2. FiLo Curator [Scale]
First Loss Curators help scale USD.AI by originating and legitimizing loans for the protocol
First Loss Curation Model
Inspired by Morpho’s modular curator model, USD.AI advances this by requiring curators to take a first-loss position in exchange for higher yield. These positions are illiquid and do not share the same benefits as the more liquid USDai & sUSDai tranches.
In other, words the "FiLo Curator" takes significantly more risk, often originates the loan, undergoes protocol governance for onboarding, and is the core risk alignment design that scales the curation & origination of the yields, while USDai takes care of the structuring and distribution.
This model ensures USD.AI can onboard loans quickly to scale, while also ensuring depositors and stakers are protected from underwriting risk
Underwriting Process and Standards
Underwriting Hierarchy (most stringent to least):
Onboarding a new underwriter
Onboarding a new industry vertical
Onboarding a new asset type (e.g., Rockchip vs B200)
First-time borrower (requires insurance structuring and bailment design)
Repeat loans to onboarded borrowers (can happen overnight)
The CALIBER tokenization framework accelerates steps 3–5 by ensuring bankruptcy remoteness and clear asset tokenization. Every deal must include a first-loss tranche to align incentives and risk exposure. We plan for a controlled rollout: a small group of participants at launch, expanding as live deal data validates the infrastructure.
USD.AI never extends credit to any business as a counterparty — only to the collateral itself.
Collateral Overview & Market Context
Unlike bitcoin mining rigs, which are highly exposed to commodity pricing and electricity costs, AI GPUs benefit from diverse global demand — sovereign AI infrastructure, foundation model training, cloud inference, HPC workloads, and leasing data centers. This broad resale and redeployment optionality reduces systemic price collapse risk.
Why BTC rigs are riskier:
Dislocation of equipment (single use cases, difficult to transport)
Extremely high revenue concentration (convexity to bitcoin price)
Dependency on input prices (energy costs)
By contrast, AI hardware serves multiple global industries and sovereign infrastructure demand, making its resale market deeper and more resilient to shocks.
Collateral Risk Management Safeguards
Aggressive Amortization
The primary safeguard is an aggressive amortization schedule, which front-loads principal repayments to reduce exposure regardless of short-term FMV swings. While industry estimates suggest 5–7 year economic lives for top-tier GPUs (NVIDIA: 7, CoreWeave: 5), USD.AI underwriters use a more conservative 3-year depreciation curve. For example, an LTV that starts at 70% could drop to ~40% after one year.
Waterfall Capital Stack
Each loan uses a layered capital stack:
Equity cushion from the borrower
First-loss tranche from the underwriter
Rapid principal amortization
This ensures that any valuation shocks or downturns are absorbed by the borrower and curator before they affect sUSDai holders. The rigid CMBS-like terms mean that missed payments immediately trigger forfeiture and recovery, adding further protection.
3rd Party Appraisals
Secondary market pricing is sourced from independent providers, including Evertas and Blockware. Additionally, the semi-fungibility of top-tier GPUs (e.g., B200, GB200) allows for standardization in valuation and risk modeling across asset classes, as these are rolling out of the factories at this moment in time (we also have the valuation calibrated via Evertas’s continuous valuation checks). Live pricing is a core feature of DeFi, but is not tenable in physical assets - which is why we originally built an oracleless NFT lending in the first place.
There are no oracles — aggressive amortization schedules, underwriters, and independent appraisals protect the protocol from collateral pricing risk
Example showcase loans:
TACOM borrow with RK3588s ($600k across two loans)
Lyceum Technologies amortization schedule ($2.6m loan)
Other Risk Management Safeguards
Operational Role
USD.AI is never the operator of the underlying hardware. It acts solely as a bailor, with ownership structured via CALIBER but operations externalized to borrowers — eliminating asset servicing risk during the loan term.
Hardware Age
Most loans are with newly installed hardware. Assets are largely semifungible, so LTVs are adjusted linearly as hardware ages. Durations always target the shorter range versus market norms to reduce exposure.
30 Day Refinancing Windows
In line with market standards, there are no liquidations on the borrower side, but only defaults. 30-day refinancing windows means hardware value would have to drop by ~40% in that window to trigger an economic default. Even then, the protocol’s pessimistic assumptions (NVIDIA’s 7-year lifespan vs. our 3-year amortization) provide an additional margin of safety. Depositors should be comfortable with this conservative risk modeling — similar to how Ethena depositors accept possible negative rebasing if the basis collapses.
Rigid CMBS-Like Logic
Like the CMBS market, USD.AI operates on rigid terms: miss a payment, forfeit the asset. This unforgiving structure is essential for scaling a deep and liquid credit market.
While amortization, FiLo and appraisals are the most important protection for the protocol, these additional safeguards provide an additional layer for USD.AI participants
Default Enforcement & Recovery
In the event of borrower default, collateral is routed through two steps:
A 30-day refinancing grace period
Then a 7-day auction
Recovery is fully automated via smart contracts — no governance or discretionary intervention. This way the sUSDai holders never hold the asset on the books, it is always placed into default and cashed up instantly.
Auction Process
We are actively onboarding hardware partners who can assist in recycling collateral for redeployment (this has been the easiest way to get buy-in from industry participants). Often times, the underwriter will provide the stalking horse bid in majority of these defaults.
Last updated