Why speculators never touch a Rillor physical contract.

If you arrive at Rillor wanting to be long or short the price of a GB300 rack with no intention of ever taking a pallet of liquid-cooled hardware off a truck, there is nothing here for you. That is not an oversight, and it is not a feature we have yet to ship. It is the design. Every contract on the Rillor book is a bilateral OTC forward with the intent of physical delivery, and none of them is ever cash-settled. A participant who only wants a price view, who wants to express a thesis on where B200 systems trade in September without ever owning one, has no instrument on this venue. That person belongs somewhere else, and we have built the rest of the market so that the somewhere else exists.

The result is a clean physical book whose price signal comes from buyers and sellers who actually deliver, and a separate, legitimate speculation surface (the Rillor Compute Index) that licensed third-party venues use to build cash-settled products Rillor does not operate. This piece explains the wall between the two, why it sits where it does in regulation, and why funds and exchanges should read the wall as the most valuable thing we offer, not a limitation.

The hard rule, stated plainly

A Rillor forward is an agreement between a verified buyer and a verified seller to deliver a specific, complete OEM GPU system at a specified future date and price. The underlying is not an abstraction. It is a Supermicro SYS-A22GA-NBRT, a Gigabyte G894-AD1-AAX5, a Dell PowerEdge XE9780, an HPE ProLiant Compute XD685, or a full GB200 NVL72 rack, with the CPUs, NICs, switches, and NVMe spelled out in the contract. At execution the buyer posts a 10% deposit into independent escrow. The balance is due at delivery. There is no line in that document that says the position can be closed out for the difference between the strike and a reference price. There is no cash settlement option because cash settlement is not a thing Rillor does, on any contract, ever.

This single rule does the work of an entire compliance regime. If the only way to realize a position is to deliver hardware or take delivery of hardware, then a participant who cannot or will not do either has no economic reason to be on the book. The speculator who wants exposure without logistics is filtered out not by a rule that says "no speculators allowed" but by the shape of the instrument itself. The instrument is useless to them.

Why the structure is the regulatory line, not a marketing posture

This is where the design and the law converge, and it is worth being precise because the precision is the point. Under the Commodity Exchange Act, the forward-contract exclusion covers, in the CFTC's own framing, any sale of a nonfinancial commodity for deferred shipment or delivery so long as the transaction is intended to be physically settled. Intent to deliver is the essential element. The CFTC's stated test is that the primary purpose of the contract must be to transfer ownership of the commodity, not solely to transfer its price risk, and that intent is assessed under a facts-and-circumstances test rather than a checkbox.

A complete GPU system is a nonfinancial commodity. A standardized agreement to deliver one at a future date, between a commercial buyer and a commercial seller, with deposit, escrow, and a performance bond all oriented around the act of delivery, is a forward contract in the most textbook sense. It is not a swap and it is not an exchange-listed future.

The distinction from futures is not cosmetic. Per the CFTC, futures are standardized, exchange-traded, clearinghouse-guaranteed instruments that can be offset by taking an opposite position. Forwards cannot be offset. If a party on Rillor buys a forward and then sells an identical one, they are not flat. They are holding two binding obligations, one long and one short, each of which must be performed. That is the mechanical reason a Rillor book cannot become a speculation venue by accident. There is no netting engine that lets a trader paper over a delivery obligation with an opposite ticket. The obligations stack. We have written more on this exact boundary in forward contracts versus futures for GPU systems, and on why we settle physically and never cash-settle in our note on physical delivery.

There is one nuance worth naming for the market-structure reader. Rillor supports pre-delivery transfer: a buyer can assign a forward to another KYC'd buyer before delivery, with Rillor and OEM approval and the end-customer-of-record captured for NVIDIA channel compliance. That is not offsetting and it is not a synthetic exit. It is a novation to another party who themselves intends to take delivery. The hardware still moves. The chain of intent-to-deliver stays intact end to end. We walk through the mechanics in pre-delivery transfer, step by step.

Why a clean physical book is worth defending

The commercial argument for the wall is the one a CME hedger would recognize immediately. On established commodity markets, cash-settled contracts are speculator-friendly because a position can be carried through expiration without ever touching delivery logistics, storage, or unexpected ownership. Physically delivered contracts pull in the opposite kind of participant: the producer or consumer who actually makes or takes the commodity and wants hedging to line up with a real physical business. Those two populations price risk differently, and mixing them changes what the printed price means.

Rillor wants the physical population, undiluted, because their prints are the raw material for everything downstream. When a tier-2 cloud commits to forty GB300 racks for Q4 delivery, that is a statement about real capacity planning backed by a deposit and a delivery obligation. When a seller writes that forward and posts a performance bond against it, that is a statement about real allocation they can actually source. The spread between those two is information. It is the obsolescence delta, the lead-time premium, the supply-tightness signal, all priced by parties with skin in the physical outcome. The moment you let in a population that prices only the abstraction, the signal degrades, because a cash-only player has no delivery constraint disciplining their bid. We let the forward curve form from real contracts, a process described in how a forward curve forms from real contracts, and that curve is only as honest as the book underneath it.

Where the speculation actually goes

None of this means speculation is bad or unwelcome. Directional risk transfer is healthy and useful. It is simply someone else's job, on someone else's venue, against a price Rillor publishes. The mechanism that makes that possible is the Rillor Compute Index: 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 API to exchanges, funds, and researchers.

This is a well-worn pattern in mature markets, not a novelty. CME runs cash-settled index futures that financially settle to the value of a licensed index. A venue licenses an index as a settlement feed and builds a cash-settled product on top, rather than operating the physical market the index measures. Equity index futures do not deliver a basket of stocks. They settle to a number. The number is the licensed index value. Rillor occupies the role of the index owner in that arrangement. A third-party exchange or DEX can license the Rillor Compute Index and stand up a cash-settled perpetual or future on RIL-GX-B300-2T or RIL-NVL72-GB200, and a speculator can be long or short that product all day. Rillor does not operate the venue, does not take the other side of those bets, and does not cash-settle anything on its own book as a consequence. Building such a product is a matter for the venue, not for Rillor: it licenses the feed and stands up the contract on its own listing infrastructure.

The precedent for a per-SKU GPU price benchmark already exists in the wild. Silicon Data's Silicon H100 index (Bloomberg ticker SDH100RT) is recalculated every business day and covers roughly 80% or more of the available global H100 rental market, including the bulk of neo-cloud providers and the major hyperscalers. That validates the demand for a transparent per-SKU compute price. The Rillor Compute Index differs in a way that matters for settlement integrity: it is computed from binding forward contracts with real deposits and delivery obligations behind them, not from rental quotes. The print is anchored to commitments, not indications. That is also why we treat the index, not the marketplace, as the durable asset, an argument made in full in the index is the moat.

Here is the separation laid out side by side.

Verified counterparties are a consequence of delivery, not a gate for its own sake

People sometimes read Rillor's KYC, independent escrow, and seller performance bonds as friction, the kind of onboarding a permissionless venue would skip. The framing is backwards. Those controls exist because every counterparty must actually perform a physical act. A buyer must be capable of taking delivery and of being the end-customer-of-record under NVIDIA channel compliance. A seller must be capable of sourcing and delivering the exact system on the contract, which is why the performance bond exists and why we explain precisely what it covers in what a seller performance bond covers.

In a cash-settled world none of that is strictly necessary. You can let an anonymous wallet take a long perp because the only thing at stake is margin, and the venue can liquidate the position if margin runs out. Delivery changes the stakes. A defaulting seller does not just owe a number, they fail to deliver a several-million-dollar rack a buyer was counting on for a buildout, and a defaulting buyer leaves a seller holding hardware they sourced on the strength of a commitment. The deposit, the escrow, and the bond are the machinery that makes a delivery promise credible. They are the price of admission to a market where the promise is physical, and they are exactly why the same controls would be theater on a purely financial venue. The same logic governs how a forward holds up when a counterparty fails to perform, with non-performance falling back on the bond and the escrowed deposit rather than on a cash-settlement option that does not exist.

Why the wall is a feature for funds and exchanges

A fund manager hearing "you cannot speculate on the Rillor book" might initially read a closed door. The opposite is true, and the reason is the same reason CME's index complex is valuable rather than limiting. The fund does not actually want to enter into delivery obligations. It wants clean directional exposure to compute prices, settled in cash, with a reference it can trust. Rillor's structure hands it exactly that, by keeping the reference uncontaminated.

If Rillor allowed speculators onto the physical book, the index computed from that book would be measuring a blend of real delivery intent and pure financial positioning, and its value as a settlement feed would erode. By walling speculation off, Rillor guarantees that the index reflects committed forward transactions and nothing else. A fund building a position against a licensed cash-settled product is therefore settling against a price formed by participants who had to mean it. That is a better reference than one polluted by players who never intended to deliver. The cleaner the physical book, the more bankable the index, the more product a venue can responsibly list on top. Everyone in the stack benefits from the discipline, which is the whole argument for how funds hedge AI compute exposure without owning a single GPU.

For exchanges and venue operators, the proposition is even simpler. You do not need to build a physical settlement layer, a delivery dispute process, an escrow relationship with every OEM, or an NVIDIA channel-compliance workflow to list a compute product. You license the index and inherit a settlement reference that already carries that integrity. Rillor handles the hard, unglamorous physical plumbing on its side of the wall, and you build the financial product on yours. What licensing entails is laid out in what it takes to license the index and list a compute product, and the broader index design lives across the /for-markets materials.

The two markets, one wall

Strip everything else away and the architecture is two markets joined by a single price and separated by a single rule. On one side is a physical forward marketplace where verified buyers and sellers transact standardized contracts on complete OEM GPU systems, with physical delivery on every one, never cash-settled, every counterparty bound to perform. You can see the live tape at /marketplace and the full SKU set at /skus. On the other side are independent, licensed venues running cash-settled derivatives against the Rillor Compute Index, where speculators express whatever view they like with no claim on a single physical rack.

The wall between them is not a compromise. It is the source of the value on both sides. The physical book stays clean because the instrument has no use for a non-delivering player. The index stays bankable because it is computed from a clean book. Speculators get a legitimate, liquid surface that is purpose-built for them. And Rillor stays exactly what it is, a physical-delivery forward market and the owner of the reference price, never the operator of a casino on its own contracts. If you are evaluating compute as an asset class to build product around, the index is where to start.