Jua
Last reviewed
Jun 7, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,912 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 7, 2026
Sources
15 citations
Review status
Source-backed
Revision
v1 · 1,912 words
Add missing citations, update stale details, or suggest a clearer explanation.
Jua is a Zurich-based artificial intelligence company building large foundation models for weather and the wider physical earth system. Founded in 2022, the company develops a family of models it calls the Earth Physics Transformer (EPT), large spatiotemporal transformer networks trained on petabytes of atmospheric and earth-observation data. Jua positions its work as a "Large Physics Model" that learns the dynamics of the atmosphere directly from data rather than solving the differential equations used by conventional numerical weather prediction. The company markets the technology through an "Earth Intelligence Platform" aimed primarily at energy traders and grid operators, with insurance and other physical-risk sectors named as longer-term targets. Jua is an early-stage company that had raised roughly $27 million across pre-seed, seed, and Series A rounds as of mid-2025.
Jua sits within the fast-growing field of AI weather forecasting, alongside research systems such as Google DeepMind's GraphCast, Microsoft's Aurora, Nvidia's FourCastNet and Earth-2, and Huawei's Pangu-Weather. Like those efforts, it applies deep learning to a problem that has historically been the domain of physics-based supercomputer simulation. Jua's distinguishing pitch is that it is building a broad foundation model for science covering the natural world, not a single forecasting product.
Jua was founded in 2022 by Andreas Brenner and Marvin Gabler, following what the founders described as several years of prior research and development on a proprietary data pipeline. Brenner served as chief executive from founding, while Gabler, who has a background in geophysics and meteorology, served as chief technology officer and is credited as the architect of the EPT models. The company is headquartered in Zurich, Switzerland, and has historically operated additional hubs in Berlin and Cape Town. Gabler's personal trajectory has been highlighted in company profiles, which describe a path from building game modifications to constructing what Jua calls a digital twin of the Earth.
In June 2025, coinciding with the company's Series A round, Marvin Gabler was appointed chief executive officer, taking over the role from Andreas Brenner, who had led the company since its founding. The company framed the transition as a move to focus leadership on scaling the platform in the energy sector. Gabler is the public face of the company's later technical announcements, including the EPT-2 launch.
Jua's stated mission is to build a foundational AI model for the natural world, beginning with weather and climate and extending toward a broader simulation of the earth system. Brenner described the goal in 2024 as "building the first foundational model for the natural world... essentially building a machine model that is learning physics."
EPT stands for Earth Physics Transformer. The family is a set of transformer-based foundation models that ingest large volumes of heterogeneous earth-system data, including reanalysis archives, satellite imagery, topography, and surface weather stations, and learn to forecast atmospheric variables. Jua describes the approach as treating the Earth as a single interconnected system rather than solving equations independently for isolated grid cells, and the company states that its models learn approximations of physical conservation laws for mass, momentum, and energy from data. This framing of a learned "physics model" is a marketing characterization, and the degree to which the network enforces physical conservation is not independently established.
Jua has published technical reports on arXiv for successive generations of the model. An EPT-1.5 technical report appeared in October 2024, and an EPT-2 technical report was posted in July 2025. The EPT-2 report lists seventeen authors, all affiliated with Jua, including Roberto Molinaro, Niall Siegenheim, and Niels Poulsen. According to company materials, the broader EPT model is trained on more than 5 petabytes of data drawn from over 120 sources and roughly 10,000 weather stations, a figure the company has cited since the 2024 seed announcement, when it also claimed the training corpus was around twenty times larger than the dataset behind GraphCast. The exact parameter count of the models has not been disclosed in the public technical reports.
EPT-2, released in 2025, is the current generation of the model and is offered as a family of variants tuned for different resolutions, refresh rates, and forecast horizons. Per the EPT-2 technical report, the core model produces native one-hour forecasts out to 240 hours (10 days), at a spatial resolution of 0.083 degrees, roughly 9 by 9 kilometres at the equator, with a deterministic variant and a perturbation-based ensemble variant called EPT-2e for probabilistic forecasting. The technical report describes initialization four times daily at the standard 00, 06, 12, and 18 UTC cycles.
Jua's product documentation lists a wider lineup of EPT-2 configurations, including rapid-refresh and high-resolution members. These vendor-stated specifications include a rapid-refresh "EPT-2 RR" variant updating 24 times per day, a roughly 5 kilometre high-resolution "EPT-2 HRRR" variant, and an "EPT-2e" extended-range configuration reaching a 60-day horizon, alongside later "EPT-2.1" members such as Helios and Europa. The headline variables Jua emphasizes in its benchmarking are 10-metre and 100-metre wind speed, 2-metre air temperature, and surface solar radiation, the quantities most relevant to renewable energy forecasting. The discrepancy between the technical report's four-times-daily core model and the documentation's 24-times-daily rapid-refresh configurations reflects different members of the family rather than a single update cadence.
| Fact | Detail | Source basis |
|---|---|---|
| Headquarters | Zurich, Switzerland (hubs also in Berlin and Cape Town) | Company / press |
| Founded | 2022 | Press |
| Founders | Andreas Brenner and Marvin Gabler | Press |
| CEO | Marvin Gabler (from June 2025; previously Andreas Brenner) | Press |
| Model family | Earth Physics Transformer (EPT); current generation EPT-2 | Company / arXiv |
| Core resolution | 0.083 degrees (~9 km), hourly steps to 240 h (10 days) | EPT-2 technical report |
| Ensemble variant | EPT-2e (probabilistic; extended-range to 60 days per docs) | arXiv / docs |
| Training data | More than 5 petabytes, 120+ sources, ~10,000 stations (company claim) | Company |
| Pre-seed | EUR 2.5 million (USD ~2.5M), October 2022 | Press |
| Seed | $16 million, February 2024 | TechCrunch |
| Series A | $11 million (~EUR 10M), 4 June 2025 | Tech.eu / press |
| Total raised | ~$27 million as of mid-2025 | Press |
Jua's commercial focus is the energy sector, where short- and medium-range forecasts of wind, solar, and temperature directly affect generation planning, grid balancing, and power trading. The company describes its Earth Intelligence Platform as a decision engine that lets energy traders and grid operators simulate weather, supply-and-demand fluctuations, and grid-level dynamics. In an August 2025 interview, Gabler said that for a Belgian customer with a small wind portfolio the company had reduced balancing costs by about 20 percent, and stated that customer-reported reductions in imbalance costs ranged from seven-figure to eight-figure values for larger portfolios. These are company-reported figures from a single named example and are not independently verified.
Insurance and reinsurance, along with agriculture, aviation, and shipping, are cited by Jua and by press coverage as additional or future markets for physical-risk modelling, consistent with the broader applicability of generative AI and simulation to the earth system. As of mid-2025, public reporting named energy trading firms as the primary active customer base, without disclosing specific insurance or reinsurance clients by name.
Jua entered the market in October 2022 with a EUR 2.5 million pre-seed round led by Promus Ventures, with participation from Session.vc and angel investors including Siraj Khaliq, a co-founder of The Climate Corporation and former Atomico partner, and Mehdi Ghissassi, a former product leader at DeepMind. In February 2024 the company raised a $16 million seed round co-led by 468 Capital and the Green Generation Fund, with participation from Promus Ventures, Kadmos Capital, Session.vc, Notion.vc, the founders of FlixMobility, and Innosuisse, among others.
On 4 June 2025, Jua announced an $11 million Series A round, reported as roughly EUR 10 million, co-led by Ananda Impact Ventures and Future Energy Ventures, with continued participation from existing backers including 468 Capital and Promus Ventures. The company said the round brought total funding to about $27 million and would be used to accelerate commercial rollout of the Earth Intelligence Platform in the energy sector. Reported figures of EUR 10 million and $11 million refer to the same Series A round expressed in different currencies.
Jua makes strong performance claims for EPT-2, and these should be read as company claims pending independent confirmation. In its July 2025 EPT-2 technical report and launch announcement, the company stated that EPT-2 achieves lower error than Microsoft's Aurora across the full 0 to 240 hour range on 10-metre wind speed, outperforms Aurora on 2-metre temperature out to about 130 hours, and beats both ECMWF's high-resolution deterministic model (IFS HRES) and its ensemble (ENS) on the variables tested, while running inference roughly 25 percent faster than Aurora and using substantially less compute. In press coverage, the company also said EPT-2 used about 75 percent less computing power than Aurora, which it described as the next most efficient system it tested. At the 2024 seed stage the company had claimed the underlying model used "10,000 times less compute" than legacy numerical systems.
Several caveats apply. The benchmark results were produced by Jua itself and published in a non-peer-reviewed arXiv report rather than by an independent evaluator. The company states that it follows the WeatherBench 2 methodology and validates against ERA5 reanalysis, IFS HRES initial conditions, and the WeatherReal-ISD station dataset, but the comparisons are limited to a small set of surface variables most relevant to energy customers and, by the company's own account, did not include DeepMind's GraphCast in the head-to-head study. Independent verification of AI weather models on the public ECMWF WeatherBench 2 leaderboard has historically been a key check on such claims, and as of this writing Jua's specific EPT-2 comparisons rest on the company's own reporting. Readers should treat the relative rankings against Aurora and ECMWF as vendor benchmark claims rather than settled results.