Borrowing on-chain looks simple from the outside: deposit collateral, take a loan, pay interest. The interesting part is what keeps that loan solvent when nobody is watching, and almost every risk in DeFi lending comes from the gap between how collateral is priced and how quickly it can actually be sold.
Utilisation is the entire rate model
Interest does not come from a central bank. It comes from scarcity inside a single pool. Lenders deposit assets, borrowers take them, and the ratio of borrowed to supplied is utilisation. The interest rate is a function of that number, and the curve is deliberately kinked.
Below a target — commonly around 80% — rates rise slowly, so ordinary demand is not punished. Above it the curve goes nearly vertical, because the pool has to attract deposits before it runs out of liquidity to lend and to withdraw. That kink is the whole mechanism: it is what makes a withdrawal possible at three in the morning with no counterparty to find.
The implication for a borrower is that your cost is not a rate you agreed to. It is a rate that moves continuously with everyone else's behaviour. A pool going from 70% to 95% utilisation can multiply your borrow rate in minutes.
Collateral is priced, not appraised
Every market has a loan-to-value limit and a liquidation threshold, and the gap between them is the liquidation zone. A position at 80% LTV against an 82.5% threshold has a 2.5-point cushion, which on a volatile asset is not much.
That is why the oracle is the most important contract in the system. The protocol does not decide what your collateral is worth; a price feed does. Everything downstream — whether you get liquidated, whether the liquidator gets paid, whether the protocol eats a loss — follows from that number. A feed on a 24-hour heartbeat in a market that moves 15% in an hour is not a detail. It is the risk.
Worth knowing how those feeds defend themselves, because the defence defines the failure mode. TWAP oracles average the price over a window, which defeats single-block manipulation but guarantees lag in a sustained move — exactly when lag hurts the most. Redundant feeds with deviation triggers update when sources disagree beyond a band, which is faster but adds a new dependency: the fallback source itself. Neither is strictly better; read which one the market you are entering actually uses, and what its worst known delay is.
The stress test every borrower should run
The liquidation threshold is quoted at today's price, but your position lives through this week's price. The arithmetic takes two minutes and should precede every deposit, not follow the first near-miss.
Worked example: you supply $10,000 of collateral with an 82.5% threshold and borrow at 70% LTV — $7,000 of debt. The cushion between your LTV and the threshold is 12.5 points, so the position liquidates when collateral falls roughly 15% (12.5 ÷ 82.5). Now the same position at 78% LTV: the cushion is 4.5 points and the wipe-out price is a 5.5% drop — inside a single ordinary day for most crypto assets. Same protocol, same asset, radically different survivability, decided entirely by the LTV you chose.
Two rules fall out of the arithmetic. First, size the buffer to the asset's volatility, not to the protocol's maximum: assets that routinely move 10% in a day deserve cushions that survive two of those days. Second, decide your deleveraging plan before you borrow — which stablecoin you would swap into, and at what cushion level you act — because the moment you need the plan is the moment you cannot think clearly enough to write one. A position with a written plan at 74% and a panic at 81% are the same position with different endings.
Liquidations, and why they cascade
When a position crosses its threshold, anyone can repay part of the debt and take the collateral at a discount — the liquidation bonus. It is permissionless and competitive, which is why bots with private mempools usually win the race.
Cascade risk is structural rather than accidental. Large liquidations push the collateral's price down, which pushes more positions below their thresholds, which triggers more liquidations. On a thin market that becomes a loop. The protocol's backstop is an insurance fund sized for normal conditions, not for a correlated unwind. And when a position becomes unprofitable to liquidate — because the debt exceeds the collateral's value even after the bonus — the protocol is left with bad debt, which is a loss to depositors rather than to the borrower.
Two design changes mitigate this. Isolated markets let a protocol list a risky asset without exposing the rest of the book to it. Efficiency modes let correlated assets borrow against each other at higher LTV, on the assumption they will not diverge — which holds until it does not.
Leverage loops: compounding every risk at once
The most popular borrow-side strategy is also the clearest illustration of why the risks above cannot be evaluated one at a time. A loop: supply $10,000 of ETH, borrow 70% of it in stablecoins, swap the stablecoins back into ETH, supply that too, and repeat. After four loops the same $10,000 controls roughly $34,000 of exposure — $10,000 yours, $24,000 borrowed.
Each loop multiplies every number in the stress test. The effective liquidation price moves much closer to spot, because the position's LTV is now a function of total debt against total collateral, both inflated. Rate risk compounds too: a pool crossing its kink now multiplies the cost of $24,000 of debt, not $7,000. And oracle lag, which at one leverage layer means a near-miss, at four layers means the near-miss arrives during the same hour you are trying to unwind.
Loops also concentrate a subtler risk: exit capacity. Unwinding four loops means selling $24,000 of ETH into the same pools whose utilisation is already elevated — the unwinding itself pushes the LTV of every other looper across their thresholds. In a crowded trade, the loop structure converts ordinary volatility into a synchronized exit. That is not a tail risk invented for this paragraph; it is the mechanical description of most lending-market crashes.
None of this makes looping irrational — it makes it a lever, and levers amplify whatever is at the other end. The discipline that separates the two outcomes is boring: run the stress test at the looped LTV, not the original one, and cap total leverage where a two-day move does not force an unwind at the worst price.
The other side of the ledger: supplying
Every borrower decision has a mirror image for lenders, and the mirror is where depositors get hurt. Yield in a lending pool is not a gift; it is the price of the risks borrowers are taking with your assets. High utilisation means high rates — and also less liquidity for your withdrawal and more stressed borrowers on the other side.
The supply-side risk that documentation buries is bad-debt absorption. When liquidations fail — the cascade scenario above — the loss lands on depositors pro rata, silently. That is why the utilisation kink matters to a lender as much as to a borrower: a pool that lives above its kink is a pool paying you for liquidity it does not really have. The questions a depositor should ask before supplying: what share of the book is the riskiest collateral, how large is the insurance fund relative to that collateral, and has this market ever taken a bad-debt event — and if so, who absorbed it. A market that answers those three plainly is doing more for your deposit than an extra point of yield.
Reserve factors and safety modules
The insurance fund mentioned above is not an abstraction — it has a funding mechanism, and the mechanism is where protocol governance reveals its priorities.
Every market charges a reserve factor: a slice of the interest paid by borrowers that accrues to the protocol instead of to lenders. A 10% reserve factor on a pool earning 4% takes 0.4% a year off lender yield and parks it as the first-loss cushion. Reading reserve factors by asset tells you how the protocol itself ranks risk — risky collateral carries a high factor (more sequestered revenue to cover its failures), blue-chip collateral a low one. When governance debates lowering a reserve factor, the substantive question is always who pays for the next shortfall.
Larger protocols layer a safety module on top: a staking pool, usually denominated in the protocol's own token, whose deposits can be slashed — seized and sold — to cover a deficit. Suppliers stake there for rewards; the reward is compensation for accepting the slash condition. That design has a known circularity worth seeing plainly: the backstop is denominated in the protocol's own token, so a failure severe enough to trigger the backstop is usually severe enough to crash the token covering it. The backstop is deepest exactly when the protocol is healthy and thinnest exactly when it is needed. It is still real protection for mid-sized shortfalls; it is not protection against the correlated unwind that defines the tail.
For a depositor, the practical read is coverage math, not vibes: reserve-factor revenue plus safety-module size versus the largest collateral bucket in the book. A protocol whose backstop could absorb the failure of its riskiest single market is sturdier than its headline APY competitors, and usually offers less APY — that difference is the insurance premium.
The market landscape
Aave and Compound are the general-purpose pools, with the deepest liquidity and widest asset coverage. Morpho separates the matching engine from the risk parameters, letting curated vaults set their own terms over shared liquidity. Euler explored permissionless listing and long/short markets before its rebuild. Spark and similar venues specialise in the stablecoin-against-Treasuries trade.
The design axis that actually matters is who chooses the risk parameters. In a monolithic pool, one governance vote sets a collateral factor that affects everyone. In a curated or isolated architecture, many small decisions can fail independently. Neither is safer in the abstract; they fail differently.
When the collateral is an NFT
NFT lending is where the assumption "collateral has a deep market" breaks completely, and every design in this corner is a patch on that break.
Peer-to-pool quotes a principal from a collection's floor price — NFTfi offers, BendDAO's pooled ETH. Peer-to-peer lets borrower and lender negotiate principal, duration, and rate, which is the only way to handle a rare item a pool cannot price. Both share one flaw: a floor price describes the cheapest unit in a collection, not the token actually posted, so a rare trait can be worth a multiple of the loan or a fraction of it.
Fractionalisation (Fractional.art, now Tessera) splits an NFT into fungible shares, giving the position a tradeable market and therefore a usable oracle. The cost is governance rather than price: someone must decide when the group sells, and the buyout rules are where these structures actually break.
Liquidation cannot be a market sell. There is no order book to absorb an NFT, so protocols either run a multi-day Dutch auction or accept a floor bid. A position can go underwater faster than it can be closed — the structural difference from fungible lending, where the venue is deep enough to liquidate inside a block.
The practical read
Treat the loan-to-value limit as a suggestion and keep your own buffer well below the liquidation threshold, because the threshold moves with the oracle and the oracle does not warn you. Read the oracle's update mechanism before you deposit, not after. And if you are the lender, understand what your yield is paying for. The pool is safe until a correlated move makes liquidation unprofitable, at which point depositors absorb the difference.
Condensed into a pre-deposit checklist:
- Utilisation and kink: where does the pool sit today, and what rate did it pay last time it crowded?
- Oracle: which feed, what update mechanism, what is the worst known lag?
- Collateral mix: what share of the book is the volatile or bridge-sourced asset?
- Liquidation history: has this market cascaded before, and did the insurance fund cover it?
- Your own buffer: stress-test arithmetic run, deleveraging plan written, both dated.
Five minutes of reading answers the first four; the fifth is the one nobody does until after.
References
Next, read Derivatives Trading 101. More in DeFi Markets.