's token flows tell you what the winner kept. The difference between the two is the whole question.

We have pointed the node back at the work it was bought for. The watcher keeps running in the background. If the picture changes, we will update this post rather than write a new one.

Frequently asked questions

What is Chainlink SVR?

Smart Value Recapture. When a Chainlink price update would make loans liquidatable, the right to liquidate straight after that update is sold in a short sealed-bid auction run on Atlas. The winning bid is paid to the protocol side and Chainlink rather than kept by the liquidator.

How much of the Aave liquidation bonus does the winner keep on Arbitrum?

Almost none. Across 121 fully priced auctions between 8 and 15 November 2026, winners bid a median of 97% of the bonus away and kept a median of $1.64, before gas and before the cost of selling the collateral.

Is running a liquidation bot on Arbitrum profitable?

In our week of data, not on Aave: the auction prices the bonus, and the gap between pool and oracle prices, to within about one percent. Small amounts remained outside the auction, on Morpho Blue and Dolomite, mostly during sharp market moves.

Why do winners sometimes bid more than the bonus?

In a fast fall, on-chain pools can trail the new oracle price for a few seconds, so the seized collateral sells for more than the price Aave used. Winners bid that extra away too, up to 128% of the bonus on 8 November.

Where is the data?

At github.com/fiftysixtech/arbitrum-liquidations-2026-11: every auction and liquidation from the week, under CC BY 4.0, with a script that checks any row against an Arbitrum node.

Sources for the auction design: Chainlink, Smart Value Recapture feeds; Aave governance, SVR v1 phase 1 activation; Aave governance, SVR expansion to Base and Arbitrum. Measurements are by Fiftysix from its own Arbitrum node, 8–15 November 2026. Data: github.com/fiftysixtech/arbitrum-liquidations-2026-11, CC BY 4.0. The watcher is observation-only and holds no keys. Profit figures from our router are floors, and losing bids are mostly not visible on-chain. Figures are as at the dates stated. Nothing here is legal, tax or investment advice.

Share on:

Next Guide: How to Become a CoW Protocol Solver →