The Compute Futures Oracle Problem

Every derivative settles against something. That something is a benchmark, and a benchmark is only as good as the observations underneath it. This is the oracle problem, and it is not new: it has been worked through, sometimes painfully, in interest rates, in energy, and in physical commodities.

What went wrong elsewhere

The instructive failures share a shape. A benchmark is constructed from submissions or from a thin set of reported transactions. Participants who report are also participants who are exposed to the benchmark. The reporting process is not independently verifiable. The incentive to shade a submission is small per observation and large in aggregate. Reform, when it comes, moves the benchmark toward observed transactions and away from opinion.

The credibility of a settlement price is inherited from the credibility of the observations beneath it.

Why compute is a hard case

Compute inherits every structural difficulty and adds several of its own. Transactions are private and often bespoke. The good is not standardized, so two observations may not describe the same thing. Contract lengths vary from hours to years. A meaningful share of capacity never transacts at arm's length at all, because it sits inside vertically integrated operators. And the population of sellers is concentrated enough that a small number of reporters could move a poorly designed index.

What a defensible benchmark would require

• Observations tied to real transactions or real firm quotes, not indicative opinion.
• A published specification, so every observation describes a comparable good.
• Independence between the entity constructing the benchmark and the parties exposed to it.
• Transparent methodology: inclusion criteria, weighting, outlier handling, and minimum sample thresholds.
• An audit trail sufficient to reconstruct any published value after the fact.
• Explicit treatment of thin markets, including the choice not to publish.

The last point is the one most often skipped. A benchmark that always prints is a benchmark that will eventually print something it cannot defend. Mature methodologies specify the conditions under which no value is published.

Where the observations come from

There is no shortcut to a credible compute benchmark. It requires an accumulating record of specified requirements met with competing, dated, firm offers — the ordinary exhaust of a procurement process run with discipline. Build the procurement layer honestly and the observation layer follows. Attempt the index first and you are publishing opinion.

We are starting with execution for that reason. Price discovery is downstream of transactions, and transactions are downstream of standardized requirements.