A common language for compute and credit
Plain-language definitions connect the hardware, operating model, and financing terms used throughout USD.AI.
Hardware and compute economics
A chip built for parallel math workloads such as AI training and inference. GPUs are the most common type.
Compute rented as whole physical servers, with no virtualization layer shared with other customers.
The average price earned per GPU-hour across all contracts, mixing reserved and on-demand pricing.
A group of GPU servers networked together to run as one system, usually for large training jobs.
Renting space, power and cooling in a third-party data center to house your own servers.
The total amount of GPU processing an operator can sell, usually stated in GPUs or GPU-hours.
How heat is removed from servers. Newer high-density systems increasingly require direct liquid cooling.
The product family a GPU belongs to, such as NVIDIA Hopper or Blackwell. Newer generations reprice older ones.
Fast memory stacked on the GPU package. Its size and speed drive which AI models a GPU can run.
NVIDIA reference server designs. HGX is the board platform OEMs build on; DGX is NVIDIA's own complete system.
Large-scale computing for science, engineering and AI workloads.
The largest cloud providers, such as AWS, Microsoft Azure and Google Cloud.
Running a trained AI model to produce answers. Demand is steadier and longer-lived than training.
A unit of power. Data center capacity and GPU deployments are often sized in MW.
A cloud provider focused on GPU compute, as distinct from a general-purpose hyperscaler.
Hourly compute pricing with no commitment. Higher rate, less predictable revenue.
Total facility power divided by IT power. Closer to 1.0 means a more efficient data center.
The standard cabinet that holds servers, power and networking in a data center.
Compute sold under a fixed-term commitment, usually at a discount to on-demand. Lenders value it more.
The exact build of a server: GPU model and count, CPUs, memory, storage and networking.
Two GPU form factors. SXM sockets onto the board for higher performance; PCIe cards plug into standard slots.
How fast a system produces AI output. A common measure of inference performance.
A measure whose meaning depends on context. For hardware, it is how many of the GPUs are in use. For Earn, it is the share of capital deployed into loans.
Loan structures and financing products
How GPU purchases are paid for, and the key terms of each loan.
The share of an asset's value a lender will fund. A 70% advance rate on $10M of GPUs is a $7M loan.
The yearly interest cost of a loan to the borrower. Distinct from the yield an investor earns.
Lending sized and secured primarily by the value of specific assets rather than the borrower's overall credit.
A large final payment of remaining principal at the end of a loan that was not fully amortized.
Short-term financing that covers a gap, such as funding GPUs before long-term debt closes.
A fee paid on undrawn amounts of a facility in exchange for the lender keeping the money available.
A loan sized against the cash flows of a signed customer contract, such as an offtake agreement.
A loan the borrower can draw down in stages, often matched to GPU delivery schedules.
Taking funds from an approved facility.
A loan used to buy equipment, secured by that equipment.
The lessor owns the GPUs and the operator pays to use them. Can be an operating lease or a finance lease.
A fixed rate stays the same for the life of a loan. A floating rate resets against a benchmark such as SOFR.
An initial stretch where the borrower pays interest but no principal, often while a cluster ramps up.
Loan balance divided by collateral value. A lower LTV means more cushion for the lender.
A contract to purchase compute capacity under stated commercial terms.
A one-time fee charged when a loan closes, usually a percentage of the loan amount.
Paying a loan back early. Some loans charge a prepayment penalty or make-whole amount.
Loans made by non-bank lenders such as funds and insurers rather than through public bond markets.
Replacing an existing loan with a new one, usually to lower cost or extend term.
A credit line the borrower can draw, repay and draw again up to a limit.
An operator sells GPUs it already owns to a lender, then leases them back to keep using them.
Secured Overnight Financing Rate, the main US benchmark for floating-rate loans.
The margin over a benchmark rate that a lender charges, such as SOFR + 6%.
How long a loan runs. GPU loans are usually shorter than the useful life of the hardware.
A loan of a set amount repaid over a fixed schedule.
Credit provided by the equipment seller or its finance arm to help a buyer purchase hardware.
Collateral, valuation and depreciation
What the GPUs are worth today, what they will be worth later, and how a lender holds a claim on them.
Recording serial numbers and locations of each financed GPU or server so collateral can be verified.
Assets subject to a lender's security rights under the loan documents.
Collateral value divided by loan balance. The inverse of LTV.
Using the same collateral to secure more than one loan, or several assets to secure one loan.
How an asset's value falls over time. For GPUs it is shaped by new chip releases and rental prices.
How long a GPU can earn more than it costs to run. Often shorter than its physical life.
The price a willing buyer and seller would agree on in an open market.
A security interest ranked ahead of all other creditors' claims on the same asset.
What collateral would fetch in a quick, distressed sale. Lenders size loans conservatively against it.
A discount applied to collateral value when sizing a loan, to absorb price drops.
A legal claim on an asset that secures repayment of a debt.
A requirement to add collateral or repay principal when collateral value falls below a set level.
Revaluing collateral at current market prices rather than the original purchase price.
The risk that newer hardware makes financed GPUs less valuable or harder to rent.
What collateral would fetch in a planned sale over a reasonable period. Higher than FLV.
The legal steps, such as a UCC filing, that make a lender's lien enforceable against third parties.
How long hardware can keep running before it fails or is too costly to maintain.
Reselling or re-leasing repossessed equipment to recover value.
The expected value of the GPUs at the end of the loan or lease term.
Where used GPUs and servers are bought and sold. Its depth sets recovery values.
A lender's legal right to take and sell collateral if the borrower defaults.
A public notice filed under US commercial law that records a lender's lien on equipment.
Credit underwriting and risk
How a lender decides whether to lend, at what price, and what can go wrong.
Contracted revenue not yet earned. A strong backlog supports a borrower's ability to repay.
Financial strategy where a chip manufacturer (such as NVIDIA) guarantees debt, lease payments, or residual values for neoclouds and AI developers so they can secure cheap funding to purchase GPUs.
The risk that the other party to a contract, such as the compute customer, fails to pay.
A promise in the loan documents, such as keeping a minimum cash balance or a maximum leverage level.
How much of a borrower's revenue comes from a few customers. High concentration raises risk.
Cash flow available for debt payments divided by the payments due. Above 1.0x means cash covers debt.
Probability of default × loss given default × exposure. The average loss a lender should plan for.
A rating of BBB- / Baa3 or higher. Many institutional investors can only buy IG debt.
Total debt divided by EBITDA. Higher means more indebted relative to earnings.
The share of a loan a lender loses after recoveries if the borrower defaults.
A rating below BBB-, also called high yield or sub-investment grade.
The likelihood a borrower fails to repay over a set period.
The share of a defaulted loan a lender gets back, mainly by selling collateral. Equals 1 minus LGD.
Yield after accounting for expected losses and the volatility of returns.
A contract clause requiring the customer to pay for committed capacity whether or not they use it.
The average time until each dollar of principal is repaid. Used to compare loans and bonds.
Securitization, ABS and capital markets
How individual GPU loans become pooled, rated securities that institutional investors can buy. This is the language of the GPU ABS thesis.
A bond paid from the cash flows of a pool of loans or leases, such as auto loans or GPU loans.
The set of loans backing a securitization.
A cap on how much of a pool can come from one borrower, customer, chip type or location.
How likely loans are to default together. Lower correlation makes pooling more effective.
Protection added so senior investors are paid first, such as subordination, reserves or overcollateralization.
The bottom slice of a securitization. It absorbs losses first and earns the highest return.
Interest collected from the pool minus interest paid to bondholders and fees. A first buffer against losses.
A commitment by an investor to buy newly originated loans on agreed terms over a period.
A middle slice of a securitization, junior to senior notes and senior to equity.
Holding more collateral in a pool than the value of bonds issued against it.
Combining many loans into one portfolio so the group's risk is lower than any single loan's.
Cash set aside in a securitization to cover shortfalls in payments to investors.
The top slice of a securitization. Paid first and rated highest, often investment grade.
Ranking some investors behind others so junior tranches absorb losses before senior ones.
A legal transfer of loans to the SPV that removes them from the seller's estate in bankruptcy.
The order in which cash from the pool pays fees, then senior, mezzanine and equity investors.
A short-term credit line that funds loans until there are enough to securitize.
Onchain credit and USD.AI terms
The digital-asset side of GPU financing, plus USD.AI's own product names.
The yearly return to a holder, including compounding. On USD.AI, staking APY is distinct from loan APR.
Compute Hardware Interest Protocol: the pricing engine for AI debt. Holders vote on underwriting and GLX parameters, such as qualifying hardware and eligibility standards. The CHIP Foundation executes approved decisions.
Support toward a 7.00% net APY for sUSDai that the CHIP Foundation intends to provide, subject to available capital, funded over time from GLX gain-on-sale proceeds. It is not a guaranteed minimum, and net APY can fall below it.
USD.AI's agentic underwriting system. It scores each GPU financing against the standards CHIP holders set, and the score sets the loan's terms. A passing score is not a financing commitment until remaining conditions are verified and approval is obtained.
Financial services run by smart contracts on public blockchains rather than by intermediaries.
The USD.AI pool where depositors' capital is deployed into GPU loans to earn yield.
USD.AI's marketplace connecting compute loans with institutional capital markets.
Lending secured by productive hardware, such as GPUs, rather than by financial assets.
Software that never takes control of users' assets. Users interact directly with smart contracts.
Recorded and executed on a public blockchain, so transactions and rules are visible to anyone.
USD.AI's published, verified reports on the loans behind the protocol.
A physical or traditional financial asset, such as a GPU loan, represented or financed onchain.
Code on a blockchain that executes automatically when set conditions are met.
A digital token designed to hold a steady value, usually one US dollar.
Locking tokens in a protocol to earn yield or take part in governance.
Staked USDai. Holders earn yield from USD.AI's GPU loan book.
Representing an asset or a claim on cash flows as a digital token on a blockchain.
USD.AI's dollar stablecoin, positioned for AI payments.
The first-in, first-out order in which withdrawal requests are processed each 30-day cycle. Requests are paid from available cash. When cash is short, unfilled requests carry into later cycles and payment can take longer.
Reserved cash, funded over time from GLX gain-on-sale proceeds, used to top up net APY when it falls below the CHIP Hurdle Rate, subject to available capital. If the reserve lacks capital, net APY can fall below the CHR.
Legal, servicing, default and recovery
What happens after closing: collecting payments, monitoring the loan, and getting money back if things go wrong.
The lender's right to demand the full loan balance immediately after a default.
Giving the lender rights to the borrower's customer contracts so it can collect payments directly.
A standby firm ready to take over loan servicing if the primary servicer fails.
A requirement to send excess cash to pay down the loan, often triggered by a covenant breach.
A bank account, often controlled by the lender, where customer payments are deposited.
Time a borrower has to fix a missed payment or covenant breach before it becomes a default.
Failing to meet a loan obligation, such as missing a payment or breaching a covenant.
The final signed loan agreements. Terms are not binding until these are executed.
A payment that is past due but not yet in default. Usually tracked at 30, 60 and 90 days.
A specific trigger listed in the loan documents that allows the lender to act.
Taking control of collateral after default so it can be sold or re-leased.
A contract between lenders setting who gets paid first and who controls enforcement.
An agreement from the data center owner allowing the lender to access and remove financed equipment.
Collecting payments, monitoring covenants and reporting to lenders over the life of a loan.
An account that receives customer payments directly, so cash reaches the lender before the borrower.
A loan where the lender can only recover from the collateral, not the borrower's other assets.
A loan where the lender can pursue the borrower's other assets if collateral falls short.
Changing a loan's terms, such as rate or maturity, to avoid or resolve a default.
The lender's right to take over operating a cluster or customer contract to protect its collateral.
Negotiating with a struggling borrower to recover as much value as possible outside of court.