WindBorne Systems
Last reviewed
Jun 7, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,880 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 · 1,880 words
Add missing citations, update stale details, or suggest a clearer explanation.
WindBorne Systems is a Palo Alto, California weather technology startup that operates a global constellation of long-duration, altitude-controlled sounding balloons and builds a transformer-based AI weather forecasting model called WeatherMesh. Founded in 2019 by a group of Stanford University graduates and led by co-founder and chief executive John Dean, the company sells in-situ atmospheric observations to forecasting agencies and commercial customers, and sells forecasts produced by its own model. WindBorne sits at the center of the broader trend in which machine-learning systems trained on historical reanalysis data have begun to rival, and on some measures exceed, the physics-based numerical weather prediction run by national agencies. The company's performance claims are self-published and, as of June 2026, have not been independently certified, although its balloon data has been validated and assimilated by the U.S. National Oceanic and Atmospheric Administration (NOAA).
WindBorne grew out of the Stanford Student Space Initiative, the university's largest hands-on engineering group, where students began experimenting around 2015 with ways to extend the flight duration of conventional weather balloons. The company was incorporated in 2019. Its founders are John Dean (chief executive), Andrey Sushko (chief technology officer), Kai Marshland (chief product officer), Joan Creus-Costa (head of AI), and Paige Brown. Dean had served as co-president of the Stanford Student Space Initiative.
The founding thesis is that roughly 85 percent of the planet, principally oceans, polar regions, and the developing world, is poorly sampled by the existing global observing system, which depends on a few hundred fixed radiosonde stations that launch single-ascent balloons once or twice a day. WindBorne's approach was to build balloons that stay aloft for days or weeks, navigate by changing altitude to ride different wind layers, and collect many vertical profiles per flight in those data-sparse regions.
WindBorne's network is named Atlas, and the company describes it as the largest balloon constellation in the world, collecting in-situ atmospheric soundings from pole to pole. The aircraft are autonomous, long-duration, high-altitude platforms that WindBorne calls Global Sounding Balloons (GSBs). Unlike a conventional weather balloon, which rises once and bursts within hours, a GSB uses dynamic altitude control to stay aloft, vent or ballast to change height, and steer by selecting winds at different altitudes. The company uses sand as ballast.
| Atlas / GSB attribute | Figure | Source |
|---|---|---|
| Balloons aloft at a given time | ~400 (June 2026); "hundreds" globally | TechCrunch; WindBorne |
| Launch sites | 15 worldwide | TechCrunch / WindBorne |
| Target full-scale fleet | 10,000 GSBs (goal by 2028) | WindBorne |
| Typical flight duration | 7 days average (12+ days typical, max 16 observed) | NOAA WPO / WindBorne |
| Demonstrated endurance | 75+ days | WindBorne |
| Vertical profiles per flight | 40 to 50 or more | NOAA WPO / WindBorne |
| Launch weight | Under 3 pounds (~1.2 kg) | WindBorne |
| Launch time | ~30 minutes by a single operator (winds up to 15 knots) | NOAA WPO |
| Data measured | Pressure, temperature, humidity, wind speed and direction | WindBorne |
| Total balloons launched (lifetime) | More than 4,000 (as of late 2025) | WindBorne / NTSB reporting |
A GSB collects data from ground level into the stratosphere and can produce dozens of vertical profiles over a single multi-day flight, against the single ascent or descent of a radiosonde or dropsonde. In a NOAA-supported atmospheric-rivers campaign from February 15 to March 21, 2022, 65 WindBorne balloons launched from Seoul, South Korea, and Maui, Hawaii, accumulated 282 days of flight time and more than 11,500 vertical kilometers of profiles. One balloon, W-273, completed a global circumnavigation 16 days after launch. WindBorne has also demonstrated deploying dropsondes into storms and what it calls the first balloon-to-buoy capability for extended ocean coverage.
The constellation's scaling has been steady rather than explosive: NOAA and WindBorne reported an average of roughly 90 GSBs aloft as of December 2025, with the company citing about 400 in flight by mid-2026 and a long-term target of 10,000 concurrent balloons.
WindBorne began as a balloon and data company, then moved into forecasting after deep learning weather models such as DeepMind's GraphCast emerged starting in 2022. Its model, WeatherMesh, is built on the transformer architecture, the same family of neural networks underlying large language models and much of modern generative AI. WeatherMesh uses an encoder-processor-decoder design: an encoder compresses raw weather fields into a latent representation, a processor advances that state forward in time, and a decoder maps the latent state back to real-world variables. The encoder and decoder use U-Net-style convolutional blocks, while the processor is built from neighborhood-attention 3D transformer layers.
WeatherMesh is trained primarily on ERA5, the European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis dataset that has become the standard training corpus for machine learning weather models, supplemented by real-time observations from U.S. and European agencies and, increasingly, WindBorne's own balloon data. This places WeatherMesh squarely within the AI for science movement of data-driven, learned simulators.
A distinguishing feature is computational efficiency. WindBorne reports that an earlier version, WeatherMesh-2, can compute a 14-day global forecast at 0.25-degree, 6-hourly resolution in about 12 seconds on a single Nvidia RTX 4090 graphics card, and that hourly global forecasts to 14 days run in under 10 seconds on an H100. The company says it trained the model on a cluster of a few dozen RTX 4090 GPUs costing roughly $100,000, and that training used about one-fifteenth the compute of GraphCast and one-tenth that of Huawei's Pangu-Weather.
| WeatherMesh version | Date | Notable details |
|---|---|---|
| WeatherMesh-2 | 2024 | ERA5-trained; 14-day forecast in ~12s on one RTX 4090; company says it out-scored GFS, ECMWF HRES, GraphCast on 8 to 24 percent of metrics, and matched or beat Microsoft's Aurora on 79 percent of evaluated targets |
| WeatherMesh-4 | 2025 | Company reports a full forecast roughly every 10 minutes operationally |
| WeatherMesh-6 | June 2026 | Hourly forecasts at 3 km resolution over the continental U.S.; first version with heavy direct ingestion of WindBorne balloon data |
WindBorne's commercial foundation is the sale of its balloon observations. The data is distributed globally over the World Meteorological Organization's Global Telecommunication System (GTS) and is assimilated into NOAA's Global Forecast System (GFS), the U.S. flagship physics-based model. Beginning in August 2025, observations from the Atlas constellation were assimilated into operational systems including NOAA's GFS and WindBorne's own WeatherMesh. The company says delivery latency averages about seven minutes from collection to the forecaster.
NOAA's Weather Program Office partnered with WindBorne to develop the balloons for collecting surface-to-stratosphere data in lesser-sampled regions, and NOAA describes the output as "radiosonde quality" with global reach at lower cost. NOAA studies have found that assimilating WindBorne data improves the accuracy of its physics-based forecasts. On the customer side, WindBorne sells balloon data to NOAA, the U.S. Air Force, and the U.S. Navy, and sells forecasts to commercial users including commodity traders and investors. The company has described its core business, spanning U.S. military and private-sector customers, as cash-flow positive.
WindBorne has raised approximately $25 million in venture funding across a seed and Series A round, and was reported at an $85 million valuation in 2024.
| Round | Date | Amount | Lead | Selected participants |
|---|---|---|---|---|
| Seed | June 2023 | $6 million | Footwork | Khosla Ventures, Pear VC, Ubiquity Ventures, Harvest Ventures, Humba Ventures |
| Series A | June 2024 | $15 million | Khosla Ventures | Footwork, Pear VC, Convective Capital |
WindBorne said the Series A would fund scaling the constellation toward 10,000 concurrent balloons by 2028, improving autonomous flight software for profile targeting and endurance, and expanding the forecast model to cover more weather variables in greater detail. As of June 2026, public reporting describes the seed and Series A but no separately announced Series B.
WindBorne's headline marketing claim is that WeatherMesh out-forecasts the world's leading agencies. CEO John Dean has written that WeatherMesh became "the most accurate AI forecasting model in the world" in 2024 and is "up to 30 percent more accurate" than ECMWF models on some measures. For WeatherMesh-6, chief product officer Kai Marshland told TechCrunch in June 2026 that the model "is as accurate five days out as a traditional forecast is the day before," especially for surface temperature, and that it updates hourly against the roughly six-hour cycle of conventional global models.
These accuracy figures are company-published benchmarks. As of June 2026, they have not been independently certified by a third party, and they should be read as vendor forecast-skill claims rather than verified results. The strongest external corroboration is narrower and more specific: NOAA's own studies show that WindBorne's balloon observations improve physics-based forecasts, and NOAA assimilates that data operationally, which is an independent validation of the data, not of the WeatherMesh model's overall skill ranking. Standardized public leaderboards such as WeatherBench exist for comparing models, but a fully independent, head-to-head certification of WeatherMesh-6 against ECMWF had not been published at the time of writing.
WindBorne's profile also drew scrutiny over flight safety. On October 16, 2025, a United Airlines Boeing 737 MAX 8 (flight UA1093, Denver to Los Angeles) suffered a cracked windshield at about 36,000 feet (flight level 360) over Moab, Utah, injuring one pilot and forcing a diversion to Salt Lake City. The National Transportation Safety Board's preliminary review indicated a WindBorne GSB launched from Spokane, Washington, on October 15 passed through the same area at nearly the same altitude and stopped transmitting around the time of impact. WindBorne said it had launched more than 4,000 balloons and coordinates with the Federal Aviation Administration for each flight; the sand ballast was consistent with the "sand-blasted" appearance of the damaged windscreen.
WindBorne illustrates the broader AI weather shift: combining a novel in-situ observing network with fast, learned models trained on reanalysis data, run on commodity GPUs rather than supercomputers, to challenge the agency-run numerical weather prediction that has dominated forecasting for decades.