Why Rillor settles physically and never cash-settles.

Every contract on Rillor ends the same way: a specific OEM GPU system arrives at a specific facility, gets racked, and gets powered on. There is no version of a Rillor forward that closes out as a cash payment in place of the hardware. We do not offer one, we will not build one, and the architecture of the platform is deliberately arranged so that the question never comes up. A buyer who signs a forward on a RIL-GX-B200-2T is committing to take eight Blackwell SXM6 GPUs on an HGX baseboard at a receiving dock, not to settle a price difference against a benchmark.

This is a position, not an accident, and it is worth defending in plain terms. A real procurement market has to terminate in real goods, because the goods are the entire point. The financial exposure that sits on top of a hardware market is legitimate and useful, but it is a second thing, owned by other people, one layer out. This piece is the argument for keeping those two things cleanly separated, and for why that separation is what makes the price Rillor publishes worth trusting.

What "physical delivery" actually means here

It is easy to say "physical delivery" and let it stay abstract, so anchor it to the hardware. A Rillor forward on a complete OEM GPU system is an obligation to deliver real iron through the OEM channel. When the underlying is an HGX B200 system, the deliverable is eight Blackwell SXM6 GPUs per baseboard, each carrying 180 GB of HBM3e, rated up to 1000 W total graphics power, with 900 GB/s of NVLink GPU-to-GPU bandwidth, integrated into an OEM server such as a Dell PowerEdge XE9680L, a Supermicro SYS-A22GA-NBRT, a Gigabyte G894-AD1-AAX5, or a Lenovo SR680a V3. At rack scale, the deliverable on a RIL-NVL72-GB200 is a liquid-cooled NVL72 connecting 36 Grace CPUs and 72 Blackwell GPUs, with 13.4 TB of HBM3e, 130 TB/s of NVLink bandwidth, and roughly 1.44 exaflops of compute. NVIDIA describes that configuration as an exascale computer in a single rack. That object weighs over a ton, draws power measured in tens of kilowatts, and needs a loading dock, a coolant loop, and an electrician.

You cannot settle that in cash and call it the same product. A wire transfer of the price difference between strike and benchmark leaves the buyer with no GPUs and the data center floor empty. For a buyer whose actual constraint is racked, networked compute by a delivery month, that outcome is not a hedge, it is a failure to procure. So the contract has to end in the rack, and the mechanics are built around that ending: a 10% deposit at execution, the balance at delivery, an independent escrow agent holding funds, a seller performance bond standing behind the obligation to ship, NVIDIA channel compliance with the end-customer-of-record captured on every contract. None of that machinery makes sense for a cash-settled instrument. All of it is required for a delivery you intend to take.

Why the legal structure depends on real delivery intent

The settlement choice is also what keeps a Rillor contract a forward rather than a futures contract, and that distinction is load-bearing.

The CFTC draws the line precisely. A forward contract, in the regulator's own glossary, is a commercial-merchandising transaction in which the parties intend that physical transfer of the actual commodity will occur. A futures contract, by contrast, is defined as an agreement that may be satisfied by delivery or offset, and that ability to close a position without ever touching the underlying is exactly the feature that makes a futures contract a futures contract. Rillor's contracts sit firmly on the forward side of that line. They are bilateral OTC forwards between verified counterparties with the intent of physical delivery, which is what brings them within the CFTC's forward-contract exclusion from the swap and futures definitions. They are not exchange-listed futures, and they are not designed to be offset into cash.

Intent is not a label you get to assert, it is a fact pattern the regulator examines. Under the Dodd-Frank forward exclusion, intent to make or take delivery is an essential element of a contract excluded from both the swap and the futures definitions, and it is evaluated under a facts-and-circumstances test of what the parties actually meant to do. Build in routine cash settlement and you have undercut the very element the exclusion rests on. The honest position and the defensible legal position turn out to be the same one: if you say you intend to deliver, the contracts have to actually deliver. Rillor's do.

The short version of where the forward sits relative to a futures contract is that the settlement mechanic is the whole difference. A forward intends delivery and is held to that intent; a futures position is built to be closed out for cash before it ever touches the commodity. Rillor is squarely the former, and the deposit, escrow, and performance-bond machinery exists to keep it that way.

Where cash settlement does belong

None of this is a claim that cash settlement is illegitimate. It is a claim about who builds it and where it lives.

Cash-settled exposure is genuinely useful. A fund that wants to be long or short the price of B300 capacity, without a loading dock or a power contract, has no use for physical delivery and every use for an instrument that pays in money. That is precisely what cash settlement is for. A cash-settled index product resolves as a debit or credit directly into the trading account, with nothing received or delivered on exercise or assignment. Exposure resolves in money, never in the underlying, and for a pure financial participant that is the correct design.

The Rillor view is simply that those products belong one layer out, and that someone other than Rillor should operate them. The mechanism that makes this clean is the Rillor Compute Index. The index is a 30-day rolling-blend forward price per SKU, computed from active Rillor contracts, owned and controlled by Rillor, and licensed as a settlement feed and an API to exchanges, funds, and researchers. Third-party venues take that feed and build cash-settled derivatives against it: perpetuals, futures, options. Rillor does not operate those venues, does not take the other side of those trades, and does not cash-settle its own contracts. The index is the moat, and the exposure layer is downstream of it.

This is the ordinary shape of the index business, not an unusual carve-out. Index providers create, maintain, calculate, and publish indices and license them to third parties who build products on them. In a typical license, the index provider's only relationship to the exchange is the licensing of the index and its trademarks, and the listed product is neither sponsored nor operated by the index provider. The index owner and the venue operating the cash-settled product are distinct parties. Rillor is the index owner. We are deliberately not the venue. If you want to understand why we treat the index, rather than the marketplace, as the durable asset, and how licensing the index to list a compute product works, the markets and index page makes that case directly.

The clean separation, stated plainly

It helps to see the two layers side by side, because the entire design is the refusal to blur them.

Read down the two columns and the discipline is obvious. Rillor lives entirely in the left column. The right column is real, it is valuable, and it is somebody else's column. We supply the price that the right column settles against, and we never step across the line to take the other side of it.

How physical delivery aligns incentives

Settlement choice is not only a legal or structural question, it is an incentive question, and physical delivery selects for the right participants on both sides.

On the buy side, a forward that can only be satisfied with hardware is uninteresting to anyone who does not want hardware. A buyer signing a RIL-GX-B300-2T or a RIL-MI355X-2T is, by the structure of the contract, someone who intends to rack and run the systems. That is the population Rillor wants in the book: tier-2 clouds planning capacity twelve months out, sovereign AI programs, enterprise buildouts. It is also why pure speculators never touch a Rillor physical contract. The deposit, the delivery obligation, and the channel-of-record requirement together make the contract useless to anyone who only wants price exposure, and that is the intended filter, not a side effect.

On the sell side, the same logic runs in reverse. A seller posting forward inventory is committing to ship, and that commitment is backed by a performance bond. The bond is not theater. It means a seller who lists a delivery month is putting capital behind the promise to make that month. Cash settlement would let a seller collect a spread on a price view without ever sourcing a single GPU. Physical settlement does not allow that. The seller either delivers the iron or the bond and contract terms make the buyer whole. That is the difference between a marketplace of people who move hardware and a casino of people who move opinions, and Rillor is built to be the former. Rillor itself sits on both sides of this discipline: it is the neutral operator, and it is also a seller and underwriter that seeds liquidity, which means it carries the same delivery obligations it asks of everyone else.

Why physical settlement is what makes the index credible

Here is the part that ties the manifesto together. Physical settlement is not only good for the contracts, it is the thing that makes the index worth licensing.

A price benchmark is only as credible as its relationship to the real market underneath it. In physically delivered markets, the underlying physical market is inherently tied to the contract through the delivery mechanism. In cash-settled markets, the contract is not physically tied to the commodity, and no one is ever at risk of having to make or take delivery. That tie is what disciplines price. When delivery is real, the contract price has to converge toward the price of the actual good, because someone is going to have to hand over the actual good.

When that discipline breaks, the consequences are well documented. Research on agricultural markets found that sustained non-convergence between cash and settled prices makes hedging less effective, sends confusing signals to the market, can threaten the viability of a contract, and ultimately leads to misallocation of resources. The lesson generalizes cleanly to compute. A forward curve for GPU systems is only useful if it tracks what buyers and sellers actually pay to move iron. The way you guarantee that is to make the contracts that feed the curve settle in iron.

So the architecture is self-reinforcing. Because every Rillor contract settles physically, the contract prices stay anchored to real procurement. Because the prices stay anchored to real procurement, the index computed from those contracts is credible. Because the index is credible, third parties are willing to license it and settle cash products against it. The moment Rillor allowed its own contracts to cash-settle, it would weaken the anchor that gives the index its value, and it would be sawing through the branch it is sitting on. We are not going to do that. For how the curve assembles from live contracts, see how a forward curve forms from real contracts.

The bottom line

Rillor is physical-delivery forward only. Every contract resolves in an OEM GPU system arriving at a facility, the legal structure depends on that delivery intent being real, and we never cash-settle. Cash-settled exposure is a legitimate product, but it is built by third-party venues against the Rillor Compute Index, which Rillor owns, controls, and licenses without operating the venues that trade on it. The two layers are separate on purpose, and the separation is exactly what keeps the price honest.

If you are sizing forward capacity and want to understand what that actually buys you, the buyer side is the place to start.