Contango and backwardation in AI compute forwards.

A forward curve is the market's opinion about the future, expressed as a price for each delivery date. In commodities, the shape of that curve has a name. When far-dated contracts trade above the near contract, the market is in contango. When they trade below, it is in backwardation. Analysts coming from energy, metals, or grains carry a strong intuition about what those two states mean, and that intuition is useful right up to the point where it is exactly backwards for AI compute.

The reason is carry. The classic model says a far-month contract should cost more than the spot because someone has to hold the physical commodity until delivery, and holding costs money: storage, financing, insurance. A GPU system has those costs too, but they are dwarfed by a cost the cost-of-carry model never anticipated. The asset depreciates while you hold it, and it depreciates on a schedule, because a faster successor ships roughly once a year. On a compute forward curve, carry is negative, and that single inversion rewrites how you read term structure.

The textbook model, stated plainly

Strip away the jargon and contango is a simple idea. If you want a barrel of oil delivered in December rather than today, the seller has to buy it now and sit on it for months. They pay to store it, they pay to finance the inventory, they insure it. They pass those costs to you, so the December price sits above today's price. CME Group puts it crisply: in contango, prices for later expiration months are higher than the nearest month, and the curve slopes upward to reflect storage, financing, and insurance. For most storable commodities, contango is the normal state.

Backwardation is the opposite shape and the more interesting one. When the near contract trades above the far, the market is telling you that having the commodity right now is worth a premium over having it later. That happens when current supply is tight. Buyers who need the physical good immediately bid up the front, while the back stays anchored to a calmer expected future. Backwardation is a scarcity signal.

Two things are worth fixing in place before we move to compute. First, these are forward contracts in the strict sense. The CFTC defines a forward as a commercial merchandising transaction in a physical commodity where delivery is intended to occur but is deferred, privately negotiated, and customizable. That is exactly what trades on Rillor: bilateral OTC forwards with physical delivery always, never cash settlement. We have written about the precise line between a forward and a futures contract in forward contracts versus futures for GPU systems, and the distinction matters here because term structure on a physical-delivery book reads differently than on a cash-settled one. Second, the textbook curve assumes the underlying asset is fungible across time. A barrel of December oil is the same barrel as a March barrel. A GPU system is not.

Why compute inverts the intuition

Hold a barrel of oil for six months and you still have a barrel of oil. Hold a B200 system for six months and you have a system that is measurably less valuable, because the market now has access to something faster. That is negative carry, and it is the structural fact that breaks the cost-of-carry model for compute.

The driver is NVIDIA's roadmap cadence, which runs roughly annually. Blackwell B200 is shipping in volume. Blackwell Ultra, the B300 and GB300 NVL72, became available from partners including Dell, HPE, Lenovo, and Supermicro in the second half of 2025, and NVIDIA states the GB300 NVL72 delivers 1.5 times the AI performance of the GB200 NVL72. Behind that sits Vera Rubin in the second half of 2026, Rubin Ultra in 2027, and Feynman in 2028. A successor predictably arrives against each generation about once a year. That predictability is the whole point. It means the depreciation is not a random shock you cannot price. It is a scheduled erosion you can build into a curve.

The secondary market quantifies how steep that erosion is. Used H100 systems that commanded extraordinary prices during the mid-2024 shortage fell to roughly 150,000 to 180,000 dollars for an 8-GPU server by early 2026, down something like 62 to 70 percent from peak, and H100 cloud rental rates compressed from 7 to 10 dollars an hour at launch toward 2 to 4 dollars by late 2025. That is the empirical signature of negative carry. The asset bleeds value on contact with its successor.

So here is the inversion. In oil, deep backwardation is unusual and signals acute physical tightness. In compute, a downward-sloping curve is partly the default, because the negative carry pulls the far contracts down even when supply is comfortable. You cannot read a compute curve the way you read a crude curve. A B200 forward book can show the near contract well above the far and have that reflect two completely different things at once: real scarcity in the front, and scheduled obsolescence pulling down the back. Separating those two effects is the analytical task, and it is the reason a compute curve carries more information per data point than a commodity curve does. We treat depreciation itself as the dominant tradable risk in depreciation is the biggest risk in a GPU fleet, and the compute curve is where that risk becomes a price.

The front-month premium as a lead-time signal

The scarcity component lives at the front of the curve, and on a compute book it expresses itself through lead time. Data-center GPU lead times move sharply with supply tightness. NVIDIA H100 lead times ranged from up to roughly four months down to about 8 to 12 weeks as the shortage eased. When B200 lead times sit in the 14 to 20 week range, a buyer who needs delivery this quarter cannot simply wait, so they bid the near contract up. The front goes to a premium over the back not because the future is cheap but because the present is genuinely hard to source.

This is the cleanest reading of backwardation on a compute curve, and it is the one that maps directly onto the commodity intuition. When you see the front-month premium widen on a SKU, treat it as a lead-time alarm. It means the OEM order book for that configuration is full and the only way to get a system soon is to outbid someone. A procurement team watching that spread has a leading indicator on allocation tightness that arrives weeks before it shows up in their own RFQ responses. We map the underlying delivery windows by configuration in GPU system lead times by OEM and configuration, and the forward curve is the financial shadow those windows cast.

The subtlety is that a front-month premium and a depreciating back curve point the same direction graphically. Both produce a downward slope. The way to tell them apart is to watch how the slope moves. A pure lead-time premium is volatile and mean-reverting. It widens when allocation tightens and collapses when the order book clears. The obsolescence component is slow and one-directional. It tracks the successor's ship schedule and does not snap back. Decompose the slope into a fast, scarcity-driven front and a slow, obsolescence-driven back, and you have separated the two forces the textbook model never had to distinguish.

The obsolescence delta as the dominant far-curve driver

Past the first quarter or two, the curve is governed by the obsolescence delta: the rate at which the SKU loses value as its successor ramps. For a B200 system priced for delivery several quarters out, the forward price encodes a forecast of how much B300 supply will have substituted for B200 demand by that date. The further out you go, the more the successor dominates, and the more the far contract trades below where naive cost of carry would put it.

This is where two SKUs with similar spot economics can have completely different curve shapes, and it is the most useful thing a compute curve tells you that no other market structure can. Consider the contrast at the heart of this market today.

The AMD Instinct MI355X is built on 4th-generation CDNA with 288GB of HBM3E and 8TB/s of memory bandwidth, with expanded MXFP6 and MXFP4 support for inference, and it carries a 1400W peak board power. It was unveiled in mid-2025, and AMD's roadmap places the next-generation MI400 in 2026. Because its nearest successor sits further out than the B200's does, the obsolescence delta that drags the far curve down is smaller in the relevant window. The MI355X curve is flatter. Its far contracts are governed more by base demand and lead time than by an imminent generation transition. We compare where the part fits against Blackwell in MI355X versus B200 and B300, and the curve shape is the financial expression of that positioning.

The B200 curve is the opposite case. With B300 already shipping from partners and delivering materially higher performance, the obsolescence delta is large and active. The far contracts on RIL-GX-B200-2T trade at a steeper discount because the market is pricing real substitution. This is a generation-transition curve, and its slope is a direct, quantified read on how fast the market believes the transition will run.

Making term structure observable

None of this is readable from anecdote. You cannot eyeball a generation-transition curve from a handful of quotes, and you certainly cannot compare two SKUs side by side from scattered RFQ responses. Term structure becomes useful only when it is measured consistently, per SKU, from real transactions.

That is what the Rillor Compute Index does. It is a 30-day rolling-blend forward price per SKU, computed from active Rillor forward contracts, owned and controlled by Rillor, and licensed as a settlement feed and API to exchanges, funds, and researchers. Because every contract that feeds it is a physical-delivery forward with a verified buyer, a verified seller, independent escrow, and the end customer of record captured for channel compliance, the inputs are real economic commitments, not indications. The index turns the curve from folklore into an observable, comparable object for the first time. For the mechanics of how contract data becomes a curve, see how a forward curve forms from real contracts.

The index is also the part that scales beyond Rillor's own book. Third-party venues license it to build cash-settled derivatives, perps and futures, against the reference price. Rillor does not operate those venues and never cash-settles its own contracts. Rillor's contracts deliver hardware. The settlement feed is the public good that lets a fund hedge exposure or a researcher study term structure without ever touching a physical system. We make the case for why that feed, not the marketplace, is the durable asset in the index is the moat. You can see the per-SKU reference prices and request licensing on the for-markets page, and browse the underlying systems on the SKU catalog.

Reading the curve for a buildout

For a procurement team, the practical value is timing. A steepening curve and a flattening curve carry opposite instructions.

When a B200 curve steepens, meaning the far contracts fall faster relative to the near, the market is raising its forecast for how quickly B300 will substitute. That is a signal to either pull delivery forward and capture the systems you need before they depreciate further, or, if your workloads can tolerate the newer part, to wait and buy into the successor at a lower forward price. Either way, the steepening tells you the transition is accelerating and that holding B200 has gotten more expensive in opportunity terms.

When a curve flattens, the opposite holds. A flattening B200 curve says the market now expects the transition to run slower, that B200 demand is stickier than feared, and that the penalty for committing to forward delivery has shrunk. For a buyer whose stack is pinned to B200's specific memory and interconnect characteristics, a flattening curve is permission to lock a longer-dated forward without paying a steep obsolescence tax.

Reading two SKUs together sharpens the call further. Put the MI355X curve next to the B200 curve. The MI355X is flat because no near-term successor is pulling its far contracts down. The B200 is sloped because B300 is. If you are weighing the two for a buildout, the curve shapes tell you not just the price but the depreciation trajectory you are buying into. A flat curve is a longer usable life at a stable forward price. A steep curve is a shorter window before the next generation reprices your fleet. That is a procurement decision the curve makes for you, quantitatively, and it is the kind of analysis the marketplace and the index are built to support. For buyers building the full timing model, how to lock 12 months of GPU capacity without overpaying spot walks the playbook end to end.

The deeper point is that contango and backwardation in compute are not noise to be smoothed away. They are the market encoding two distinct facts, lead-time scarcity in the front and scheduled obsolescence in the back, into a single observable shape. The commodity reader's instinct to treat a downward curve as simple tightness will mislead. The compute reader's job is to decompose it. Once you do, the curve stops being a chart and becomes a forecast you can act on.