For the complete documentation index, see llms.txt. This page is also available as Markdown.

Compute indices

GPU compute rental price indices, in USD per GPU-hour.

What is this endpoint for?

Get a live feed or historical values for any Kaiko Compute Index.

Compute Indices measure the market price of GPU compute rental, expressed in USD per GPU-hour. Values are calculated from provider offer data collected from compute marketplaces (e.g. Squaretower for NVIDIA H100), using a volume-weighted methodology (VWM or VWAP) with time-weighted aggregation (TWAP) over a rolling calculation window.

Indices are published as a real-time feed and as scheduled fixings.

Endpoints

gateway.equ.kaiko.io:443 kaiko.equities.EquitiesService/StreamComputeIndex
gateway.equ.kaiko.io:443 kaiko.equities.EquitiesService/GetComputeIndex

Full history available

Unlike the Reference Rates replay endpoints, GetComputeIndex is not limited to a rolling replay window: the complete publication history can be requested. Omit start_time and end_time to receive the last published value.

Each request returns at most 100 datapoints. For longer ranges, page by advancing start_time past the tsEvent of the last datapoint received.

Request parameters

Subscribe

Parameter
Description
Examples

index_codes

The list of Kaiko Compute Index tickers.

["KK_CMP_H100ST_RT"]

Requesting multiple tickers at the same time

To configure multiple tickers in the same call, provide the index_codes as a comma-separated list, e.g. ["KK_CMP_H100ST_RT", "KK_CMP_..."].

Historical

Parameter
Description
Examples

index_code

The Kaiko Compute Index ticker.

KK_CMP_H100ST_RT

start_time

Optional. Start of the range in ISO 8601. Omit for the last published value.

2026-08-17T00:00:00Z

end_time

Optional. End of the range in ISO 8601.

2026-08-17T12:00:00Z

Response fields

Field
Description

indexCode

The ticker identifying the index.

commodity

The type of publication. Either real-time (SIC_REAL_TIME) or fixings (e.g. SIC_HOURLY_FIXING).

interval

The time period in which observation data are considered for the calculation of the index.

.startTime

The start time of the data interval.

.endTime

The end time of the data interval.

quote

The quote asset used for the index denomination (usd).

bases

The list of base assets included in the index and their weight, e.g. h100_st.

exchanges

The data sources involved in the computation, e.g. sqrt (Squaretower).

price

The value of the index in the quote denomination (USD per GPU-hour).

methodology

The calculation methodology, e.g. vwm_twap, vwap_twap.

details

The different partitions included in the price computation.

.partition

The sequential number of the partition in the calculation window (0 = oldest).

.price

The volume-weighted price of the partition.

.volume

The total observation weight in the partition.

.count

The number of observations in the partition.

.underlyingTrade

The representative observation details of the partition.

.datetime

The datetime of the observation.

.exchange

The provider code.

.volume

The weight of the observation.

.id

The ID of the observation.

fallbackInfo

An object containing fallback information.

.fallbackUsed

true when the value is forward-filled (no new observations in the window).

tsEvent

The exact time of price publication.

sequenceId

A unique identifier for the publication.

lastIngestTime

The timestamp of the last source observation ingested.

Default values are omitted from the JSON output

Standard gRPC JSON behavior: fields at their default value are omitted — a missing partition means partition 0, fallbackInfo: {} means fallbackUsed: false, and an empty sequenceId is not shown. Treat absent fields as defaults, not as errors.

Forward-fill

When a calculation window contains no new observations, the index re-publishes the last computed value with fallbackUsed: true. Consumers can use this flag to detect stale values.

Weights as volume

Observation volume reflects provider weighting supplied by the data source, not traded quantity. underlyingTrade exposes one representative observation per partition for auditability; the full underlying dataset is not redistributed.

Request examples

cURL requests are intended for testing purposes only.

cURL requests are intended for testing purposes only.

GetComputeIndex returns {"datapoints": [ ... ]}; the stream pushes one datapoint per publication.

Response Example

Captured from the live production feed (note the omitted defaults: no partition on the single bucket, fallbackInfo: {}, no sequenceId):

Was this helpful?