Why a compute market needs an explicit reference#
Quotes for different accelerators, regions and delivery terms are not directly comparable. A single average can conceal incompatible hardware or stale offers. An index used for daily marks or settlement must define the economic exposure it represents before collecting prices.
The design goal is to avoid giving one provider unilateral control over the settlement price. Adding multiple inputs is only a starting point: correlated sources, duplicated quotes and low-volume executions can still produce a misleading reference.
Inputs need meaning as well as timestamps#
The brief proposes four types of controls. Volume filters limit the influence of economically insignificant activity; freshness rules address stale quotes; region filters preserve geographic comparability; and outlier handling reduces anomalous observations. None of their numeric thresholds is finalized here.
| Input | Potential information | Question before inclusion |
|---|---|---|
| Approved-provider quote | A provider offer under stated terms | Is it current, executable, capacity-backed and comparable? |
| Executed marketplace activity | Observed execution price and size | Is it genuine activity and which fees or incentives affect it? |
| External benchmark | Independent reference context | Does it match hardware, region, term and pricing basis? |
| Quality metadata | Age, source, volume and confidence | Can weak inputs be identified and excluded consistently? |
What the production methodology must publish#
The investor brief calls for methodology, input windows, confidence bands and an emergency fallback rule before production trading. An integration also needs versioned definitions, publication timestamps, market identifiers and a clear stale-state signal. These additional interface details are implementation requirements derived from the design, not current endpoints.
Changing weights or source eligibility can change position marks. A governance process therefore needs an announced version, effective time and audit trail. Quietly substituting a benchmark during an outage would make it difficult to explain a liquidation or final settlement.
Staleness, manipulation and basis risk#
A feed can fail by disappearing, publishing late or producing plausible but economically wrong values. Confidence checks should inform circuit breakers rather than treating every numeric response as tradable. The final policy must say whether quoting, funding, liquidations or settlement pause under each failure mode.
A buyer also faces basis risk: its actual workload cost may move differently from the chosen reference. Region, availability, memory, delivery period and negotiated terms can all differ from the index basket. A compute hedge should not be described as guaranteeing a fixed total workload bill.
What the current API actually returns#
GET /api/markets no longer returns price records at all. It reports status unavailable with an empty markets array, because the prototype has no verified supply or reference index to publish. Two proxies supply external context instead: /api/gpu-prices for current provider rates and /api/stock-history for listed-company closes. Their timestamps describe when the proxy fetched the data, not when any provider requoted.
There is no index publication endpoint, source-weight feed or live oracle attestation in the inspected server. Do not calibrate trading decisions or production monitoring against a third-party price reference republished without weighting, filtering or confidence bands.
Common questions
What are the final index weights?
They are not specified in the investor brief or implemented in the local API. They must be defined, published and tested before production use.
Does an asOf timestamp prove a live quote?
No. The price proxy timestamps its own fetch, and the markets route publishes no prices at all. Quote freshness requires evidence about the underlying observation.
Sources & implementation status
Based on the NeuraMesh investor brief and the project sources below. Source review: . Design explanations are not a deployment audit, legal opinion or commitment to launch.
Read our source and status conventions. Suggest a correction.