Compute Derivatives Basis Risk Primer
What risks are there for buyers and sellers of compute derivatives?
Please see relevant disclosures here.
This post attempts to present a taxonomy of basis risks for compute derivatives. Considering how nascent the market for compute derivatives is, consider the following to be a reasonable, albeit likely imperfect, attempt at surveying the terrain. Caveat emptor applies.
Everyone involved in the nascent market for compute derivatives is worried about basis risk. And the failure mode for the first compute futures contracts is fairly predictable: a hedger trades one, discovers that it does not track his book, and does not come back. Without that hedger, the volume that was supposed to mature the contract never arrives.
That failure is a basis risk failure, and it is the risk every camp building a futures market is implicitly making a bet about. Most of the conversation gets the shape of the risk wrong. The imported intuition is that basis is high when a contract is new and declines as the contract seasons. That intuition is correct for corn and crude. It is wrong for compute, because it classifies compute basis as an early market liquidity problem, when the binding constraint is a structural tracking error problem created by hardware turnover.
This primer lays out a taxonomy of basis risk, identifies the one property that separates a basis you can reserve against from one you cannot, and explains why that property is what determines whether a compute contract lives.
The thing hedging cannot remove
A hedge does not eliminate price risk. It converts price risk into basis risk. You hold an exposure, such as a book of GPU capacity, a financing commitment against a neocloud, a forward purchase obligation, and you offset it with a traded contract. The contract is never the exact good you hold. The residual gap is the basis, and the variance of that gap over your hedge horizon is the risk that survives the hedge. Whether a derivatives market is useful reduces to whether that residual is small enough, and stable enough, that hedgers accept it in exchange for the price protection the contract provides. If they don’t, they leave, and the contract never earns the liquidity that was supposed to save it.
Basis risk is therefore not a nuisance term sitting beside the real market risk. In a functioning derivatives complex, it is the risk. And it is not one thing.
Basis is not one risk
What gets lumped together as “basis” decomposes into distinct sources, each with its own behavior and, critically, its own stabilizing mechanism.
Liquidity, or microstructure basis, is the gap from immature price discovery: wide bid-ask, uncertain hedge ratios, thin open interest, paper that does not yet anchor to physical because not enough has traded. This is the term that genuinely decays in the way the conventional intuition describes, but it decays as a function of cumulative volume, not calendar time. And it turns out that this distinction matters enormously.
Locational basis is the gap between the reference point of the contract and the location of your exposure. Henry Hub is not the price at Algonquin; Cushing is not a coastal barrel; a nodal power price is not the price two buses away.
Quality or grade basis is the gap between the deliverable specification and the spec you hold. Light sweet crude versus a heavy sour barrel; deliverable grade wheat versus what is in your bin.
Calendar or term structure basis is the gap from maturity and roll mismatch. It’s the relationship between spot, the front contract, and the deferred curve, governed by storability and the cost of carry.
Settlement basis is the gap that survives because of how the contract resolves. A physically deliverable contract forces convergence: at expiry, paper and physical must meet, and arbitrage closes the gap. A cash-settled contract converges only to its index, and the gap between that index and your exposure is set by index methodology, with nothing forcing it to zero.
Obsolescence basis is the gap created when the economic identity of the underlying drifts over time. This one does not appear in classical commodities at all, and it is the crux of the compute case.
Every basis but one has an anchor
An interesting question that arises here is whether a given basis has a stabilizing anchor: a convergence mechanism, an arbitrage bound, a delivery specification, or a repeatable observable differential that holds it within a regime.
An anchored basis is haircuttable. It can be violent in the moment, such as when gas basis blows out, ERCOT scarcity spikes, or WTI prints negative. But within a stable regime, it is bounded by something physical or contractual, so you set a haircut, reserve against its variance, and move on. The entire classical commodity hedging toolkit of basis swaps, locational differentials, and quality adjustments, exists to manage anchored basis.
Obsolescence basis has no such anchor, because the reference object itself, such as an H100, decays relative to the exposure set. There is no arbitrage that closes the gap between an H100-hour and the book of a hedger who is migrating to B200s, because the two are not the same economic good and no trade converts one into the other. The differential is not repeatable; it is a one-way drift that resets at each hardware generation. You cannot haircut that, because there is no fixed level to haircut against. The thing you would reserve against is the variable that is moving. This is why obsolescence basis is corrosive in a way that locational basis is not, even though locational basis can be far larger in any given month. Magnitude is not the problem; the absence of an anchor is.
With all of this established, the taxonomy of basis risks resolves into a single matrix.
The matrix
Ratings are relative within each row. The “Anchored?” column indicates whether the basis has a stabilizing mechanism that bounds it within a regime (haircuttable) or lacks one (not).
Reading the table
Three things become legible once the basis types are laid out this way.
Classical commodities are clean on the diagonal. Each conventional commodity has essentially one dominant basis and is otherwise quiet. Crude’s is quality. Gas and power live on locational and calendar basis. Gold is clean on everything because it is fungible, physically deliverable, and timeless. These differences explain why risk managers in each market specialize: a power desk obsesses over nodal congestion and shape; a crude desk obsesses over grade differentials. Each market hands you one hard problem.
Compute is the only column that lights up every row. It inherits power’s locational and term structure problems, crude’s and grain’s quality heterogeneity across SKUs and tenancy, a settlement gap that is methodological rather than physical, and the obsolescence drift on top. No prior market required one contract to manage all six at once.
The anchor line is the line between manageable and corrosive. Every row above the bottom has a stabilizing mechanism, even when it is large or shifts between regimes. The obsolescence row is the only “No,” and it is the only row where the standard apparatus fails at the root. And it fails because there is no level to reserve against. That single cell, not the simultaneity, is the reason that compute is not simply power markets all over again.
DRAM and NAND are the warning case, not the proof. Memory is the only other column with anything in the obsolescence row. It belongs here as an adjacent precedent for what happens when the reference object decays under the contract. The conceptual point survives regardless: turn the obsolescence terms up and the difficulty compounds.
The two clocks
Compute basis runs on two clocks that the conventional basis declines over time story collapses into one.
The first is the liquidity clock, which runs on cumulative volume. As trading accumulates, price discovery matures, the hedge ratio firms, bid-ask compresses, and the microstructure term decays toward zero. Within its domain, the conventional intuition is correct.
The second is the hardware clock, which runs on the generational cadence of the underlying. An H100-hour is not the economic good a B200-hour is, and neither is whatever Rubin-class capacity clears eighteen months out. At each generational boundary, a new wedge opens between the reference SKU and the book a hedger actually carries. The level of that wedge is not reliably directional. It depends on whether the hedger owns older capacity, newer capacity, reservations, spot exposure, a regionally constrained fleet, or customer obligations benchmarked to performance rather than SKU. What is reliable is that the irreducible tracking error resets upward and hedge effectiveness gets knocked down. The wedge reappears every generation regardless of its sign.
Superimpose the two clocks on a single calendar axis and you get a misleading picture of smooth decline. Separate them and the real shape appears: effectiveness rises on the liquidity clock, then drops on the hardware clock. This is a sawtooth, not an exponential to zero. Inside a single generation, for a well-specified contract on a liquid reference SKU during its peak deployment window, basis behaves and effectiveness climbs as the conventional story says. Across generations, it does not. Convergence is not a property of time. It is a property of where you are in the hardware cycle.
The failure mode
The two clocks are not independent, and their interaction is what actually kills contracts.
If the obsolescence wedge is large enough, hedgers discover the contract does not track their book. Usage stalls. Open interest never accumulates. And because the liquidity clock runs on cumulative volume, the microstructure term never gets the volume it needs to compress. Unanchored tracking error suppresses the hedging utility that would have built the liquidity that would have tightened the microstructure basis. The contract dies because the wrong people traded it at once, found the hedge did not hold, and did not come back. Liquidity was endogenous to a tracking quality the contract failed to deliver.
This is the dynamic the memory comparison gestures at and the dynamic compute faces in sharper form: a faster generational clock multiplied by a larger per-generation economic delta. No liquid commodity complex has previously had to price a basis where both terms are large at once. That is the genuinely novel object, and it must be defeated before liquidity can occur.
Generic benchmark versus vintage curve
The most important design consequence follows directly. A well-specified, generation-local contract, such as an H100 reference during peak H100 deployment, is not obviously doomed. Its obsolescence wedge is small within its deployment window, because the reference good and the installed base are the same generation. What inherits the full obsolescence problem is the generic contract: the “GPU-hour” or “AI compute unit” that pretends H100, B200, GB200, Rubin, tenancy, interconnect, region, and reservation terms can be blended into one stable reference good. That contract bakes the drift into its own definition and never escapes it.
So the answer to “just list generation-specific contracts” is not that it defeats the argument. It is that it changes the market’s shape. You do not get one eternal Henry Hub for compute. You get a succession of vintage specific curves, each with a useful life tied to a hardware deployment window, separated by rollover discontinuities where one generation’s curve hands off to the next. The market’s job is not to eliminate obsolescence basis. It is to quarantine it by generation and build transition instruments, such as calendar spreads across vintages, roll-management hedges, and generation migration swaps, around the discontinuity. That is a more demanding and more interesting structure than a single benchmark, and it is the structure a serious contract designer should be building toward.
If you enjoy this newsletter, consider sharing it with a colleague.
I’m always happy to receive comments, questions, and pushback. If you want to connect with me directly, you can: