Forward contract versus spot versus OEM allocation, compared.

If you procure GPU systems, you have historically had two ways to source them. You buy spot, taking whatever is available now at whatever it costs now, or you join an OEM allocation queue and wait for a slot that may or may not arrive in the quarter you planned around. Both are real. Both have a place. But for the most common situation in this market, a buyer who knows roughly how many systems they need and roughly when they need them, neither one is actually a good fit. They were the only options, which is a different thing.

The forward contract is the third option, and it is the one that fits that situation. A Rillor forward binds a specific OEM system SKU to a delivery month at a price fixed today, with a 10 percent deposit held by an independent escrow agent and a seller performance bond standing behind delivery. This piece compares all three across the dimensions procurement actually cares about: price certainty, counterparty risk, flexibility, and total cost. The conclusion is not subtle. If you have a known delivery date, the forward dominates.

What each option actually is

The three options are not three flavors of the same thing. They are structurally different transactions with different risk profiles, and the regulator's own vocabulary makes the distinction clean.

Spot is a purchase for immediate delivery at the prevailing price. The CFTC defines the spot price as the price in the marketplace for actual cash commodities to be delivered through customary market channels. You pay today's number for today's hardware. The number is whatever the market will bear at that moment, and you have no claim on what it will be next quarter.

OEM allocation is a queue position. When you request allocation for, say, a block of Supermicro SYS-A22GA-NBRT or Dell PowerEdge XE9680L systems, the OEM acknowledges your demand and places you in line for capacity that is itself constrained upstream by TSMC CoWoS packaging and HBM supply. It is a commercial expectation, not a binding instrument. There is no fixed price attached, no deposit in escrow, and no contractual delivery date you can hold anyone to.

A forward contract is a binding agreement to take delivery of a specified quality and quantity of goods at a specified future date. That is, almost verbatim, the CFTC definition of a forward, and it is precisely what a Rillor contract is: a bilateral OTC forward with the intent of physical delivery, which is why it falls under the CFTC forward-contract exclusion rather than being an exchange-listed future. The SKU is fixed, the delivery month is fixed, the price is fixed, and the obligation runs both ways.

One important contrast with futures. Almost all futures contracts end without physical delivery of the underlying. A forward more frequently results in actual delivery, and a Rillor forward results in physical delivery every time. You are not buying a paper position you will later close out. You are buying the machines. If you want the deeper version of that distinction, we wrote it up in anatomy of a Rillor forward contract.

Price certainty: the dimension that decides most cases

Spot prices for AI compute move, and they move a lot. H100 rental ran 7 to 10 dollars per GPU-hour at launch, with a median around 7.76 dollars in mid-2023, and fell to roughly 1.95 dollars median by late 2025. That is a more-than-threefold compression over the life of one generation. The direction happened to be down for the renter, but the point is the magnitude of the move, not its sign. A buyer committing capital across multiple tranches on spot has no idea what tranche two through six will cost.

Allocation does not solve this. A queue position tells you nothing about the price you will pay when your slot comes up, because OEM list and channel pricing can shift between request and fulfillment, and your delivery date is itself uncertain. You are exposed on both axes at once: when, and how much.

A forward locks one number today for a delivery month you choose. For a buyer planning a Blackwell buildout, that certainty is the whole game. A single GB200 NVL72 rack is 72 Blackwell GPUs, 36 Grace CPUs, 13.4 TB of HBM3e, 1,440 PFLOPS of NVFP4, and 130 TB/s of NVLink bandwidth delivering roughly 30x the LLM inference throughput of an H100. That is a multi-million-dollar line item per unit. When the unit price is that large and your delivery date is fixed by a model-training schedule or a data center power-on date, paying a known price beats paying an unknown one, even if the unknown one might turn out lower.

Counterparty risk: the part allocation pretends does not exist

Here is the dimension where the three options separate hardest, and where allocation is weakest despite feeling safe because it comes from the OEM directly.

A bare forward contract, the kind two commercial parties sign on their own, is non-standardized, not exchange-traded, and carries significant counterparty default risk. There is no clearinghouse standing behind it. If your counterparty fails to perform, you have a lawsuit, not a remedy. This is the structural weakness of forwards in general, and it is exactly the gap Rillor is built to close.

Rillor closes it with two mechanisms borrowed from how serious commodity and contract markets handle default. First, an independent escrow agent holds the 10 percent deposit until contract conditions are met, so neither side is trusting the other with cash. Second, the seller posts a performance bond, a surety instrument that compensates the buyer if the seller fails to deliver. Together these give a buyer the protection of a cleared market without requiring an exchange. We walk the escrow and bond mechanics in detail in anatomy of a Rillor forward contract.

Compare the three head to head:

The line that surprises procurement leads is the allocation column under default protection. A queue position can be re-routed when supply tightens, including toward a larger or earlier-committing customer, and you have no contractual recourse because you never had a contract. The 2026 supply picture, with hyperscaler reservations pre-committing large blocks of capacity, makes that re-routing risk concrete rather than theoretical. Allocation feels safe because it comes from the manufacturer. It is the least protected of the three.

Flexibility: a forward is an asset, an allocation is a dead end

Plans change. A model gets delayed, a region's power comes online late, a budget gets re-prioritized. What happens to your sourcing commitment when it does.

On spot, there is no commitment to manage, which is the one genuine advantage of spot. On allocation, a slot is typically non-transferable, so if your need evaporates you simply forfeit your place in line and whatever soft commitment you made. There is no secondary value.

A Rillor forward is transferable. Before delivery, you can transfer the contract to another KYC'd buyer with Rillor and OEM approval, with the NVIDIA channel compliance and end-customer-of-record handling done properly so the OEM's business unit stays whole. That turns a commitment into an asset with optionality. If your need changes, the forward is something you can move rather than something you have to eat. This is also the foundation for financing GPU capex against the contract itself, because a transferable, escrow-backed instrument is something a lender can actually underwrite.

Cost, honestly

A forward is not free, and we will not pretend it is. Rillor's take rate is 2 percent all-in, split 1 percent buyer and 1 percent seller, plus roughly 1,000 dollars of escrow cost per contract. On a contract worth several million dollars, the escrow line is a rounding error and the 2 percent is a known, disclosed number you can put in a model.

Now price the alternative honestly. The cost that matters is not the spread on the forward, it is the cost of getting the timing wrong without one. If you are building on spot and your second tranche has to be bought into a tight market, you pay the front-month premium on every system, and the front month is exactly where compute is most expensive. If your allocation slips a quarter, your data center sits powered and staffed with no hardware in it, your customers wait, and your competitor who locked a delivery month does not. A slipped quarter on a planned buildout costs far more than 2 percent. The forward is cheap insurance against the most expensive thing in procurement, which is unplanned schedule risk.

There is one more thing the forward buys that does not show up on the invoice. Every Rillor forward feeds the Rillor Compute Index, the 30-day rolling-blend forward price per SKU computed from active contracts. As a buyer you transact against real, observable forward prices rather than a vendor's quote of the week. You can see that curve forming on the marketplace and read the per-SKU detail on the skus catalog.

A decision matrix by buyer type

The forward dominating for known-date buyers does not mean spot is dead. Match the instrument to the situation.

• Buy spot when your need is genuinely immediate, small, and one-off. A research team that needs a handful of H200 systems this month for a fixed experiment should buy spot. There is nothing to plan around, so price certainty buys you nothing, and the speed of spot is its real value.
• Buy spot, briefly, when you are bridging a gap until a forward delivers. Spot is a fine stopgap. It is a poor foundation.
• Contract forward when you have a roadmap. A tier-2 cloud planning 12 months of capacity, a sovereign AI program staging a multi-rack buildout, or an enterprise standing up a training cluster all share one trait: a known delivery date and multiple tranches. That is the exact profile the forward is built for. See for buyers for the worked version.
• Contract forward, emphatically, when the unit is large and the schedule is hard. A GB200 NVL72 or a fleet of HGX B200 nodes (180 GB HBM3e per GPU, 7.7 TB/s of bandwidth, 1000W TGP, eight GPUs per board) tied to a power-on date is the worst possible thing to leave on spot or to a queue position. Lock it.
• Use allocation only when an OEM relationship gives you something a forward does not, such as bespoke configuration or a co-engineering arrangement, and you accept that you are holding a queue position, not a contract. For most standardized systems, the forward is the better expression of the same demand.

The short version: if you can name your delivery month, you should be on a forward. The market gave you spot and allocation because those were the only tools it had. There is now a third, and for the buyer with a plan, it is not close.

If you want to see how this maps to your own pipeline, start with the buyer overview and the standard contract terms.