When an OEM or reseller writes a forward on Rillor, the price the buyer sees is only half of what they are evaluating. The other half is whether the systems actually arrive in the delivery month, in the contracted configuration, at the contracted facility. The seller performance bond is the instrument that turns that second half from a hope into a number. It is the seller's own capital, posted at execution, that pays out if the seller fails to deliver. It is the reason a buyer can treat your quote as a commitment rather than an option you may decide to honor.
This is an OEM-facing piece, so we will be concrete. The bond is not a fee, not a penalty, and not an insurance product you buy from a third party in most cases. It is collateral against your own delivery obligation on a specific SKU and a specific delivery month. What follows is what the bond covers, what it deliberately does not, how its size is set, and why posting one lets you quote a forward price an unbonded seller cannot match.
What a performance bond is
In commodity and derivatives markets, a performance bond is a good-faith deposit posted to guarantee a participant's performance against potential future losses. CME Group uses exactly this language: the term performance bond is interchangeable with margin, and it exists to make sure a counterparty can stand behind an open position. There is an initial amount posted when the position opens and a maintenance level that has to be kept in place, with a call issued if equity falls below it. The instrument is old, well understood, and boring in the way infrastructure should be.
The construction industry runs the same idea with a slightly different shape. There, a performance bond is a three-party instrument among the obligee (the party that requires the bond, usually the buyer or project owner), the principal (the party obligated to perform), and the surety (the entity guaranteeing performance). The AIA A312 form is the industry-standard document that defines how a claim is filed and how the obligee is made whole. The point of citing both is that the structure Rillor uses is not novel. It is the established mechanism for the same problem: one party has to perform in the future, and the other party wants assurance that capital is at risk if they do not.
A Rillor seller performance bond is the compute-hardware version of that mechanism. You, the seller, are the principal. The buyer is the obligee. Rillor administers the bond and the independent escrow agent holds the collateral. The thing being guaranteed is delivery of a complete OEM GPU system, on time, to spec.
What the bond covers
The bond covers the failures that are inside your control as the seller. There are three, and each has a defined remedy. We separate them on purpose, because lumping them together is how buyers and sellers end up arguing in the worst possible moment.
Late delivery
The contract names a delivery month. If you miss it, the late-delivery remedy runs. This is not an automatic forfeiture of the full bond. It is a graduated draw that compensates the buyer for the cost of the delay, typically calibrated to the spread between the contracted forward price and the buyer's cost to cover the gap (short-term rental, spot purchase, or rescheduled deployment) over the period of lateness. The bond is the source of those funds. The intent is to make the buyer economically whole for the delay, not to punish a seller who delivers a week late on a clean contract.
Partial delivery
You contracted to deliver a quantity of a SKU. You deliver some of them. The partial-delivery remedy treats the shortfall as its own mini non-delivery: the buyer is made whole on the units that did not arrive, at the contracted price, and the bond covers the difference between that price and the cost to re-source the missing units. The units that did arrive settle normally. This matters for rack-scale contracts where a single rack short on a multi-rack GB200 NVL72 order can hold up a whole training cluster.
Outright non-delivery
The hardest case, and the one the bond exists for. You do not deliver at all. The non-delivery remedy makes the buyer whole at the contracted price and funds Rillor's effort to re-source the systems from another seller on the book. If the re-source price is higher than the contracted price, the bond covers the difference up to its face value. The buyer's deposit is returned. The end-customer-of-record captured at execution stays attached to whichever seller ultimately delivers, so NVIDIA channel compliance is preserved through the substitution.
The common thread across all three is that the bond is a delivery guarantee, not a price guarantee in the abstract. It guarantees that the specific contracted system, for example a Supermicro SYS-A22GA-NBRT or a Dell PowerEdge XE9680L carrying NVIDIA HGX B200 boards at 180 GB of HBM3e and roughly 8 TB/s of memory bandwidth per GPU, shows up. If you want to see how the underlying contract structures the SKU, delivery month, and substitution mechanics that the bond sits on top of, the anatomy of a Rillor forward contract walks every field, and how a forward curve forms from real contracts covers how those terms turn into a priced curve.
What the bond does not cover
A bond that covers everything covers nothing, because its price becomes unmanageable and its meaning becomes muddy. Two categories are deliberately excluded, and both are handled by separate, named clauses.
The first is force majeure. If a fab loses a quarter of output to a natural disaster, a port closes, or an export-control change makes a contracted shipment unlawful, that is not a seller performance failure. The force-majeure clause governs it: the contract is suspended or unwound under defined terms, deposits are returned, and the bond is not drawn. This keeps the bond a clean signal. A drawn bond means a seller-controllable failure, full stop, which is exactly the information a buyer and the index need.
The second is buyer-side default. If the buyer cannot take delivery, will not pay the balance, or fails KYC re-verification before delivery, that is not your problem to cover, and the bond does not pay for it. The buyer's 10% deposit and the buyer-default clause handle that side. We treat the two sides symmetrically. The governing principle: each party's collateral backs only that party's own obligations.
If the failure is something the seller controls (lateness, shortfall, non-delivery), the bond pays. If it is force majeure or a buyer failure, the bond does not, because a different clause already owns that risk. Keeping the bond narrow is what keeps it credible.
How bond size is set
The bond is sized off two inputs: the contract notional and your track record as a seller.
Notional is the obvious lever. A bond is expressed as a percentage of the total contract value, so a larger or higher-value order carries a larger bond in absolute dollars. The percentage is not arbitrary. CME's margin model is designed to cover at least 99% of anticipated price changes over a liquidation period, and the same logic applies here: the bond should be large enough that, in the realistic worst case, it covers the cost of re-sourcing the systems if you fail. For compute hardware, the relevant volatility is real. Blackwell-class systems are still early in their production ramp, allocation is tight, and the cost to cover a missed delivery is set by the market on the day of the failure, not the day of execution. A bond has to be sized against that uncertainty rather than against a stale snapshot, because a seller who fails in a tight month forces the buyer to re-source into the most expensive part of the cycle.
Track record is the second lever, and it is the one you control over time. A seller writing a first contract on Rillor posts a higher bond percentage, because there is no delivery history to price against. A seller with a clean record of on-time, in-spec deliveries posts less, because the empirical probability of a draw is lower. This is the same maintenance-margin logic CME uses, applied to reputation: demonstrated performance lowers the collateral the market requires from you.
A note on form. The bond does not have to be posted as cash in every case. CME permits performance-bond requirements to be met with cash or non-cash collateral, with only the daily mark-to-market settled in cash. Rillor follows the same principle: the bond can be cash held by the escrow agent or an acceptable non-cash equivalent, depending on the seller and the contract. What does not change is that the collateral is real, segregated, and callable.
Why bonded supply quotes tighter
This is the part that matters for your margin, so it is worth being precise about the mechanism. A forward price has two components in the buyer's head: the price of the hardware, and the price of the risk that the hardware does not arrive. When you do not post a bond, the buyer prices in that second component themselves. They widen the price they are willing to pay, or they discount your quote, to compensate for the default risk they are now carrying. You eat that spread whether you see it itemized or not.
When you post a bond, you move the default risk off the buyer's books and onto your own collateral. The buyer no longer has to price it, so the spread they were holding back collapses into your quote. A bonded seller can therefore quote closer to fair value and still win the contract, because the buyer is comparing a bonded price against an unbonded price that already had a risk premium baked in. The bond is not a cost center. It is the thing that lets you compete at a tighter price than a seller who refuses to post one.
This is also why standardized, bonded supply is the precondition for a real reference price at all. GPU systems have lacked a standardized, deliverable reference price for valuation, hedging, and long-term planning, which is a gap every mature commodity closed long ago with standardized contracts and posted collateral. Bonded contracts are how that infrastructure gets built: they make each contracted price a credible, deliverable price rather than a quote with an asterisk. Those prices feed the Rillor Compute Index, the 30-day rolling-blend forward price computed from active Rillor contracts. A bonded contract is a clean data point. An unbonded one is noise.
| Scenario | Who carries delivery risk | Effect on forward price |
|---|---|---|
| Unbonded seller | Buyer prices in default risk | Wider spread, discounted quote |
| Bonded seller | Seller's collateral absorbs it | Tighter spread, competitive quote |
| Force majeure event | Force-majeure clause, no draw | Contract unwound, deposits returned |
| Buyer default | Buyer deposit and buyer clause | Bond untouched, seller protected |
How the bond interacts with escrow
The bond does not work alone. At execution, two pools of capital are committed at once. The buyer posts a 10% deposit. You post the performance bond. Both are held by the independent escrow agent, not by Rillor and not by the counterparty. That is the two-sided structure: the buyer has skin in the game against walking away, and you have skin in the game against not delivering. Neither side is asked to trust the other; both are asked to trust the escrow agent and the contract, which is a much easier thing to do.
The CFTC's definition of a forward contract is the regulatory backbone here. A forward is an agreement between a commercial buyer and seller to deliver a specified quality and quantity of goods at a specified future date, with delivery made to a named counterparty. In a futures clearing house, the clearing house picks the delivery counterparty. In a forward, the contract names them. That distinction is why Rillor is a physical-delivery forward and never a cash-settled instrument, and the bond fits that frame exactly: it guarantees physical performance to a named buyer, not a cash payment in lieu of delivery. The index Rillor publishes is licensed to outside venues that may build cash-settled products against it, but Rillor's own contracts always settle in hardware, and the bond is what makes that promise enforceable.
How a claim is adjudicated and paid
A claim runs the same way a construction-bond claim does under the A312 form, adapted to hardware delivery. The buyer (the obligee) declares the delivery failure to Rillor. Rillor verifies the failure against the contract: was the delivery month missed, was the quantity short, did nothing arrive. This is a factual determination against named fields, not a negotiation, which is the whole reason the contract is standardized.
Once the failure is confirmed and it is established that no excluded cause (force majeure, buyer default) applies, the remedy attached to that failure type runs. For late or partial delivery, the buyer is compensated from the bond for the measurable cost of the gap. For non-delivery, Rillor moves to re-source the systems from another seller on the verified book, and the bond covers any cost-to-cover above the contracted price up to its face value. The buyer's deposit is returned. Through all of this, the end-customer-of-record stays attached, so the substitute delivery remains NVIDIA channel compliant.
The buyer ends one of two ways: made whole in cash for a delay or shortfall, or re-sourced into the systems they contracted for, with the price difference absorbed by the bond. That is the entire promise. The buyer either gets the hardware or gets the money to go buy it elsewhere at the price you locked, and your collateral is what funds the difference. If you want to weigh bonded forward supply against listing inventory through other channels, why an OEM should list forward inventory on a forward market frames the broader trade, and the marketplace shows the live contract surface.
What this means for an OEM writing supply
Posting a bond is the difference between offering an option and making a commitment, and buyers pay for commitments. The bond covers exactly the failures you control, leaves force majeure and buyer default to their own clauses, and sizes itself down as you build a delivery record. In exchange, you quote tighter than any unbonded competitor because the buyer stops carrying default risk, and your clean deliveries feed an index that prices your future supply more accurately. The bond is not friction on the deal. It is the structure that makes your price believable.
Put your systems on the forward curve.
List forward inventory to a book of KYC'd, delivery-capable buyers. Standard channel pricing, forward demand visibility, end-customer of record on every contract.
List as an OEM →