sUSDai — Risk Report
Moderate-to-elevated risk · 4.5/10
| Yield | Exit methods | Effective time-to-cash | Age | Chains |
|---|---|---|---|---|
| NAV growth from GPU-loan interest — about 7.8% annualised over the last 76 days; live figure on the dashboard | Epoch FIFO redemption queue (issuer-documented 30-day; the async structure is on-chain verified, the interval is not), or two Fluid pools plus Curve on the secondary | Up to 30 days on the queue and longer if it rolls; minutes on secondary at retail size | About 10 months | Arbitrum (canonical), plus bridged on Ethereum, Plasma and Base |
Summary
sUSDai is the yield-bearing, credit-risk-bearing leg of USD.AI. You stake USDai into an ERC-7540 async vault that lends to neoclouds and AI-infrastructure operators against installed GPU hardware, run on the MetaStreet lending engine. All of the yield and all of the collateral exposure live here — not in USDai, which is the clean PYUSD reserve leg covered in its own report. In plain terms this is a tokenized private-credit fund collateralized by AI compute hardware, not a stablecoin, and it should be sized like one.
The 4.5/10 reflects a genuinely-structured credit product carrying genuinely exotic collateral. On the supportive side: real overcollateralization at a stated 70–80% LTV, a debt-service reserve funded at each loan’s close, a warrantied collateral-value schedule with an institutional reinsurance layer above it, a Cantina audit with no critical or high findings, and an admin chain whose 48-hour delay is binding rather than decorative. Working against it: the collateral depreciates on a technology cycle and cannot be marked from chain, two borrowers represent close to the entire loan book, delinquency and default rates are not disclosed anywhere public, and the primary exit is a 30-day queue that the protocol will let roll rather than force-liquidate loans to meet.
The vault grew materially over the summer — total assets of $354.99M as of 2026-08-12, up about 18% in three weeks — and secondary depth grew faster, which is what moved the liquidity axis this cycle. What did not change is the binding constraint. Concentration, disclosure and the redemption gate are where this asset’s risk actually sits, and none of them improved.
What’s backing sUSDai
We verify this on-chain rather than taking the protocol’s number. As of 2026-08-12, totalAssets reports $354.99M, and it decomposes exactly: $92.37M of idle USDai sitting in the vault as the redemption buffer, plus $262.62M of drawn GPU loans. ⚠️ That “adds up exactly” check is weaker than it sounds, and it is worth being precise about why. The drawn-loan figure is not measured — it is totalAssets minus idle. It ties to the total by construction, so the arithmetic reconciling is not evidence of anything. The only independently sourced quantity in the decomposition is the issuer API’s loan figure, and that is the one that disagrees.
⚠️ Re-read 2026-08-27, the gap is $11.1M — 3.094% of assets — and the finding is no longer the gap itself but that it has not closed. The vault’s totalAssets implies materially more deployed than the issuer’s API reports.
⚠️ It stepped on 2026-08-05, from 0.512% to 2.708% — roughly $1.6M to $11.1M — and twenty-two days later it is still there, at 3.094%. A step alone would not be notable, and this report should not treat it as the anomaly: there is a control. An earlier and larger step on 2026-06-05 (1.111% → 4.425%) reverted within six days. So the shape that matters is the absence of re-convergence, not the step.
One explanation is eliminated rather than merely doubted: committed-undrawn escrow. The escrow read $0.00 for the 54 days before the step, so it cannot account for it. ⚠️ What remains is explicitly unranked, and this report will not pick between them — a timing difference in what the API counts as deployed, an accounting treatment the vault applies and the API does not, and a genuine divergence are all consistent with what can be observed. Naming a favourite here would be inventing a cause for a measured effect.
Which side is wrong has not been established, and this report does not guess. A continuously amortizing loan book can legitimately produce timing differences between an issuer’s snapshot and a vault’s accounting. But the direction matters: the vault reports more assets than the issuer reports lending out, and that is the direction where the error, if it is one, sits in the holder’s disfavour over.
The shape of the book has changed since this report was first published. In May, a large slice of capital sat in a DepositTimelock escrow — committed to specific loans but not yet drawn by the borrower. That escrow is now empty: the pipeline drew down in full, drawn loans rose to $262.62M, and there is no committed-undrawn dry powder left in it. That is an operational fact rather than a problem, but it is informative. The book is now fully deployed, and the idle buffer — a little over a quarter of assets — is the only liquidity standing between a redemption request and a loan repayment schedule.
Loan terms, from protocol documentation rather than chain: 70–80% stated LTV, interest tiers running from 7–9% on investment-grade offtake up to 12–15% on on-demand compute, 36-month amortizations with 30-day payment cadence and a single 30-day grace period at an accelerated rate before default. A debt-service reserve of roughly 10% of each loan is funded at close. Collateral is represented on-chain as loan NFTs, with legal enforcement running through Delaware LLC SPVs and off-chain UCC bailment.
One thing the vault does not make obvious: sUSDai is a four-chain asset. The token is live at the same address on Arbitrum, Ethereum, Plasma and Base, and just over a third of all shares sit off Arbitrum — about 118M of roughly 323M as of 2026-08-18, with Ethereum holding around 60% of that bridged float, Plasma most of the rest and Base a small tail. The accounting ties, and a holder can check it without trusting us: the Arbitrum contract’s own bridgedSupply() reads 118,420,253 against a sum of 118,420,250 across the three bridged chains — a residual of three shares out of 118.4M. The bridged deployments are deliberately thin. Arbitrum’s implementation is 24,345 bytes, sitting at the EIP-170 contract-size ceiling; Plasma’s is 9,261 and Ethereum’s and Base’s 8,249, and their function sets decode to a plain ERC-20 with mint. The economics — NAV, the loan book, the redemption queue — all live on Arbitrum, so a bridged share is a token representation of the Arbitrum vault rather than a separate vault with its own book. Bridging runs on LayerZero V2 through an OFT adapter, not a bespoke bridge.
The caveat is worth stating plainly: a bridged holder depends on the bridge as well as on the vault. Arbitrum is where the value is determined and what the Cantina audit covers; holding the share anywhere else puts a messaging layer between you and it.
Why GPU collateral is scored at 4.5
This is the heaviest discount on the rating and the reason the underlying axis sits well below the vault’s structural marks.
Hardware depreciates on a technology cycle. Each AI hardware generation devalues the one before it. These are 36-month amortizations against hardware with a useful economic life on the order of three to five years, so the collateral is racing the loan. That is a fundamentally different profile from a T-bill or a liquid crypto collateral pool, and it is not something a higher LTV cushion fixes.
You cannot compute LTV yourself. The chain shows loan principal; it does not show the appraised value of the hardware behind it. The 70–80% figure is a trust input from the protocol and its appraisers, not a verifiable on-chain ratio. Recovery, if it is ever needed, means physically de-installing hardware from a data center, enforcing through an SPV, and selling into a thin secondary market for used AI compute.
Concentration and disclosure are the live problems. Public risk write-ups state that two borrowers account for very nearly the entire pipeline — one of them duosEdgeAI, a subsidiary of Nasdaq-listed duostech, on a loan around $98M. A single borrower default would therefore be a protocol-level NAV event, not a diversified loss absorbed by the spread. And no numeric delinquency or default rate is published. The protocol reports aggregate active-loan totals; it does not report what share of them are paying late. For a credit product that is a material diligence gap, and it is the single thing whose disclosure would most change this report.
The mitigants are real but bounded: the debt-service reserve, a staked-CHIP first-loss layer whose capacity relative to the book is not disclosed in granular form, and a Barkr warrantied collateral-value schedule set at origination with an institutional reinsurance layer covering liquidation shortfalls up to 80% below that warrantied value. That reinsurance is a genuine loss-severity mitigant. It does not reduce default probability, and it covers collateral shortfall rather than borrower default as such.
NAV, yield, and a trap on the block explorer
sUSDai is a vault share, so the metric is discount to net asset value, not price against a dollar. The deposit-side price read on-chain at $1.1046 on 2026-08-12, against $1.0875 on 2026-05-28 — 76 days apart, which works out to roughly 7.8% annualised, smooth, with no jumps. That is modest for a loan book grossing more than that, and the gap is explained by fees plus the drag of holding a quarter of assets idle as the redemption buffer. Secondary prices have tracked NAV closely in calm conditions.
sUSDai has two share prices, and which one you want depends on what you are asking
⚠️ There is no single “NAV” here. The vault publishes two prices and they answer different questions. Read 2026-09-12 on Arbitrum:
| getter | reads | what it answers |
|---|---|---|
depositSharePrice() | 1.111760 | what a new depositor pays |
convertToAssets(1e18) | 1.111760 | ⚠️ identical to the deposit price |
redemptionSharePrice() | 1.107063 | ✅ what an exit settles at |
totalAssets / totalSupply | 1.566819 | ⚠️ wrong — 41.5% above the redemption price |
⚠️ So if you are modelling an exit, use redemptionSharePrice(). The gap is 42.4 bps, and convertToAssets — the getter a careful reader would naturally reach for, and the one most explorers surface — returns the entry price.
Quote the gap, timestamp the level. The deposit price accrues continuously by construction, so the absolute numbers move between any two reads: 1.104646, then 1.111753, then 1.111760 over a few days. The ~42bp spread has been far more stable than either level, which makes it the more useful figure to carry.
✅ The divergence is deliberate and USD.AI documents it. Their withdrawal-estimates page states that the deposit price “continuously prorates expected upcoming repayments across all loans in the book” while the redemption price “is based on settled repayments and realized vault accounting.” The stated purpose is symmetric and runs in favour of existing holders: a new depositor “enters at a price that already reflects the yield expected to accrue” so they cannot buy in just before a repayment and capture a step someone else earned, and a redeemer “does not receive value before it has actually been received by the vault.” It is offered as an alternative to a fixed redemption fee, not as a haircut.
⚠️ That it is documented does not make it costless to a holder. Two prices are a thing you have to model, the redemption leg settles through a 30-day FIFO queue, and the priority-exit auction is not live on-chain. Disclosed complexity is still complexity.
And do not compute NAV as totalAssets / totalSupply. That gives about $1.57 and is badly wrong in the flattering direction. ⚠️ The naive division is the obvious thing to try and it returns a plausible-looking number — which is exactly what makes it dangerous for a reader doing their own homework.
The reason the two differ is that 203.42M shares represent only about $224.70M of a $355M pool — a gap of roughly $130M between assets held and claims outstanding. To be clear about what is and is not known here: the assets side reconciles completely, as set out above, so nothing is missing or overstated. What is unexplained is the claims side. The most likely reading involves deposits that have been received but whose shares have not yet been minted under the vault’s async settlement model, which would fit an 18% AUM increase in three weeks — but the vault exposes no getter that confirms it, and we do not assert a cause we cannot verify. We state the trap and the correct getter; the cause is an open item, tracked and not resolved.
Exit liquidity
Two paths, both with friction, and they tighten together under stress.
Primary redemption is an epoch-based FIFO queue. ⚠️ The structure is on-chain verified and the timing is not. sUSDai is a genuine ERC-7540 asynchronous vault — pendingRedeemRequest, claimableRedeemRequest and maxRedeem are all present — so requests queue, settle asynchronously, and can roll to a later epoch. That part is measured. ⚠️ The 30-day epoch length is documented by the issuer, not verified on-chain: no epochDuration, currentEpoch, epochLength, noticePeriod, withdrawalDelay or cooldownDuration exists in the deployed bytecode. Treat 30 days as the issuer’s stated figure. The structural risk is the same at seven days or thirty — it is the rolling, not the number. On the issuer’s stated schedule, requests queue through roughly 29 days, process on day 30, and fill in order until available USDai is exhausted. The protocol explicitly does not force-liquidate GPU loans to meet redemptions. That protects remaining holders and the loan book, at the cost of exit certainty for the holder leaving: under pressure your withdrawal can partially fill and roll into the next epoch, possibly more than once. The planned QEV priority-exit auction is still not live. The idle buffer is the only on-demand liquidity behind the queue.
Secondary depth roughly doubled this cycle, and that is what moved the Liquidity axis from 4.5 to 5.0. Across the venues that carry it: Fluid sUSDai/USDC now holds about $17.93M, a new Fluid sUSDai/USDT pool adds about $17.66M, Curve sUSDai/USDC holds about $1.55M, and a long tail across Uniswap v4 and Maverick adds roughly $0.32M — about $37.5M total against $17.05M three weeks earlier. Depth as a share of assets improved from 5.68% to 10.55% despite the vault growing 18%. The previous 4.5 rested on the argument that liquidity was thin for the size on either path; with two Fluid pools of comparable depth, the single-venue leg of that argument no longer holds.
It does not go above 5.0, for four reasons. There is no centralized-exchange listing. Pool TVL is two-sided, so executable depth is roughly half the headline figure. Exiting even 5% of the vault would move the market against you. And the primary exit is still the 30-day queue. The figures above are pool TVL — the same basis used previously, which is what makes the comparison honest — not a measured slippage ladder.
One consequence worth stating for anyone using sUSDai as collateral elsewhere: check what oracle the venue prices it with. If a lending market marks sUSDai at market price and the secondary detaches under stress, leveraged positions can liquidate even while the loan book’s NAV is intact.
Admin, audit and issuer
The 48-hour delay on this asset is binding, not theatre, and that distinction is worth making explicitly because plenty of protocols have a timelock that can be shortened at will. sUSDai’s ProxyAdmin (0x0b3296b6…) is owned by a TimelockController at 0x0EEA1EE0…639b with a minimum delay of 172,800 seconds — exactly 48 hours — re-verified on-chain 2026-08-12. Critically, the timelock holds its own admin role, so the delay cannot be shortened and no new role can be granted without first passing through 48 hours. Proposal authority sits with a 3-of-3 Safe multisig with three known signers, and no role revocations have occurred since deployment. USDai’s ProxyAdmin (0x2ddf39c7…) is owned by the same timelock, so the two legs share one authority chain — this is the same test the Saturn reports apply, and sUSDai passes it.
Practically: a malicious or buggy upgrade is visible on-chain for two days before it can land. Two days is a real observation window, though it is shorter than the 30-day redemption queue — so for a holder relying on primary redemption, the window lets you see the change coming, not necessarily exit ahead of it. The secondary market is the exit that fits inside 48 hours.
Contract & Admin is scored 3.0, cut from 5.5 on 2026-09-09. ⚠️ No new evidence about sUSDai produced this. The finding is the one stated in the paragraphs above and on this page since 18 August: the 48-hour delay governs the upgrade path only, while role grants — including whatever gates minting — are an undelayed 3-of-3. The score simply had not followed its own rationale. ✅ Re-verified on 2026-09-09: the timelock’s minimum delay is exactly 172,800 seconds and the ProxyAdmin is genuinely owned by it, but the governance and operations Safes turn out to hold the identical three owners, so the two-Safe arrangement separates no duties.
The vault’s own surface is more complex than a plain ERC-4626 — async request-and-fulfill deposits and redemptions, a redemption queue, blacklist and freeze capability, and a Chainlink oracle dependency — and Cantina (Spearbit) audited it. ⚠️ That extra layer is NOT charged for here, and the omission is deliberate rather than an oversight. A wrapper is scored equal to its underlying on the underlying’s own axis rather than notched below it, so this sits at USDai’s 3.0 and no lower. If anything that makes 3.0 generous for the wrapper, since sUSDai carries everything USDai carries plus a vault of its own.
Issuer sits at 5.0, matching USDai, since it is the same entity, the same governance and the same audit — and it moved with USDai on 2026-08-26. ⚠️ The finding is that USDai’s 48-hour timelock governs upgrades but not issuance: the three keys behind both Apyx-style Safes can mint USDai without delay, either by granting BRIDGE_ADMIN_ROLE (they hold DEFAULT_ADMIN on all four chains) or via setPeer on the LayerZero OFT adapter they own, which is itself a mint. That reaches this vault through its underlying, since sUSDai’s assets are USDai: undelayed issuance dilutes the pool a share is a claim on, even though it touches neither the vault contract nor the PYUSD reserve. See the USDai report for the measurement, its coverage limits, and why this is 5.0 rather than lower. Permian Labs carries MetaStreet lineage and a $13.4M Series A from Framework, Dragonfly, Coinbase Ventures and Arbitrum. It is held down by a roughly ten-month track record, single-operator concentration, and no stress event to date.
Who this is for
- Someone explicitly underwriting GPU-credit risk who wants that exposure tokenized, and who is sizing it as a small, illiquid sleeve rather than a yield allocation.
- Holders comfortable with a multi-week exit and with the possibility that the exit rolls to a second epoch under stress.
Who this is NOT for
- Anyone treating this as a stablecoin substitute or a cash-equivalent yield token. It is neither. The USDai report covers the leg that actually is a dollar.
- Anyone who needs to exit size quickly. About $37M of two-sided pool depth against a $355M vault means executable size is a fraction of that, and the alternative is a 30-day queue.
- Anyone using it as high-LTV collateral on another protocol. The same exit asymmetry that affects a retail holder applies to a forced unwind, and the oracle question above decides whether a NAV-intact position can still be liquidated.
- Anyone who needs to see delinquency data before underwriting credit. It is not published.
What to watch
- Any disclosure of delinquency or default rates. The single most score-relevant thing that could appear. The book has not yet run through a default cycle.
- Borrower concentration. New borrowers diversifying the pipeline would ease the volatility axis; further concentration into the existing two would tighten it.
- Whether redemptions ever roll. An epoch that partially fills is the practical test of the redemption gate, and it has not happened yet under observable conditions.
- Whether the residual re-converges. ⚠️ This is now the single most informative thing to watch on this asset. The June step closed in six days; this one has stood for twenty-two. A return toward zero would retire the question; another month at 3% would make it structural rather than a timing artifact. - The claims-side gap. If the protocol exposes a getter for pending deposits, or the gap closes on its own, the accounting picture resolves. If it widens without explanation, that is a structural re-rate.
- Anything queued in the timelock. Scheduled operations are visible for 48 hours before they can execute; the live dashboard surfaces pending operations with their countdown.
- Secondary depth holding. The new Fluid USDT pool is what carried the liquidity upgrade. If it drains back toward single-venue concentration, the 5.0 does not hold.
Live dashboard
A stacked-bar decomposition of total assets, the credit-pipeline framing, NAV trajectory, secondary-market depth across venues, and the 48-hour pending-upgrade watch are on the live dashboard. ⚠️ One scope note on that last panel, because it is the one a holder would rely on: the pending-upgrade watch covers scheduled upgrades only. It reads CallScheduled on the timelock — and the mint path never touches the timelock, so an undelayed issuance will not appear there and gives no advance warning. The dashboard itself now says so; this sentence previously did not.
Sibling
- USDai — the PYUSD-backed reserve leg, carrying no GPU-loan exposure and a materially different risk profile
This report is built from publicly available documentation and independent on-chain reads, most recently on 2026-08-12, with the multi-chain supply figures verified 2026-08-18. We hold no privileged information about the issuer or borrowers. Per-loan detail is partly reconstructable from chain, but borrower identity, GPU specifications and appraised collateral values are off-chain and not independently verifiable. Corrections welcome: info@tidresearch.com.
Revision history: 2026-08-27 — redemption epoch attributed; scores held at 5.0. The 30-day epoch is the issuer’s documented figure, not an on-chain fact — no epochDuration, currentEpoch, epochLength, noticePeriod, withdrawalDelay or cooldownDuration exists in the deployed bytecode. What is verified is the ERC-7540 asynchronous structure (pendingRedeemRequest, claimableRedeemRequest, maxRedeem), so requests queue, settle asynchronously and can roll. The reconciliation residual is $11.1M, 3.094% of assets, against $8.40M / 2.35% at the prior read. It stepped on 2026-08-05 from 0.512% to 2.708% and has not re-converged in twenty-two days — an earlier and larger step on 2026-06-05 (1.111% → 4.425%) reverted within six. Committed-undrawn escrow is eliminated: it read $0.00 for the 54 days before the step. Remaining explanations are unranked. The undelayed-mint finding on USDai does not extend to this vault — a pair simulation reverts identically (0xe6c4247b) from both the operations Safe and a dead address, so it fails a precondition before any access check and nothing about the mint gate is established here. 2026-08-27 — Issuer 6.0 → 5.0, following USDai, same entity and governance; Overall held at 5.0 on vault-specific drags. 2026-08-18 — chain footprint: sUSDai is live on Arbitrum, Ethereum, Plasma and Base, not Arbitrum-only. 2026-08-12 — Liquidity 4.5 → 5.0, Overall held at 5.0. 2026-07-23 — loan book roughly doubled; borrower concentration surfaced, two borrowers close to the entire pipeline.