Why an AI weather company sells balloon observations by the kilometre

TechCrunch reported on 12 October that the Air Force had awarded WindBorne a data contract worth up to $49 million.[1]

The interesting business question is what a customer buys before there is a forecast. WindBorne’s September explanation supplies a concrete answer: a sounding equivalent represents 10 kilometres of vertical travel while collecting atmospheric measurements. It is a unit for the amount of observation data delivered, not a count of balloons or an accuracy score.[2]

A balloon can contribute more than one sounding

A conventional radiosonde collects a profile as its weather balloon rises. WindBorne’s Global Sounding Balloons repeatedly move through altitude bands during longer flights. The company says its Atlas platform measures temperature, humidity, pressure and wind, uses altitude control to change the sampled layer, and sends observations through satellite communications.[3]

This matters because the same balloon can contribute repeatedly, and several balloons can contribute to the same customer’s specified area and period. WindBorne says orders can specify geography, altitude, time and other parameters. Billing is attached to the eligible data delivered across the constellation.[2]

Take an original, invented example. Three balloons collect eligible profiles covering 18, 27 and 35 vertical kilometres in a customer’s requested area. Together that is 80 kilometres, or eight sounding equivalents. A fourth balloon outside the requested area contributes nothing to this example’s order, however far it travels.[2]

If only half of the third balloon’s vertical travel qualifies, the total becomes 18 + 27 + 17.5 = 62.5 kilometres, or 6.25 sounding equivalents. The calculation illustrates the company’s unit; the inputs and eligibility scenario are ours. It is not a quotation for eight balloon launches, and no dollar price can be inferred without a commercial rate.

Observations, predictions and blends are different goods

WindBorne’s WM-6 announcement describes a model that incorporates observations through AI-based data assimilation. One ingredient is an innovation: the difference between a measurement and the model’s background estimate at the corresponding location. Its global ensemble produces 128 members at roughly 25-kilometre resolution. The ensemble provides multiple possible evolutions from varied starting conditions.[4]

For a simple illustration, an observation of 15°C against a background estimate of 12°C gives an innovation of +3°C. That does not instruct the forecaster to add three degrees everywhere. The useful problem is how information about one place should affect the estimated atmospheric state around it. This example explains the input relationship; it does not reproduce WeatherMesh’s learned update rule.

MetaMesh works a stage later. The company says it blends forecasts from several AI and physics-based models, adjusts weights as performance changes, and refreshes when new inputs arrive. It offers hourly steps out to 15 days. The published evaluation distinguishes station-based and gridded verification, rather than presenting every forecast location as an independently instrumented station.[5]

The asset can be useful to somebody else’s model

The observations API returns balloon measurements as records and supports polling for new or updated data. That delivery model helps explain the commercial separation: a customer can consume observations into its own pipeline without treating WindBorne’s forecast as the final answer.[6]

Our business interpretation is that the defensible asset begins with repeatedly collecting useful measurements that others do not already possess. A forecast can monetise those measurements internally; a data service can monetise their usefulness to other forecasters. A blend can then compete on the finished prediction delivered to an operator.

Those revenues would have different dependencies. Observation sales depend on usable coverage and delivery; forecast sales depend on performance at the customer’s required variable and lead time. Calling them all an “AI weather product” conceals those differences. The sounding equivalent makes the first transaction unusually tangible: a defined quantity of the atmosphere measured, aggregated and delivered.

Report an error or suggest a correction