Tom Dunleavy (@dunleavy89) gives a useful framework for a factor-decomposition of yields and rates in DeFi. However, in my view, his conclusion for onchain HY repo, 12.55%, is overstated.
The framework is a reasonable one to apply. The directional question (is DeFi lending priced for the risk taken?) is exactly the question institutional allocators should be asking. The answer is more interesting than either "yes, ignore the data" or "no, demand 12%." It depends entirely on which DeFi lending you're talking about, and the empirical loss data tells a clean story across distinct tiers that the original essay collapses into one.
Tom refers to a Duffie-Singleton credit spread decomposition. The model is intuitive and actually quite simple:
Where r is the risk-free short rate, h is the hazard rate (default intensity), and L is the expected fractional loss given default. The credit spread under D-S is one term.
Tom adds oracle manipulation, governance attack, composability cascade and stablecoin depeg to the mix.
R = Risk-free base (10Y UST) + Technical expected loss + Oracle manipulation risk + Governance / admin key risk + Composability cascade + Regulatory asymmetry + Stablecoin depeg tail + Liquidity premium + Risk Premium
In my view, however, four of these are not separable independent premia. They are the causes of default events that are already captured in the expected loss term.
The Kelp event makes this concrete: it is a composability cascade and a governance / admin key risk materialized. If we count Kelp's 200m in the loss data, as Tom does, the composability premium is already embedded in the 1.50% expected loss line. He cannot then add another 1.25% for "composability cascade" and "governance risk" on top.
Therefore, my view is that Tom's 12.55% is highly overstated.
Stripped of the double-count, an illustrative decomposition applied to DeFi becomes:
Yield = Rf + Expected Loss + Risk Premium + Liquidity Premium
You could argue, for institutional contexts, a regulatory access cost, defensible as a separate line for compliance overhead.
Expected Loss encompasses every loss mechanism: Exploit, oracle manipulation, governance failure, composability cascade, depeg, etc. The risk premium compensates for the non-diversifiable systematic component above expected loss. The liquidity premium compensates for inability to exit. With those four components, the fair-value question becomes tractable.
Risk Free Anchor
Tom starts with the 10Y Treasury at 4.29%. This is the wrong reference. DeFi lending is overnight repo with continuous repricing, there is no duration term to compensate. The correct Rf is SOFR, currently 3.6%. Comparing DeFi lending yields to 10Y UST imports duration risk that DeFi suppliers do not bear. Steakhouse Prime, and even curated HY vaults all reprice every block. The benchmark is overnight-equivalent.
Risk Premium
Non-diversifiable or model error components over expected loss. Tom uses 1.5% as a catch-all. For Prime markets, the systematic component above expected loss is small. Loss events are largely the result of gap risk rather than composability. This factor is already captured in the expected loss term. We'll allow 25bps for the sake of the argument.
High Yield markets arguably feature more correlated sources of potential risk - though we might not go as high as 1.5%, we can accept it again for the sake of the argument. Notably, Drift and Kelp, although correlated on the "type" of loss, were nonetheless not part of a single default clustering event that in TradFi contributes to HY spreads.
Liquidity Premium
TradFi private credit pays a positive illiquidity premium of roughly 200 basis points relative to public benchmarks. This compensates for multi-year capital commitment and the absence of meaningful secondary markets. DeFi has the opposite property. The correct adjustment should be zero or even modestly negative. The original essay charges +50 basis points. For the sake of the argument we can concede a +50bps premium in High Yield vaults, as in times of stress liquidity does lock up. Though still for significantly less time (hours or days vs months), and none (or 5bps concession) in Prime.
Tiering
Prime DeFi
Steakhouse USDC is the benchmark for noncustodial isolated-market exposure. Collateral is BTC and ETH only, with a significant haircut between LLTV and bad debt. Across all Steakhouse-curated Morpho vaults on all chains, since January 2024, the empirical loss rate is 0.000000426%.
A conservative forward-looking PD × LGD for Prime, allowing for gap risk under correlated stress where liquidators compete for block space: ~5 basis points.
HY DeFi
Steakhouse High Yield, with multiple exposures to longer tail collateral and meaningful composability cascade surfaces: Let's accept Tom's loss rate of 1.5%. Though it's worth noting Steakhouse High Yield vaults have empirically faced no bad debt either.
I Clauded some historical estimates for comparable credit asset classes in TradFi to get a sense of where this stacks:
Temporary supply-rate spikes during utilization stress are exactly what variable-rate money market mechanisms are designed to do, namely, incentivize pulling more supply into the system. In the Aave example, this rate could be argued was set too low to execute that role.
Morpho's IRM is far more aggressive with the effect of washing out borrowers who don't want to take the pain or properly compensating new lenders for the stress and illiquidity, for solvent but tapped out collateral markets.
Using R = Rf + EL + RP + LP
Prime
R = 3.6% + 0.05% + 0.25% + 0.05% = 3.95%
Steakhouse USDC is currently within 50bps of this target, sometimes less sometimes more depending on the supply/demand balance between borrowers and lenders. The historical undershoot on a 30D basis is still puzzling but may be a reflection either of subdued borrower demand for variable-rate repo vs fixed or could also capture a negative liquidity premium (these markets are objectively more liquid than term loans against the same collateral).
With SOFR at 3.6%, a 1.5% risk premium and a 50bps liquidity premium, Tom's 12.55% actually solves back to an implied expected loss of 7%. What Tom is pricing with 12.5% is a different type of beast, a savage DeFi lending market with an expected 7% loss rate.
Conclusion
None of the above is intended to justify that DeFi cannot or should not do better on any of the risk factors he identified. But, similarly, I believe it's important to be precise when we refer to things and not throw out the baby with the bathwater.
The key advantages of DeFi (transparency, composability, settlement finality) have a dark side if not applied correctly, because losses materialize immediately, spread quickly and are irreversible and large.
DeFi rate pricing has significantly matured over the past few years. However, I agree with Tom that to reach the end-state where these advantages can properly flourish, we still have a long way to go to understand how to operate this beast.
Understandably a lot of people are spooked about DeFi at the moment. The threats are only getting larger. We have to turn the ratchet up constantly.
But the rate for DeFi lending is not 12.55%.
Thoughts/comments/errors welcome
