NVIDIA channel compliance inside a forward contract.

NVIDIA does not sell accelerated systems into a vacuum. It allocates through authorized server makers, and it tracks where those systems land. That single fact governs almost everything about how high-end GPU hardware can legitimately change hands. A buyer who acquires an HGX B200 or a GB300 NVL72 rack outside that tracked channel does not just risk a procurement headache. They risk a void warranty, no vendor support, and damaged standing the next time they ask an OEM for allocation.

A forward market sits in tension with all of that by design. The whole point of a forward is that the contract is standardized and transferable. You can buy a position today and hand it to another buyer before delivery. Done carelessly, that transferability is exactly the mechanism that creates a gray channel. The work Rillor does, and the reason a single contract field carries so much weight, is to make a transferable forward stay inside the authorized channel from execution through delivery. This is a procurement explainer of how that works.

How NVIDIA's channel actually works

NVIDIA brings its data center products to market through a connected ecosystem of partners. The NVIDIA Partner Network is organized into Registered, Preferred, and Elite tiers across competencies including Compute, DGX AI Compute Systems, and Networking. Underneath that partner framing is a simpler operational reality. GPUs reach end customers through authorized OEMs, and the path each unit takes is visible to NVIDIA.

The HGX H200 launch made the authorized list explicit. NVIDIA named ASRock Rack, ASUS, Dell Technologies, Eviden, GIGABYTE, Hewlett Packard Enterprise, Ingrasys, Lenovo, QCT, Supermicro, Wistron, and Wiwynn as the partner server makers that could deploy the platform, with systems beginning to ship in the second quarter of 2024. The H200 itself was the first NVIDIA GPU with 141GB of HBM3e at 4.8 terabytes per second, a substantial jump in capacity and bandwidth over the prior generation. The point for procurement is not the spec sheet. It is that the GPUs only became available through that named set of builders. There was no fourteenth, unlisted route.

That structure has not loosened with Blackwell. A complete HGX B200 NVL8 system still arrives as a built configuration from one of those OEMs: a Supermicro SYS-A22GA-NBRT, a Gigabyte G894-AD1-AAX5, a Dell PowerEdge XE9680L, an HPE ProLiant Compute XD685, a Lenovo SR680a V3, an ASRock Rack 4U8X-EGS2/DLC. Each is the same eight-GPU baseboard wrapped in a different vendor's chassis, head node, cooling, and service contract. The GPUs inside are NVIDIA's. The system around them, and the channel relationship that delivered it, belongs to the OEM.

When a system is resold outside that chain, three things break at once. The OEM loses sight of where its hardware went. NVIDIA loses the end-customer signal it uses to plan and police allocation. And the buyer inherits a unit with no provenance, which is precisely the condition under which support and warranty disappear.

Why the channel of record is not optional

The warranty consequence is the sharpest one, because NVIDIA states it plainly. Under the NVIDIA limited warranty, coverage applies only to the original purchaser of a product acquired from an NVIDIA-authorized channel, and it does not extend to anyone who acquires the product on a used basis. The warranty is described as for the original owner only, non-transferable, non-assignable, and voided when the product is transferred to another party.

Read that against how a careless forward might work, and the problem is obvious. If a position were simply assigned from one buyer to another, with the hardware nominally changing owners along the way, the receiving party could find themselves holding exactly the used, transferred unit the warranty terms exclude. That is not a marginal risk on a system that can run into the high six figures. A single HGX B200 NVL8 node commonly carries a system price in the range of 350,000 to 500,000 dollars depending on OEM, CPU, NVMe, and cooling, and a GB300 NVL72 rack runs into the low millions. Buying that with no warranty and no support path is not a discount. It is an unhedged liability.

The allocation consequence is quieter but, for repeat buyers, larger. OEMs and NVIDIA both watch where systems end up. A buyer who shows up clean, with traceable destinations on every unit, builds standing that pays off in the next allocation round. A buyer associated with hardware that surfaced in a gray channel does not. For a tier-2 cloud or a fund-backed operator planning multiple rack deliveries, channel hygiene is not a compliance chore. It is the thing that keeps the supply line open.

The end-customer-of-record field

A Rillor forward is a bilateral OTC forward on a complete OEM system, written with the intent of physical delivery. It is standardized in structure but, like any forward, customizable in its terms between the parties. The CFTC defines a forward contract as an agreement between a commercial buyer and seller to deliver a specified quality and quantity of goods at a future date, with terms more personalized than standardized futures. CME Group draws the same line: forwards are customizable over-the-counter agreements where delivery and counterparty terms are set directly between the parties, distinct from exchange-traded futures. Rillor's instrument lives on the forward side of that line. It is a forward and not a future, and it always settles physically rather than in cash, which is why every contract resolves to delivered metal at a named destination.

Standardization is what makes a forward tradable. Six fields define every contract.

The last field is the channel-compliance keystone. The end-customer-of-record is the authorized destination, the entity that NVIDIA and the OEM will see as the customer for that system. Because it is captured at execution, the contract never exists in a state where its physical destination is unknown. A position is not an abstract claim on a GPU somewhere. It is a claim that resolves to a named, authorized buyer of record.

This is also why Rillor's SKU catalog ties each Rillor SKU back to concrete OEM configurations rather than to a bare GPU. You are not contracting for "a B200." You are contracting for a deliverable system from an authorized builder, destined for a named end customer, which is the only form in which the channel recognizes the transaction.

KYC is a precondition, not a postscript

Both sides of a Rillor forward are verified before they can execute. KYC is run on the buyer and on the seller, and execution does not happen until both pass. This ordering matters more than it might seem.

Verifying the buyer is intuitive. You want to know the end customer is a real, authorized entity that can legitimately take delivery and become the customer of record. The buyer onboarding path walks through what that actually involves. But verifying the seller is just as load-bearing for channel compliance. A seller offering forward inventory has to be an entity whose supply is itself authorized, whether that is an OEM listing its own forward allocation, a reseller with a legitimate channel relationship, or Rillor itself acting as a seller and underwriter to seed liquidity. If the seller's supply is gray, no amount of buyer verification cleans it. Channel integrity is only as strong as the weaker verified side.

Because both verifications precede execution, the contract is born compliant. There is no window in which an unverified party holds a live position. Compare that to a spot transaction stitched together after the fact, where provenance is reconstructed from invoices and hope. The forward structure front-loads the check, which is exactly where a channel control belongs.

Transfer without breaking the chain

Transferability is the feature that makes a forward useful and the feature that, handled wrong, would manufacture a gray channel. Rillor allows pre-delivery transfer of a position to another buyer, and the controls around that transfer are what keep it authorized. Here is the channel-compliance core.

A transfer is not a private assignment between two parties. It runs through three gates.

1. The incoming buyer is KYC'd. The new buyer is verified to the same standard as any buyer at original execution. An unverified party cannot receive a position.
2. The end-customer-of-record is re-captured. The contract's destination field is rewritten to the new authorized end customer. The system is never in a state where its tracked destination is stale or unknown.
3. The OEM re-approves. Both Rillor and the relevant OEM sign off on the new destination before the transfer completes. The OEM whose channel relationship delivers the hardware confirms the new end customer is acceptable to it and, through it, to NVIDIA.

Only after all three clear does the position move. The effect is that a transfer behaves, for channel purposes, like a fresh authorized sale rather than a used-goods handoff. The end customer of record at delivery is a KYC'd, OEM-approved entity, which is the condition the warranty and channel rules require. The transferable forward and the authorized channel are reconciled precisely because the destination is re-established, not inherited, at every hand-off.

Which OEMs map to the record

The end-customer-of-record only means something because there is an OEM behind it whose NVIDIA relationship the record attaches to. The authorized server makers Rillor sources from are the OEMs themselves: Supermicro, Gigabyte, Dell, HPE, Lenovo ISG, ASRock Rack, Aivres, ASUS, Pegatron, Inventec, Fujitsu, NationGate, Altos Computing, and Netweb, with NVIDIA as the channel and KYC authority and as the OEM of DGX systems.

Each OEM brings its own configurations under the same Rillor SKU. A RIL-GX-B200-2T contract can resolve to a Supermicro SYS-822GS-NBRT, a Gigabyte G893-SD1-AAX5, a Dell PowerEdge XE9685L, or an HPE ProLiant Compute XD685, with an Intel Xeon 6980P or AMD EPYC Turin head node, ConnectX-7 or ConnectX-8 fabric, BlueField-3 DPUs, and Micron 9550 PRO NVMe. The Rillor SKU standardizes the contract. The OEM mapping standardizes the deliverable. The record connects the named end customer to the specific OEM whose channel will fulfill it, which is what lets NVIDIA see the destination through its own partner.

This is also why an OEM has a clean reason to list forward inventory rather than treat the forward market as a threat. The record keeps every unit inside the channel the OEM already operates in, with the end customer captured and re-captured under OEM approval. We make that case in detail in why an OEM lists forward inventory. For OEMs, the forward is a way to sell future allocation into verified demand without ever losing the channel-of-record control they are required to maintain.

What the buyer actually keeps

The reason all of this matters to the person writing the check is concrete. Buying through a Rillor forward, with the end-customer-of-record captured at execution and re-captured on any transfer, means the buyer takes delivery as the original purchaser from an authorized channel. The warranty terms that void on used or transferred hardware do not bite, because the system is delivered to a named end customer of record, not assigned around as a used unit. The OEM support contract attaches to a recognized customer. And the buyer's allocation standing with both the OEM and NVIDIA stays intact, because their name is on the record where it belongs.

Set against the supporting structure of a Rillor forward (the independent escrow agent, the 10 percent deposit at execution with balance at delivery, and the seller performance bond), channel compliance is the piece that protects the asset's value after it arrives. Escrow protects the money in flight. The performance bond protects against seller default. The end-customer-of-record protects the warranty, support, and standing that make the delivered system worth owning. The pricing that results from all of this active contracting feeds the Rillor Compute Index, the owned forward-price feed Rillor licenses to exchanges, funds, and researchers, but the contract on the buyer's desk is and always will be a physical-delivery forward with a known, authorized destination.

If you are an OEM weighing whether listing forward inventory keeps you inside your NVIDIA channel obligations, it does, and the record is how.