KETJU Research

← The Register

synthetic-yield

USD.AI (sUSDai)

Rejected
Max sleeve
—
Reviewed
2026-08-14 · v1
Next review
2027-07-31
Research basis
Individual research
Chains
Arbitrum One · hybrid

Rejected venues wait the longest for re-review; a rejection has to earn another look before the scheduled date.

Rejected for sUSDai; this verdict does not silently classify base USDai as the same credit claim. Two tokens share the name. USDai is the non-yielding liquid wrapper; sUSDai, the token the yield marketing sells, is a share in a credit fund making non-recourse loans to AI infrastructure operators at 70-80% loan-to-value, secured on GPU hardware. The lender of record is GPU Finance Ltd.; the Cayman-registered USD.AI Foundation operates the protocol, and core contracts sit on Arbitrum behind a multisig whose signers and timelock the docs do not publish. The yield story depends on the date. As late as the CHIP token launch in late 2025, roughly 99% of backing sat in Treasuries with about 10% of deposits lent, per Pine Analytics, and the docs disclose a 4.5% PayPal incentive on the PYUSD float flowing into sUSDai yield: early holders earned mostly T-bills plus a subsidy. Utilization has since risen toward 60%, with roughly $261M of $430M in deposits out in loans at the 2026-08-14 review, so the yield is now credit compensation. The 13-17% headline is the protocol’s mature-stage marketing; live APR was 8.90% the same day. The collateral structuring is real and named: UCC Article 7 warehouse-receipt NFTs (legally untested), Alliant property and casualty cover, Barkr residual-value reinsurance, a debt service reserve, amortization toward roughly 40% LTV, ITAD resale partners. None of it has been through a default. There have been zero defaults, and the docs themselves decline to model coverage carve-outs and insurer counterparty risk. The exit decides the verdict. Redemptions run on a global 30-day FIFO epoch; the docs say queues can extend across multiple epochs and loans are never liquidated to meet redemptions. The on-demand secondary exit was a $3.76M Curve pool against roughly $430M of deposits at the review. The CHIP backstop the docs call the last line of defense was a $46M market cap token, down about 84% from its April 2026 high. TVL peaked at $702M on 2025-11-21; about $270M left between January and April 2026, and no source explains why, so the review holds it open. The name is the finding. The homepage sells a fully-backed synthetic dollar while the docs’ fine print says sUSDai is not a stablecoin and redemption may not be possible at all. A client who sizes it like cash has mistaken the wrapper for the exposure. It is a venture-adjacent private credit fund with a 30-day gate, and honest ways to buy AI exposure exist.

The research file

The mechanism

Two tokens sit behind the name, and the distinction carries the file. USDai is the base token: users exchange a currently supported stable asset for USDai. It pays no yield. The current FAQ names USDC and USDT as redemption examples; the earlier PYUSD-only description is not a durable product definition. Direct minting and redemption at the contract is currently permissionless, but the docs say it will be restricted to whitelisted, KYC-verified market makers and institutions in Q2 2026, and the technical overview already describes minting and burning as something KYC’ed institutions do. sUSDai is the staked version, an ERC-4626 vault token with ERC-7540 async redemption whose share price appreciates. The docs state plainly that sUSDai ”is not a stablecoin, but is a free floating token, representing shares in an assortment of targeted lending positions and unallocated USDai.”

Yield comes from two streams: interest on GPU-backed loans, and yield on idle reserves. On the second, the docs’ partners page discloses that PayPal pays a 4.5% annual incentive on PYUSD held in the protocol, on up to $1B of backing for 2026, and that the incentive accrues at the protocol level and flows through to sUSDai yield. Part of the headline yield is a PayPal marketing subsidy on the reserve float, not credit compensation.

The loans are non-recourse, secured per the docs ”by the physical GPU assets themselves along with their contracted cashflows,” originated at 70 to 80% loan-to-value, with a debt service reserve account of about three months of peak debt service, 30-day repayment cycles, and terms around three years. Third parties report amortization brings effective LTV toward roughly 40% after a year. Borrowers sit in bankruptcy-remote SPVs; the lender of record named in the terms of service is GPU Finance Ltd. The CALIBER framework is real as described: hardware is documented as warehouse receipts under UCC Article 7 and tokenized as NFTs representing a claim on specific hardware, and loans are issued against the tokenized receipts. CoinRank notes this application of UCC Article 7 is untested in court.

Origination runs through what third parties report as a FiLo Curator model: first-loss curators underwrite and originate loans and post first-loss capital. The docs describe underwriting, appraisal, and monitoring, with Aravolta providing hardware monitoring of usage, uptime, and location, but do not name the curators. Marketing copy on usd.ai calls it ”Decentralized Underwriting: Independent experts vet every borrower.”

On rates, the 13 to 17% figure is the protocol’s own mature-stage marketing (docs FAQ: 13-17% APY, ”up to 20%”). The site showed 8.90% current APR and 11.86% expected on 2026-08-14; Pine Analytics put the then-current APR at 6.87% at the CHIP launch. On the borrower side, third parties report smaller GPU operators pay roughly 15 to 20% APR.

The exit

Redemptions run on a 30-day epoch with a FIFO queue, a global protocol timer rather than a per-user countdown. Unstaking is an async ERC-7540 requestRedeem; redemptions are serviced when the strategy admin calls serviceRedemptions() with USDai available. There is no instant exit at the protocol. The docs say that under high utilization queues may extend across multiple epochs, and that the protocol does not prematurely terminate or liquidate active GPU loans to satisfy redemptions; loans run to maturity or scheduled paydown. The terms of service say it without varnish: ”redemptions may be delayed, restricted, or may not be possible at all. You may not be able to redeem your tokens when you want or at the value you expect.”

Redemption pricing uses the conservative NAV, counting settled repayments only, while deposit pricing uses the optimistic NAV: a redeemer waits for cash that has actually arrived. QEV, an auction to bid for queue position that pays fees to those who stay, is planned but not live; as of the docs read on 2026-08-14 the FIFO queue is what exists.

Secondary liquidity is thin relative to size. GeckoTerminal showed the Curve sUSDai/USDC pool on Arbitrum at about $3.76M of liquidity, with sUSDai at $1.0835; sUSDai has an appreciating NAV, so spot price alone does not give the premium or discount to NAV. Pendle listed four USDai and sUSDai markets on Arbitrum, launched 2026-02-19, with $79.5M total liquidity and a best fixed APY of 9.46%, but Pendle PTs are a fixed-maturity instrument, not an at-will exit at par. Against roughly $430M of deposits, the on-demand exit is a single-digit-millions Curve pool.

In a redemption rush the design outcome is a lengthening FIFO queue serviced only by loan amortization, roughly 3 to 4% of outstanding principal returning per month per the docs, plus whatever reserves are unallocated, with the Curve pool as the only immediate door.

The collateral

A loan is backed by specific GPU clusters in insured datacenters, tokenized as UCC Article 7 warehouse receipts, plus the borrower’s contracted revenue, debt service reserve cash, and a parent pledge; the terms of service describe stepping into the parent’s shoes on default. Origination LTV is 70 to 80%, amortizing, with third parties reporting roughly 40% effective LTV after a year of paydown.

The insurance is layered and named. Alliant Insurance Services places property and casualty cover on the clusters. Barkr provides value reinsurance, a residual-value policy: if a defaulted borrower’s GPUs sell below a warrantied price on a predefined residual-value schedule, the insurer pays the delta. The NAV math in the technical overview models this payout, but the docs themselves caveat that ”coverage carve-outs, claim-eligibility conditions, and insurer counterparty risk are not modeled.” Recovery is capped at the frozen claim; the protocol keeps no upside. After the on-chain auction of the collateral NFT, ITAD (IT asset disposition) partners handle physical retrieval and resale.

The whole mitigant stack is built around the premise that the hardware loses value fast; Pine Analytics states it directly, that GPU collateral depreciates quickly. No public source gives realized recovery data, because there has been no public default.

Third parties name Hydra Host, Lyceum, Compute Labs, and TACOM as borrowers or operators. The docs describe a Proof of Reserves dashboard with per-borrower detail, but this review did not capture a per-borrower breakdown. The homepage claimed a $236M active loan pipeline, and DefiLlama showed $261M borrowed at the review. The protocol’s own terms carry the enforcement caveats: bankruptcy automatic stay can delay foreclosure; cross-border collateral under UCC, Canadian PPSA, and other named regimes complicates enforcement; and the terms disclaim any assurance ”that the net proceeds of any such enforcement will be sufficient to repay in full.”

Who controls it

The terms of service are a contract with USD.AI Foundation, a Cayman Islands foundation company that operates the protocol. GPU Finance Ltd. is the lending entity. Permian Labs Inc. of New York, founded by David Choi (CEO) and Conor Moore (COO), is the developer, per IQ.wiki. The terms impose mandatory individual arbitration with a class-action waiver, and state that USDai and sUSDai have not been and will not be registered under the U.S. Securities Act.

Strategy operations, meaning deploying loans, harvesting yield, and servicing redemptions, require STRATEGY_ADMIN_ROLE, which the docs say is scheduled offchain and executed by a multisig, to become governance-driven in the future. A Timelock Controller and a Chip Governor contract exist on Arbitrum. The docs’ disclosures page asserts that no role has discretion to pick redeemers, redirect payments, or alter yield allocation outside protocol rules. Multisig signers, threshold, and timelock delay are not published in the docs read for this review.

The terms reserve wide discretion: USD.AI Foundation may, ”in its sole and absolute discretion,” modify the protocol at any time, including fee structures, reward rates, and yield parameters, and may suspend or terminate service ”for any reason… or for no reason.”

The hub is Arbitrum, where all core contracts live; USDai and sUSDai exist as LayerZero OFTs on Ethereum, Base, and Plasma, and CHIP also on Solana. Using the service requires agreeing you are not sanctioned, not on OFAC or similar lists, and not using a VPN to disguise location; no clause barring US persons as such was found. Staking USDai into sUSDai is permissionless today, but direct mint and redeem of USDai is being restricted to KYC-verified market makers and institutions in Q2 2026 per the docs, which pushes retail exit toward secondary markets.

The record

DefiLlama first shows TVL on 2025-05-18 at about $0.9M. Growth was fast and incentive-led: $62.7M by late August 2025, two $250M deposit caps filled by 2025-09-11, and a peak TVL of $702M on 2025-11-21. Permian Labs raised a $13M Series A on 2025-08-14 led by Framework Ventures, with Dragonfly, Arbitrum, DCG, Delphi, and Fintech Collective; YZi Labs, formerly Binance Labs, invested 2025-08-26.

Then the decline. DefiLlama monthly closes: $684M in December 2025, $641M in January 2026, $467M in February, $311M in March, $250M in April, $293M in May, $196M in June, $160M in July, and $174.9M on 2026-08-14, plus $261M in outstanding loans DefiLlama books as borrowed. The usd.ai homepage showed $430M of total deposits the same day, consistent with DefiLlama’s TVL plus borrowed of about $436M. Deposits are down roughly 40% from the peak while the loan share of backing has risen. No reporting attributes the outflows to a specific event; the points-farming origin of much of the TVL, documented by Pine Analytics, and the CHIP token’s launch and decline are the obvious candidates, unproven.

The utilization history changes what early holders actually owned. Pine Analytics documented that around the CHIP launch in late 2025 about 99% of backing sat in US Treasuries with roughly 10% of deposits deployed to loans: depositors were then earning mostly T-bill and incentive yield under an AI-credit label. Figures at this review, roughly $261M borrowed against $430M of deposits, imply utilization near 60%. The credit exposure this memo describes is now the majority of the book, more than at any earlier point.

CHIP launched at a reported $300M fully diluted valuation; CoinGecko shows an all-time high of $0.1402 on 2026-04-23, an all-time low of $0.0216 on 2026-08-09, and $0.0229 on 2026-08-14, down about 84% from the high, at a $45.8M market cap. The docs describe CHIP staking as a ”last line of defense… ultimate risk buffer in case of major depeg,” in the Aave and Maker style. At today’s market cap that backstop is small against $430M of deposits.

No default, loss, or gate beyond the standing 30-day epoch design was found in any source as of 2026-08-14, and nothing new surfaced between the original 2026-07-31 review and this deepening. Coingape reported over $20M of cumulative yield paid. On audits, the docs’ audit page hosts eight reports as file attachments without firm names or dates in the page text; search results attribute reports to Cantina (2025-05-12), Quantstamp (2026-02-03), and others dated 2025-10-29, 2025-11-27, and 2026-03-12, but this review could not open the files to confirm firms or scope. A live bug bounty runs on Cantina with its cap behind a login; IQ.wiki also mentions an Immunefi program.

The name

The material read at this review bears out the naming finding, with one precision. The protocol itself draws the line the branding blurs, but only in the fine print. usd.ai markets USDai as a ”Fully-backed synthetic dollar” under the brand line ”The dollar that builds AI.” The docs state that sUSDai is not a stablecoin, not instantly redeemable at par, and not pegged 1:1 to any other asset, and the terms of service warn in capitals that holders ”MAY LOSE SOME OR ALL OF YOUR DEPOSITED FUNDS.” The homepage headline and the risk page describe two different instruments, and the client sees the headline.

The precision: the wrapper-versus-exposure point lands on sUSDai, the yield product the 13 to 17% marketing sells. The base USDai token is closer to a genuine PYUSD wrapper; the yield token is the credit fund. The defense that USDai is fully backed is true and irrelevant to the thing a client would buy for yield. The sizing risk is reinforced by the exit facts: a 30-day global epoch, a multi-epoch queue under stress, and a $3.76M Curve pool against $430M of deposits.

Open questions

Audit firms, dates, and scope. Eight reports exist as attachments and the page text names no firms; the third-party attributions to Cantina and Quantstamp are secondhand, and the PDFs must be read before any memo cites an audit record. The Cantina bug bounty cap is behind a login.

Per-borrower concentration was not captured. The Proof of Reserves dashboard reportedly shows loan-by-loan detail, and the four named borrowers come from third-party coverage, not the docs. Whether any FiLo curator first-loss capital is posted today, and how much, is not quantified in the docs pages read. The Barkr residual-value schedule of warrantied prices by hardware age is not public, and the insurance claim is only as good as that schedule and the carve-outs the docs decline to model.

Multisig signers, threshold, and timelock delay for the strategy admin and the Timelock Controller need an on-chain read. Whether sUSDai trades at a premium or discount to NAV is unknown: the Curve spot of $1.0835 means nothing without the same-moment redemption share price, which also needs a contract read. Whether the Q2 2026 KYC restriction on direct mint and redeem is already enforced at the contract was not verified.

The cause of the roughly $270M deposit decline from January to April 2026 is unexplained; a points unwind and CHIP disappointment are plausible, not established. Stablewatch’s deep dive returned 403 at both URLs and is cited only via search excerpts.

Sources

The claims above trace to these. Where a number could not be independently verified, the thesis says so.

Inherited controls

The verdict above grades the protocol layer. Every position also inherits the asset it holds and the chain it settles on. The least safe layer sets the position’s grade, and the position table names which one that is.

ChainVerdictGradeControl constraint
Arbitrum OneApproved · limits hybrid a single sequencer orders >99% of transactions and admin keys can upgrade bridge contracts on a ~7-day timelock.
The memo is public. The watching is the product: the terminal reads your clients’ wallets against this Register and flags the events above when they fire. $49 per advisor per month, first 14 days free. Start the trial.