How a forward curve forms from real contracts.

A forward curve is one of the most misunderstood objects in markets, because the word "curve" makes it sound like a model. It is not. A forward curve is a set of prices that buyers and sellers actually agreed to pay for the same thing, delivered at different points in the future. Plot those agreed prices against their delivery months, connect the dots, and you have a curve. Nothing is fitted. Nothing is forecast. Every point on the line is a transaction that two counterparties signed.

For AI compute hardware, that distinction is the whole game. The question every procurement lead, fund analyst, and market researcher is now asking is what an HGX B200 system will cost to take delivery of in three, six, or nine months, while the next generation ships against it. A forward curve built from executed Rillor contracts answers that question with prices people committed real deposits behind, not with a survey of opinions. This piece walks through how each point on the curve is created, what the shape of the curve is telling you, and how the curve becomes the input to the Rillor Compute Index.

What a forward curve actually is

Start with one SKU. Take the HGX B200 system, the 8-GPU Blackwell node that ships through every major OEM. It is a real, deliverable product: NVIDIA's DGX B200 reference pairs 8 Blackwell GPUs for 1,440 GB of HBM3e at 64 TB/s, fifth-generation NVLink at 14.4 TB/s aggregate, and 144 PFLOPS of FP4 Tensor Core throughput. The same architecture is sold as Lenovo's ThinkSystem SR680a V3, the Dell PowerEdge XE9680L, the Supermicro SYS-A22GA-NBRT, and the Gigabyte G894-AD1-AAX5, among others. When we talk about a forward curve for "the B200," we mean a standardized Rillor SKU (RIL-GX-B200-2T) that resolves to one of these qualified OEM configurations on delivery.

Now imagine that across the last month, verified buyers and sellers executed forward contracts on that SKU for several different delivery months. A contract for May delivery cleared at one price. A contract for August delivery cleared at a different price. A contract for November cleared at a third. Each of those is a single transaction. The forward curve is simply those clearing prices, indexed by their delivery month and laid side by side.

That is the definition exchanges use too. As CME Group puts it, the forward curve is a graph of prices across delivery dates, and the curve can shift between shapes as expectations change. The only thing that makes Rillor's curve different from an exchange forward curve is what stands behind each point. On Rillor, each point is a bilateral OTC forward with physical delivery intended on both sides, not a cash-settled futures print.

Each point is an agreement, not an opinion

This is the part that separates a Rillor curve from most "price indices" you will see quoted for GPUs. A great deal of what circulates as compute pricing is survey data: someone called ten brokers, took a midpoint, and published it. That is an opinion curve. It moves when sentiment moves, and nobody is on the hook for any of the numbers.

A Rillor curve point is a contract. Two KYC'd counterparties, a standardized contract, a 10% deposit posted at execution, a seller performance bond, an independent escrow agent holding the balance until delivery, and NVIDIA channel compliance with the end-customer of record captured. To create a point on the curve, somebody had to be willing to commit a deposit to buy and somebody had to be willing to bond a promise to deliver. The price that results is information that was paid for. We wrote up the contract mechanics in detail in the anatomy of a Rillor forward contract, and the reason we settle physically rather than in cash in why Rillor settles physically and never cash-settles. Both are load-bearing for what follows, because the integrity of every curve point comes from the obligation behind it.

Reading the shape: contango and backwardation

Once you have points across delivery months, the curve has a shape, and the shape carries information. There are two canonical shapes, borrowed from every other forward market.

Contango is an upward-sloping curve: later delivery months trade above near months. In classic physical commodities this reflects the cost of carry, the financing, storage, and insurance you incur to hold the asset until a later date, plus an expectation of rising prices. If it costs money to hold a barrel of oil for six months, the six-month forward has to compensate the holder for that, so it sits above spot.

Backwardation is a downward-sloping curve: later delivery months trade below near months. As Britannica's markets reference frames it, backwardation appears when near-term demand outruns available supply and buyers pay a premium to secure scarce units now rather than wait, a convenience yield for having the thing today. The deferred months sell off relative to the front because the market expects the squeeze to ease.

Most other forward markets spend their lives oscillating between these two states as expectations shift. Compute is unusual in that it has a strong structural bias toward backwardation, and the reason is specific to this asset class.

Why compute curves bend downward

A barrel of oil delivered in November is the same barrel delivered in May. A B200 system delivered in November is competing, on that delivery date, against a more capable generation that has begun shipping. That is the difference, and it dominates the shape of a compute forward curve.

NVIDIA's published roadmap is an annual refresh cadence: Blackwell Ultra (the GB300 NVL72 generation) in the second half of 2025, Vera Rubin in the second half of 2026, Rubin Ultra in 2027, Feynman in 2028, with each step roughly multiplying prior capacity. The successor to the B200 is already real and already in the channel. Blackwell Ultra (B300) carries 288 GB of HBM3E and 15 PetaFLOPS of dense NVFP4 per chip, against 192 GB and 10 PetaFLOPS for the standard B200, roughly 50% more memory and 50% more compute per chip, at up to 1,400W versus 1,200W. The standard B200 and the B300 trade concurrently within the same Blackwell generation.

So when a buyer prices a B200 for November delivery, they are pricing in the fact that B300 supply will have partially substituted for B200 demand by then. The buyer who could run their workload on either generation discounts the deferred B200 aggressively. The buyer whose pipeline is pinned to B200's exact memory and interconnect profile pays closer to the front. The forward market surfaces that disagreement, and the net effect is a curve that bends downward over time. We treat that downward bend as a first-class, modelable quantity, the obsolescence delta, in the forward curve, the spot price, and the obsolescence delta. The headline for this piece is simpler: obsolescence expectations are the single largest force shaping a compute forward curve, and they almost always push it into backwardation.

This is also why a compute curve is not static. If B300 supply slips, near-term B200 scarcity intensifies and the front lifts, steepening the backwardation. If a hyperscaler cancels a large B200 commitment, the front softens and the curve flattens. The curve is a live readout of how the market is weighing supply, demand, and the obsolescence clock, month by month.

A worked example: B200 across three delivery months

Make it concrete. Suppose that over a recent window, the following B200 (RIL-GX-B200-2T) forward contracts executed on Rillor. All prices are per complete 8-GPU OEM system, indicative and rounded for illustration.

Three points. Three real agreements. Connect them and you have a downward-sloping curve, backwardation, with the deferred months trading below the front. Read it back in plain language:

• The market will pay roughly $352,000 today to lock a B200 system for May.
• For the same system three months later, it will pay about $21,000 less, a 6% discount.
• For November, the discount widens to about $40,000, just over 11% below the front.

The roughly two percent per month decay is not a coincidence and it is not a fitted parameter. It is the market's collective pricing of the obsolescence delta: how much less a B200 is worth to take delivery of as B300 availability improves through the year. A procurement lead reads this curve and gets an immediate, defensible answer to "should I lock May or wait for November." A fund reads the same curve and gets a depreciation schedule for the asset class. Both are reading the same executed prices.

Note what the curve is not telling you. It is not a prediction that B200 systems will be "worth" $312,000 in November in some abstract sense. It is the price at which buyers and sellers were willing to contract today for November delivery. That is a sharper, more honest number than a forecast, because someone committed a deposit to each end of it.

From curve to index: the Rillor Compute Index

A raw curve made of individual contracts is the right primitive, but it is noisy. Contracts execute at irregular intervals, in different sizes, with slightly different delivery windows. A single large contract at an off-market price can distort the picture if you read points naively. To turn the curve into a reference price that an exchange or a fund can settle against, you need a disciplined aggregation. That is the Rillor Compute Index.

The index is a 30-day rolling-blend forward price per SKU, computed from active Rillor contracts. "Rolling-blend" means it blends the executed forward prices observed over a trailing 30-day window, weighting by recency and size, so that a single print neither sets the level nor is ignored. "Per SKU" means each standardized Rillor SKU has its own index value: the B200 has one, the B300 (RIL-GX-B300-2T) has another, the GB300 NVL72 rack (RIL-NVL72-GB300) has another. The result is a single, clean, reference forward price per SKU that updates as new contracts execute, sitting on top of the noisier underlying curve. The full field-by-field methodology lives in how the Rillor Compute Index is computed.

The index is computed from, owned by, and controlled by Rillor. It is licensed as a settlement feed and API to exchanges, funds, and researchers. Third-party venues build cash-settled derivatives, perpetuals and futures, that reference it. Rillor does not operate those venues and never cash-settles its own contracts. Rillor's contracts are physical-delivery forwards, full stop. The index is the downstream layer that lets the broader market take exposure to the price discovered on Rillor without ever touching a physical contract. If you build or operate a venue, you can see what it takes to reference the feed on the markets page.

Why a physically settled curve resists manipulation

This is where the choice to settle physically pays off in the data, and it is the reason a researcher should trust this curve over a cash-bet curve.

The CFTC's long-standing distinction between forwards and futures turns on intent to deliver. A forward, in the regulator's framing, is a commercial merchandising transaction in a physical commodity, between counterparties who intend that physical transfer of the actual commodity will occur, with delivery deferred for commercial reasons. In a cash-settled futures contract, by contrast, a party may make or take delivery but need not expect to, and frequently does not. The transaction can be closed out for cash before delivery ever comes due. That is the regulatory line, and it is also the manipulation line.

Consider what it takes to push a curve point in each regime. To move a cash-settled price, you place bets; if you have enough capital, you can paint a print and walk away, because nobody has to deliver anything. To move a Rillor curve point, you have to execute a real forward: post a 10% deposit, pass KYC, and either bond a commitment to deliver complete OEM systems or commit to take delivery of them with the end-customer of record captured under NVIDIA channel compliance. You cannot paint the tape with an order you never intend to honor, because every point on this curve is an obligation to actually move hardware. The cost of distorting the curve is the cost of really buying or really selling the systems, which is exactly the friction you want protecting a reference price. We go deeper on the safeguards in what makes the Rillor Compute Index hard to manipulate.

That property is the foundation of the index as a business. A settlement feed is only worth licensing if the underlying prices are hard to fake. Physical delivery is what makes them hard to fake. The curve is honest because the obligations behind it are real.

What to do with the curve

For a procurement lead, the curve is a timing tool. Read the slope, compare the cost of locking the front against the discount on deferred months, and weigh that discount against your own substitution flexibility and the risk that near-term scarcity steepens further. For a buyer evaluating whether to commit forward at all rather than chase allocation, the buyers page lays out the contract path; the underlying SKUs are catalogued at the SKU index, and live contract activity is visible on the marketplace.

For an analyst or fund, the curve is a depreciation schedule and a hedging instrument in one. The slope is a directly observed, contract-backed estimate of the obsolescence delta for each SKU. The index built on top of it is what you settle a cash-settled product against, if you license the feed.

For everyone, the discipline is the same: every number on this curve was paid for. A forward curve formed from real contracts is the closest thing the compute market has to ground truth, because it is made entirely of decisions that two counterparties were willing to stand behind.