Pendle + sUSDai — a fixed rate on GPU-backed credit
Data as of 24 Sep 2026.
Verdict: usable only with a tight limit. Pendle is the sturdier half here: the market core is immutable, it has no exploits of its own, and the PT rate is enforced by code. The weak half is sUSDai. It isn't a stablecoin — it's a share in a credit pool, ~58% of which is loans to thirteen Nvidia GPU lessors. Those loans are marked by the lender itself, exits go through a 30-day queue, and not a single loan has run even half its amortisation schedule. A PT holder eats any loss in that book in full: the rate is only fixed in sUSDai units.
What it is
Three wrappers stacked. USDai is a dollar token backed by PayPal/Paxos PYUSD. sUSDai is a vault that lends those dollars against GPU servers; the share price accrues interest. PT-sUSDai on Pendle is a zero-coupon bond: bought at a discount, redeemable for sUSDai at maturity. Issuer: USD.AI Foundation (Cayman Islands); developer: Permian Labs (Delaware); public launch 20 Aug 2025 (CoinDesk, 14 Aug 2025). Home chain: Arbitrum.
Sizes as of 24 Sep 2026. Pendle: $1.28bn TVL on DefiLlama, −90% from a $13.39bn peak (19 Sep 2025). The sUSDai vault: $507.0m in assets and 322.3m shares (on-chain read on Arbitrum, totalAssets/totalSupply). DefiLlama shows $207m for USD.AI. The ~$300m gap equals loans outstanding ($291.4m, app.usd.ai/reserves, 22 Sep): the tracker only sees the liquid sleeve. Not a contradiction, but read "TVL" here with care.
Pendle TVL: points on 1 March and 1 September, plus the peak and the latest print. Source: api.llama.fi/protocol/pendle, read 24 Sep 2026. The decline is deposit outflow, not losses — there are no exploit events in the series.
Findings
1. A fixed PT rate hedges sUSDai yield changes — not losses in its loan book
This is the core point. In Pendle's contract the redemption index is max(current rate, stored index) and never goes down (PendleYieldToken.sol, line 406, pendle-core-v2-public, read 22 Sep 2026). So 1 PT at maturity delivers the same number of sUSDai regardless of what each share is worth in dollars by then. If the sUSDai rate drops because a loan is written off, the PT holder receives the same count of cheaper shares. Pendle's docs say plainly that PT redeems 1:1 into the accounting asset, not the underlying.
The asset class already has precedents. USD0++, January 2025: the issuer introduced a $0.87 redemption floor, and PT-USD0++ was worth ~$0.92 at maturity (Leviathan News, The Big Whale). Resolv USR, 22 Mar 2026: a compromised minting contract printed ~80m unbacked USR, and the Curve price fell to $0.025 (CoinDesk, 23 Mar 2026; Resolv post-mortem). The recovery plan carves out Pendle position holders as a separate class (Bitget News). Pendle itself was not exploited in either case — the loss came through the underlying.
2. sUSDai is ~42% PayPal dollars and ~58% GPU loans. The book is young, and the lender marks it
Vault composition: $507.0m of assets (on-chain, 24 Sep) minus $291.4m of loans outstanding (issuer figure, 22 Sep) leaves ~$216m of liquid USDai — 42%. The other 58% is loans. Per the issuer's "Lighthouse" report (8 Jun 2026) there were 13, the largest at $98.1m, or ~48% of the book at that point (arithmetic on two issuer figures). No September breakdown has been published. An independent data point: LlamaRisk on 3 Apr 2026 counted 27 loans totalling $18.1m, 6.4% of the vault. So the credit sleeve went from 6% to 58% in five months. No defaults yet — because 36-month amortising loans haven't run even halfway.
Collateral valuation: the protocol is oracleless by design — collateral is priced by independent appraisals, and the valuation is maintained by a service provider (docs.usd.ai). Loans sit on the books at outstanding principal until a default is declared — no mark-to-market. In tranched credit, book marks typically lag realised losses by around six months. So "no losses to date" says nothing about the quality of this book.
3. Exit means the secondary market, because redemptions run through a 30-day queue
The docs say outright that sUSDai is not a stablecoin and isn't instantly redeemable at par. Redemptions run on global 30-day epochs in a FIFO queue, and active GPU loans are not terminated or liquidated early to meet them. Since April 2026, redeeming USDai into PYUSD is open only to KYC'd market makers (docs; LlamaRisk, 3 Apr 2026).
The real exit today is secondary pools. Per a 22 Sep snapshot (GeckoTerminal), Fluid and Curve pools on Arbitrum had ~$37m of depth at the vault rate, no discount. DefiLlama's price on 24 Sep was $1.1140 vs convertToAssets 1.11441 — no gap. Selling PT through Pendle on 22 Sep SDK quotes: 100k PT at −0.19% price impact, 750k at −1.36%, no route at 1m. But $37m of secondary depth is ~11% of sUSDai float. In a stress, the positions currently posted as collateral on Fluid (~$130m, DefiLlama, 22 Sep) will be heading for the same pools.
4. Who can change the rules: Pendle is a 2-of-4 multisig; USD.AI runs three multisigs, timelock unconfirmed
Pendle: the market core (AMM curve, PT/YT/SY within a market) is immutable. The router, factories, oracles and SY adapters are governed, and newer SY adapters are mostly upgradeable proxies. They're controlled by a 2-of-4 core-contributor multisig that can set pauses and caps (OAK Research, 5 Dec 2025; single source). That lever has already been used defensively: on 3 Sep 2024, during the exploit of Penpie ($27m), a protocol built on top of Pendle, Pendle paused its contracts, says it protected ~$105m, and resumed within a day (The Block).
USD.AI: per LlamaRisk (3 Apr 2026) — three team multisigs (3/3, 2/4, 3/4), no timelock, upgradeable proxies, a separate pause role, a blacklist. The docs now mention a Timelock Controller, but it hasn't been verified on-chain. A timelock is table stakes for any DeFi protocol, and until it's confirmed this is the biggest hole. The terms of use (Cayman law) state plainly that redemptions may be delayed, restricted or not possible at all.
5. The wrapper stack is already at the ceiling
USDai → sUSDai → PT is exactly three layers. A sensible ceiling for this kind of structure is three; a fourth is a hard no. So any borrowing against PT-sUSDai, or looping it, is out of bounds — by rule, not by taste.
Independent risk views diverge. Chaos Labs (19 Oct 2025) did not recommend listing. LlamaRisk was against on 20 Oct 2025 and in favour on 3 Apr 2026, with conservative parameters. On-chain share price: 1.03288 (21 Sep 2025) → 1.11381 (21 Sep 2026) → 1.11441 (24 Sep 2026). That's +7.84% realised over a year, against the issuer's headline "Current APR 8.26%, Expected 11.71%" (usd.ai/susdai, 22 Sep).
What breaks it, and at what number
| Event | What happens to the holder | How to check |
|---|---|---|
| First declared default by a large borrower (48% of the book in one name as of June) | sUSDai rate falls for the first time; PT at maturity delivers the same count of cheaper shares. Barkr cover (80% of appraised value, reinsured by Munich Re) only applies to loans after 6 Feb 2026 (usd.ai/insights; Reinsurance News) | on-chain convertToAssets, monthly |
| Outflows the secondary pools can't absorb | sUSDai trades below the vault rate and PT follows it down. Redemption is a queue that "doesn't liquidate loans" | sUSDai/USDC on Fluid vs vault rate, same day; a gap above 1% for two days is the signal |
| Issuer code exploit or malicious upgrade (upgradeable proxies, timelock unconfirmed) | Tail up to 100%. Class precedent: Resolv, 22 Mar 2026 — one privileged key, ~80m unbacked tokens | incident feeds; on-chain timelock check |
| Pendle pause | Costs time, not money. Precedent 3 Sep 2024: one day | event-driven |
The call
On 23 Dec 2026 the sUSDai rate (convertToAssets, Arbitrum) will be in the 1.1311–1.1389 range — 1.5–2.2% above 1.11441 on 24 Sep 2026. Below 1.1311 means a write-down or loss of PYUSD promo yield. Above 1.1389 means the promised 12% has become real — and the claim that "realised is ~7.6–7.8% and the headline is inflated" was wrong. Any decline in the rate versus the previous read pulls the verdict immediately, regardless of the call.
Unknowns
Obtainable:
- Whether there's a timelock on sUSDai/USDai upgrades, and how long — readable from the proxy admin on-chain (Arbiscan returned 403 on 22 Sep). This decides whether the verdict stays "with limits" or drops to "no".
- The September loan book, especially the largest borrower's share after June — pending the issuer's next report.
- The Network Firm "Proof of Loans" attestation (15 Jul 2026) — the file wouldn't open. Until it's read, treat the backing as unattested.
- Current redemption queue length — "Withdrawal estimates" page in the docs.
Unknown to anyone:
- What used B300s will fetch in the first real foreclosure. There has been no GPU loan default in this class yet; recovery values exist only as dealer estimates.
- Whether the SPV's bankruptcy remoteness holds up in court. No tokenised credit structure has been through a court-run bankruptcy yet.
Limits, and what pulls the verdict
Asset class. This is a single-sector credit bet, not a dollar. It can't be counted as cash or as a stablecoin.
Exit depth. Size PT no larger than Pendle can sell at ≤0.5% price impact. On 22 Sep that was ~270k PT (linear interpolation between the 100k and 750k quotes). Re-quote before entry.
Concentration. A common ceiling is up to 10% of a portfolio per RWA issuer; each holder can set a lower one.
No fourth layer. No borrowing against PT or sUSDai, no loops: the stack is already at the three-layer ceiling.
Pulls the verdict immediately (to "no"):
- any decline in the sUSDai rate (
convertToAssets) versus the previous read — i.e. the first booked loss; - sUSDai trading more than 1% below the vault rate on secondary markets two days running;
- the issuer suspending or extending redemptions beyond two epochs (60 days), or introducing a redemption floor à la USD0++;
- an exploit of Pendle's core or of the sUSDai SY adapter; a Pendle pause longer than 72 hours;
- no on-chain confirmation of a timelock on sUSDai upgrades by 23 Dec 2026 — then "no" for new money.
Tightens the limit without changing the verdict: largest borrower above 50% of the book in a new issuer report; liquid sleeve of the vault below 30%; Pendle TVL below $600m (half today's level).
Next checkpoint: 23 Dec 2026 — the rate call resolves and the timelock check should be closed.
Sources: docs.usd.ai (depositor/susdai, borrower, technical overview, ToS) · usd.ai/susdai, usd.ai/insights (Lighthouse, 8 Jun 2026) · app.usd.ai/reserves (22 Sep) · LlamaRisk 20 Oct 2025 and 3 Apr 2026 · Chaos Labs 19 Oct 2025 · docs.pendle.finance · github pendle-core-v2-public · OAK Research 5 Dec 2025 · The Block (Penpie, 2024) · CoinDesk 14 Aug 2025 and 23 Mar 2026 · Resolv post-mortem 22 Mar 2026 · Bitget News (Resolv plan) · Leviathan News, The Big Whale (USD0++) · Reinsurance News · DefiLlama · GeckoTerminal.
Research for information only. Not investment advice.