Legal name: Jua AG
Jua Deep Dive: The Earth Physics Transformer as a Large Physics Model for Energy Trading
Jua (Zurich, 2022) trains the Earth Physics Transformer for weather/energy forecasting, closed an $11M Series A in 2025 and launched the Athena trading agent in 2026, rivaling DeepMind and NVIDIA earth models.
Executive Summary
Jua (Jua.ai AG) is an earth system foundation model company founded in Zurich in 2022, with additional teams in Berlin and Cape Town. Its product has exactly two layers: an atmospheric prediction model called EPT-2, the Earth Physics Transformer, and an energy market agent called Athena built on top of it. The company positions itself as building foundation models for physical reality, with the atmosphere as the first domain, serving energy traders, utilities and hedge funds.
The core specifications of EPT-2 are checkable: native hourly time steps, a 0 to 240 hour horizon, and a spatial resolution of 0.083 degrees, roughly nine kilometres at the equator, with an ensemble variant EPT-2e, a rapid refresh variant updating 24 times per day, and a European high-resolution variant at 5.5 kilometres. The company states that training data exceeds five petabytes across more than 120 sources and about 10,000 ground stations, and that its research is peer-reviewed at ICLR and NeurIPS.
The evaluation methodology is the most notable part. Jua open-sourced StationBench, validates against more than 10,000 real ground stations, applies no post-processing and no station fine-tuning, and claims that EPT-2 beats ECMWF HRES on 10 metre wind, 100 metre wind, 2 metre temperature and surface solar radiation across the full 0 to 240 hour range, with the ensemble variant beating the 50-member ECMWF ENS mean at virtually every lead time. If that holds, the implication is that a model running on a single GPU outperforms a numerical weather prediction system that has run for four decades, on the variables the energy sector cares about.
Commercial disclosures include serving more than 100 gigawatts of energy assets, with named customers including TotalEnergies, Shell, Enel, Statkraft, RWE, EDF, Hydro-Quebec, Adani Energy, Vitol, Origin Energy and ESB. Funding ran from a 2.5 million euro pre-seed in October 2022 led by Promus Ventures, through a 16 million dollar seed in February 2024 co-led by 468 Capital and Green Generation Fund, to an 11 million dollar Series A in June 2025, roughly 10 million euros, co-led by Ananda Impact Ventures and Future Energy Ventures, for about 27 million dollars raised in total. Founders are Andreas Brenner and Marvin Gabler, with Gabler becoming CEO in June 2025.
In physical AI terms the position is clear. Jua implements the learn-physics-from-data route in meteorology, which places it in geophysical physical AI. The value is not solver fidelity but inference speed and refresh frequency: traditional numerical prediction is capped at two to four runs per day by compute cost, while an AI model can produce forecasts hourly or faster. For energy trading, refresh frequency is itself an information advantage.
One framing point matters. The technology narrative is strong and third-party evidence is thin. Every performance claim comes from the company's own technical report and website charts, which the site itself labels illustrative, and no independent body has reproduced the full benchmark. Customer outcome figures, such as a roughly 20 percent reduction in balancing cost for a Belgian customer, are single-case company disclosures. That does not make the claims wrong, but assessment must separate company statements from independently verified results.
Company Overview
Origins and Team
Jua was founded in Zurich in 2022 by Andreas Brenner and Marvin Gabler. Brenner is a serial entrepreneur who served as CEO from founding until June 2025, while Gabler is a researcher in meteorology and machine learning who led the EPT technical direction and took over as CEO when the Series A closed. The timing is worth noting because it coincided with the financing close, which usually signals a shift from technical validation to commercial expansion, with founder roles adjusted accordingly.
Headcount sits in the 51 to 200 range according to third-party sources, with teams in Berlin and Cape Town besides Zurich. Building a model compared against ECMWF, DeepMind and Microsoft at that scale makes talent density the core asset. Industry coverage describes the company competing with Google and DeepMind for AI talent, which cannot be verified, though it reflects a European reality: graduates of the Zurich technical university ecosystem have supported a cohort of earth science and AI startups.
Funding History
The three rounds ran as follows. October 2022 pre-seed of 2.5 million euros led by Promus Ventures with Session.vc and angels including Siraj Khaliq, co-founder of The Climate Corporation, and Mehdi Ghissassi, formerly a product leader at DeepMind. February 2024 seed of 16 million dollars co-led by 468 Capital and Green Generation Fund with Promus Ventures, Kadmos Capital, Session.vc, Notion.vc and Innosuisse participating. June 2025 Series A of 11 million dollars, roughly 10 million euros, co-led by Ananda Impact Ventures and Future Energy Ventures with existing investors following on.
Cumulative funding is about 27 million dollars as of mid-2025. That is tiny for a foundation model company, since training a single large language model costs more. The reason it works lies in the domain: meteorological reanalysis data such as ERA5 is public, highly structured and physically constrained, so model scale need not reach language model levels. This is why weather became one of the first domains where AI for science produced commercial companies.
Customer Base
The disclosed customer list covers major European power and trading institutions: TotalEnergies, Shell, Enel, Statkraft, RWE, EDF, Hydro-Quebec, Adani Energy, Vitol, Origin Energy and ESB. Their nature varies: integrated energy majors such as Shell, TotalEnergies and Enel; hydro and wind dominated utilities such as Statkraft and Hydro-Quebec; a trading house in Vitol; and vertically integrated utilities and retailers such as EDF, ESB and Origin. What they share is that weather directly determines marginal cost or spread, which makes them willing to pay for better forecasts.
The claim of serving more than 100 gigawatts of energy assets is company framing. Against European wind and solar installed capacity, 100 gigawatts is roughly the entire renewable fleet of a mid-sized European country, and if accurate it would indicate the customer base has moved from pilots into production. The company does not disclose how the figure is counted, whether contracted capacity, assets actually connected to the models, or nominal capacity of customer portfolios, which is a diligence question.
Technology
The Cost Structure of Numerical Weather Prediction
Understanding Jua's technical value requires understanding why numerical weather prediction is expensive. It discretizes the atmosphere into a three-dimensional grid and solves momentum, mass and energy conservation equations plus parameterizations for radiation, cloud microphysics, boundary layer turbulence and land surface processes at every grid point. Accuracy comes from the equations themselves, and the price is computation: a global high-resolution model needs a supercomputer, a single run takes hours, and operational frequency is therefore capped at two to four runs per day at 00, 06, 12 and 18 UTC.
That cap is a real loss in energy applications. Diurnal wind cycles, convective cloud formation and front passage timing all shift wind output and power prices within hours. If forecasts refresh only four times a day, a trading desk must extrapolate and rely on experience between updates. The value of an AI model is not only accuracy but speed and frequency: when a model completes a global forecast in minutes on a single GPU, refresh can move to hourly or better.
Jua's claim follows this logic: rather than starting from equations, learn atmospheric evolution from data, encode physical constraints in the architecture, and integrate forward in a latent space faster than the physics itself unfolds.
EPT-2 Architecture and Specifications
EPT-2, the Earth Physics Transformer, is a Transformer-class atmospheric foundation model. Published specifications are: native hourly time steps, a horizon to 240 hours or ten days, spatial resolution of 0.083 degrees, about nine kilometres at the equator, and a core model initialized four times daily at 00, 06, 12 and 18 UTC per the technical report.
Product documentation lists a model family: EPT-2e as the perturbed ensemble for probabilistic forecasting with documented extended-range configurations out to 60 days; EPT-2 RR as the rapid refresh variant updating 24 times daily through satellite data assimilation; EPT2-HRRR as a European regional high-resolution variant at 5.5 kilometres and hourly; and later EPT-2.1 members such as Helios and Europa. The family design reflects a simple engineering truth: no single model satisfies global coverage, high resolution, high refresh rate and long horizon simultaneously, so combinations are required.
One technically notable feature is dynamic lead time conditioning, where a single model generates forecasts at different lead times through conditional inputs rather than training separate models per horizon. This lowers training cost and improves inference flexibility, and it is standard practice among atmospheric foundation models.
StationBench: The Evaluation Methodology as an Asset
The biggest problem in AI weather modeling is incomparable evaluation. Each vendor uses different reanalysis data, variables and weighting and can conclude that it is best. Jua chose to open-source its methodology as StationBench, validating against real ground station observations without post-processing or station fine-tuning, and publishing both data sources and algorithms.
The rigor deserves credit. Using observations rather than reanalysis as truth avoids validating models with models. Avoiding station fine-tuning avoids leakage. Publishing the method allows third-party reproduction. The company claims superiority over ECMWF HRES at every lead time from 0 to 240 hours on 10 metre wind, 100 metre wind, 2 metre temperature and surface solar radiation.
Two caveats apply. First, the website comparison chart is labeled illustrative, with full data in the technical report, which the company publishes itself. Second, the choice of variables is itself a position: wind speed, temperature and solar radiation matter most to energy customers, so superiority there is most persuasive for that audience, but does not extend to all meteorological variables such as precipitation, cloud cover, humidity or the positional error of extreme events. For energy trading the variable selection is reasonable; for the broader claim of the most accurate weather model, it is insufficient.
Physics Transfer: From Atmosphere to Airfoils and Shock Waves
The most interesting part of the narrative is the claim that the same foundation model can be fine-tuned for other physical domains, with airfoils and shock waves cited publicly. The technical logic is that atmospheric equations and compressible flow equations share a mathematical structure as conservation law systems, so latent representations may transfer.
If that path works, the company moves from weather model vendor to physical foundation model vendor, expanding from energy into aerospace, automotive and defense. But public information contains only a one-line capability description, with no paper, no benchmark and no customer case. This is a clear case of narrative running ahead of evidence, and should be treated as option value rather than folded into current valuation.
Athena: What the Agent Actually Does
Athena is the energy market agent built on EPT-2. Per the official description it reads EPT-2 forecasts, the market's own models, prices and user positions, and then indicates what is about to move, outputting briefings, alerts, backtests and power forecasts.
Interaction has two layers. The natural language layer accepts textual objectives and returns briefings, benchmark comparisons, backtests and custom widgets in roughly 90 seconds to five minutes. The programmatic layer exposes more than 25 models through a REST API with Apache Arrow support for large payloads and a Python SDK installed via pip, integrating ENTSO-E European grid data. The design is sensible: quant teams need API and SDK access into existing pipelines, while traders need a natural language surface for fast conclusions.
The boundary deserves clarity: Athena does not trade automatically. It aggregates information and supports judgment, connecting the chain from weather forecast to generation output to price impact to portfolio exposure, while the order remains a human decision. This matches agent positioning in industrial settings, where agents advise and people decide in high-consequence environments.
Technical Shortfalls
First, independent validation is absent. All performance conclusions come from the company's own technical report and website, and no independent body, whether a meteorological service, academic group or industry benchmark, has reproduced the full results. This is common for early-stage companies, but it means the claim of most accurate rests on company endorsement alone.
Second, physics transfer has no evidence. Migration from atmosphere to airfoils and shock waves is narrative rather than result, with no verifiable intermediate output.
Third, extreme event capability is unclear. What energy systems fear is not average wind bias but extreme weather such as storms, cold snaps and prolonged calm. Such events are rare in training data, and extrapolation is a recognized weakness of AI models. No dedicated extreme event evaluation has been published.
Fourth, explainability is limited. Trading and dispatch decisions need to know why a forecast is credible, and latent representations in Transformer models are hard to interpret physically. The physics-constrained architecture mitigates this, but no quantified physical consistency metrics such as conservation violations have been disclosed.
Product and Commercial Model
Product Matrix
The product has three layers. The model layer is the EPT-2 family: deterministic for single-value forecasts, EPT-2e for probabilistic, EPT-2 RR for high-frequency refresh and EPT2-HRRR for European regional resolution. The platform layer is the Earth Intelligence Platform, which combines Jua models with third-party models including ECMWF AIFS, ECMWF IFS, NOAA GFS and Microsoft Aurora in one interface, offering dynamic weather maps, point forecast analysis, power trading analytics in beta and natural language queries. The application layer is Athena, an agent workspace for trading desks.
One pragmatic decision stands out: including competitors' models in the same platform. For customers, comparing multiple forecast sources in one interface beats assembling data pipelines; for Jua, it means retaining customers through accuracy rather than data monopoly. It is a confident approach and it raises switching costs, because customer workflows become embedded in the platform.
Pricing and Delivery
The API uses usage-based pricing with a Python SDK and multiple output formats including JSON, CSV, GRIB and map tiles. Enterprise deals run through sales with no public price and no self-serve signup, which suits a customer base where procurement is long and service levels and data licensing require negotiation.
Usage-based pricing also matches the cost structure of AI models, since inference cost grows with call volume and revenue grows with it. This captures more of the value of heavy usage than flat subscription pricing would.
Quantified Customer Outcomes
The company disclosed one specific case: a Belgian customer with a small wind portfolio reduced balancing costs by about 20 percent after adopting the service, and for larger portfolios customer-reported reductions in imbalance costs ran from seven to eight figures. These figures come from an August 2025 interview, are a single named example and are not independently verified.
The value proposition is logically sound. Balancing costs arise from imbalance settlement caused by forecast error, so better forecast accuracy directly reduces that expense. For a 100 megawatt wind portfolio, if annual balancing costs run into several million euros, a 20 percent reduction is a material saving that easily justifies a software subscription.
Choice of First Market
Jua chose energy rather than insurance, agriculture or aviation because energy has the highest weather sensitivity and an existing trading market that prices information. Power spot and balancing markets settle daily, so forecast accuracy converts to money immediately, whereas insurance pricing cycles run annually and agricultural decisions run seasonally, giving much slower feedback.
The cost of that choice is a market ceiling. Global weather services for energy trading and utilities are finite in size and already served by established vendors such as DTN and Meteomatics. Breaking through requires entering new markets such as insurance, aviation and agriculture, or new physical domains such as airfoils and shock waves. The first is commercial expansion and the second technical, and neither has public progress.
Competitive Landscape
Relationship With ECMWF: Replacement or Complement
ECMWF is the global benchmark for medium-range forecasting, and its HRES deterministic and ENS ensemble products are widely used by national meteorological services and industry. Jua's claim that EPT-2 beats HRES on energy-critical variables positions it as a replacement.
Reality is closer to complement. ECMWF's value is not only forecast products but also the observation assimilation system, model physics and the public-good nature of global meteorological infrastructure. AI models still rely on assimilation for initial conditions, with most initialized from ECMWF IFS analyses, which means AI companies depend on the very institution they claim to surpass. Jua's EPT-2 RR claim of achieving rapid refresh through satellite assimilation indicates movement upstream, but the full capability of an assimilation system cannot be replicated quickly.
The Model Race With Technology Giants
Google DeepMind's GraphCast and GenCast, Microsoft's Aurora, NVIDIA's FourCastNet and Huawei's Pangu-Weather form the first tier of AI weather models. Most come from corporate research labs, are open, publish papers and have abundant compute, and share a common weakness: they are not productized. A model in a paper does not come with service levels, an API or customer support.
Jua's relative advantage sits exactly there, as one of the few companies turning AI weather models into commercial products. The technical gap may be small while the productization gap is large. For a corporate lab, a weather model demonstrates compute and research capability; for Jua it is the entire business, and that difference in incentives usually determines long-term depth of product refinement.
The risk is symmetrical. If Google or Microsoft decides to productize weather models, with Microsoft already offering some weather data services through Azure, Jua's technical edge would be offset by compute and channel advantages. This is the common problem of startups built on a direction that giants research.
Energy Weather Incumbents
In the specific market of energy weather services, competitors include DTN, a long-standing provider of energy and agricultural weather data, Meteomatics with high-resolution weather data APIs, and in-house meteorological teams at trading firms. Their advantage is customer relationships and historical data, and their weakness is model technology. Jua's entry is to win accuracy-sensitive customers with better models and faster refresh, starting at trading desks and expanding into utilities.
Differentiation
Jua's differentiation comes down to three points: transparency in evaluation through open-source StationBench, which is scarce in a market full of best-in-class claims; refresh frequency of 24 times a day against two to four for traditional systems, a structural difference driven by compute economics; and degree of productization through the Athena workspace, API and SDK, which research lab models lack.
The shortfalls are equally clear: small scale with about 27 million dollars raised, missing independent validation, customers concentrated in hedge funds and trading desks that are price sensitive and less sticky, and potential entry by giants.
Financing and Capital Structure
Structure of the Three Rounds
The 2.5 million euro pre-seed in October 2022 led by Promus Ventures was typical technical validation funding for team building and first-generation model training. The 16 million dollar seed in February 2024 co-led by 468 Capital and Green Generation Fund was a sixfold jump, corresponding to validation of the first EPT model and initial customers. The 11 million dollar Series A in June 2025 co-led by Ananda Impact Ventures and Future Energy Ventures brought in two impact and energy transition funds, showing the narrative shifted from AI technology to energy transition infrastructure.
Cumulative funding is about 27 million dollars with valuation undisclosed. At a typical 15 to 20 percent dilution for an 11 million dollar round, post-money valuation would fall roughly between 55 and 73 million dollars, but that is inference and unconfirmed.
What Capital Efficiency Implies
Building a model compared with ECMWF and DeepMind on 27 million dollars is extremely capital efficient. Three reasons explain it: the public availability of meteorological data such as ERA5 reanalysis, the modest scale required because atmospheric degrees of freedom are far below those of language, and a lean team of 50 to 200 people. This efficiency is typical of European deep tech, doing more focused work with less money.
It also means limited resilience. If giants start a price war or offer comparable models free, Jua lacks the capital for a protracted fight. Its path is to build stickiness quickly by embedding in workflows and to build data advantages through proprietary observations and customer feedback loops, shifting competition from model accuracy to product and data.
Revenue and Profitability
Revenue, gross margin and losses are not disclosed. Based on customer scale of 100 gigawatts and more than ten leading customers with typical software pricing, annual revenue might sit in the several million to low tens of millions of dollars range, but that is inference. The 11 million dollar Series A suggests the company is still distant from breakeven, with the round covering 18 to 24 months of operations.
DISCREPANCIES
1. Currency and amount of the Series A. The official framing is 10 million euros, co-led by Ananda and Future Energy Ventures, while third-party databases record 11 million dollars. Both refer to the same round, with the difference coming from currency conversion and timing. This report uses the dual form of 11 million dollars, roughly 10 million euros. Source tier: OFFICIAL (investor announcement) and third-party database in parallel, difference being conversion.
2. Cumulative funding does not equal the sum of rounds. A third-party source cites about 27 million dollars raised, while summing the rounds gives roughly 2.5 million euros of pre-seed, about 2.5 million dollars, plus 16 million plus 11 million, totaling about 29.5 million. The gap may come from convertible instruments, Innosuisse grants not counted, or a third-party snapshot predating the Series A close. This report uses the third-party figure of about 27 million and lists the round details. Source tier: third-party statistic (MEDIA) and official disclosure in parallel.
3. Attribution of founding. Some sources say Marvin Gabler founded the company, others name Andreas Brenner, and others describe Brenner as founder with Gabler as CEO from June 2025. The accurate account is that the two are co-founders, with Brenner serving as CEO from 2022 founding until June 2025 and Gabler leading the EPT technical direction and taking over as CEO at the Series A close. This report uses the co-founder description. Source tier: OFFICIAL (company leadership page) over one-sided MEDIA accounts.
4. Two numbers for model refresh frequency. The technical report describes the core EPT-2 initialized four times daily at 00, 06, 12 and 18 UTC, while product documentation states EPT-2 RR updates 24 times per day. Both are correct and refer to different family members. Some coverage presents the 24 times figure as a capability of EPT-2 itself, creating the misimpression that the core model refreshes hourly. This report separates the core model from the rapid refresh variant. Source tier: OFFICIAL (technical report and product documentation) in parallel, difference being member distinction.
5. Ensemble size. One source describes EPT-2e as a 30-member ensemble, while another states it beats the 50-member ECMWF ENS mean at nearly every lead time. The first is the size of the Jua ensemble and the second is the size of the comparison, not the same thing. This report labels both when ensembles are mentioned. Source tier: OFFICIAL and third-party product documentation.
6. Strength of evidence for beating ECMWF, DeepMind, Microsoft and NVIDIA. The website comparison chart is company-produced and explicitly labeled illustrative, with full tables deferred to the company technical report, which the company publishes and which no independent third party is known to have reproduced in full. This report retains the conclusion with the label of company-reported and not independently verified. Source tier: company technical report, OFFICIAL but self-evaluated, below independent third-party assessment.
7. Third-party misstatement of resolution and horizon. One third-party source describes Jua as providing one kilometre resolution and 20 day forecasts for aviation. The official specification is 0.083 degrees globally, about nine kilometres, with a 240 hour horizon and a European variant at 5.5 kilometres. Neither one kilometre nor 20 days has an official basis. This report uses the official specification. Source tier: OFFICIAL over third-party encyclopedia.
8. Verification status of customer outcome figures. The roughly 20 percent balancing cost reduction for a Belgian small wind portfolio and the seven to eight figure imbalance cost reductions for larger portfolios are single-case company disclosures without stated measurement period, baseline definition or customer verification. This report retains the figures labeled as not independently verified. Source tier: company interview, UNVERIFIED.
Risks and Diligence Gaps
First, missing independent validation is the largest valuation uncertainty. The most accurate claim underpins pricing power and the funding narrative, yet rests on self-evaluation. If a third party such as an academic group, meteorological service or independent benchmark reaches different conclusions, the commercial narrative would take a material hit. Diligence should request the full StationBench evaluation scripts, test set construction and raw scores by model.
Second, extreme weather capability is untested. The genuine risk to energy systems comes from extreme events, and AI weather models extrapolate weakly on sparse samples. No dedicated evaluations of cold snaps, storms or prolonged calm have been published. Such capability matters more to utility dispatch than to short-term trading, and its absence limits deeper penetration into utilities.
Third, customer stickiness and switching costs. Decision support tools at trading desks have relatively low switching costs, since a competitor with better accuracy and lower price requires only a data pipeline change. Jua raises stickiness through platform integration, including competitors' models, and workflow embedding such as position monitoring and backtesting in Athena, but the effect requires renewal rate data that the company does not disclose.
Fourth, risk of giant entry. Microsoft, Google and NVIDIA all build weather foundation models and own cloud channels. If one productizes and bundles with cloud services, Jua's pricing and channels would be squeezed. Its defense is depth in the energy scenario, delivering the power forecasts, balancing cost optimization and ENTSO-E integration that general model vendors will not build.
Fifth, narrative risk in physics transfer. Migration from atmosphere to airfoils and shock waves is the main source of valuation imagination but has no public evidence. If the path does not work over time, the company would be repriced as a weather services company with a lower multiple.
Sixth, data dependency. AI weather models rely on numerical prediction analyses for initial conditions and on public meteorological infrastructure for observations. Changes in upstream data policy, such as an ECMWF shift in data openness or tighter satellite licensing, would affect capability. This is systemic across the AI weather field.
Seventh, European concentration. The customer list is dominated by European energy companies, with Hydro-Quebec in Canada and few United States trading houses. The fragmentation of United States power markets, where each independent system operator has different rules, is an expansion obstacle, and the United States is the world's largest energy trading market.
Composite Assessment
Technology scores 8.0. EPT-2 specifications of 0.083 degrees, hourly steps, 240 hours and a family of variants reach the international first tier, the open-source StationBench methodology is a genuine contribution, the physics-constrained architecture is directionally correct, and transfer from atmosphere to other physical domains has academic grounding. Deductions for missing independent validation, undisclosed extreme event capability and no evidence for physics transfer.
Business scores 7.0. The customer list covers major European energy institutions, service across more than 100 gigawatts is persuasive under company framing, Athena's two-layer interaction design is pragmatic, and usage-based pricing matches the cost structure. Deductions because revenue is undisclosed, the market is concentrated in Europe and trading desk stickiness is unproven.
Finance scores 6.5. Reaching a first-tier model on 27 million dollars is extremely capital efficient, and the three-round investor structure combining technology and energy transition funds is sensible. Deductions for small absolute capital, limited resilience to price competition and no disclosed path to profitability.
Product scores 7.5. The three-layer structure of model family, platform and agent is complete, integrating competitors' models into one interface is a smart stickiness move, and API and SDK engineering is mature with more than 25 models, Arrow support and ENTSO-E integration. Deductions because enterprise features such as permissions, audit and service levels are not described and products for aviation and insurance are absent.
Market scores 7.5. AI weather models are moving from papers into production, energy transition raises weather sensitivity as a long-term tailwind, and the company leads corporate labs in model productization. Deductions for a finite market size, incumbents such as DTN and Meteomatics holding installed customers, and giant entry risk.
Overall 7.3. This is among the most technology-dense and commercially earliest companies in this series: a team of 50 to 200 people produced a model comparable with ECMWF on energy-critical variables and put the credibility question on the table through an open evaluation method. But all performance conclusions remain self-endorsed, customer outcomes are single cases, and physics transfer is narrative. The judgment on physical AI is that the company has demonstrated the feasibility of learning physics from data in the atmospheric domain, and that the path is proven; whether the company itself grows from the most accurate weather model into a physical foundation model platform depends on two things, whether third-party validation supports its claims and whether transfer beyond energy can produce evidence.
References
- Jua website, accessed 2026: EPT-2 and Athena product lines; claim of serving more than 100 gigawatts of energy assets; customer names including TotalEnergies, Shell, Enel, Statkraft, RWE, EDF, Hydro-Quebec, Adani Energy, Vitol, Origin Energy and ESB; StationBench validation against more than 10,000 ground stations with no post-processing; claims of beating ECMWF HRES on 10 metre wind, 100 metre wind, 2 metre temperature and surface solar radiation across 0 to 240 hours; EPT-2e beating the 50-member ECMWF ENS mean at virtually every lead time; aggregate skill chart with EPT-2 at 100, GenCast 84, Aurora 79, FourCastNet 3 at 73 and IFS at 68, labeled illustrative
- Jua website model documentation: EPT-2 as a global hourly model family with deterministic, ensemble, rapid refresh and high-resolution members; refresh up to 24 times daily; ensembles extending to 60 days; power forecasts in megawatts for wind and solar
- Jua article, Jua Tops 2026 Energy Forecasting Software Benchmarks, July 2026: superiority over ECMWF HRES on four variables across all lead times; EPT-2 RR maintaining hourly cadence in production; actual generation power forecasts refreshing every 15 minutes across Germany, Great Britain, France, the Netherlands and Belgium covering solar, onshore wind, offshore wind, total wind, total renewables, load and residual load; Athena natural language layer responding in 90 seconds to five minutes, programmatic layer exposing more than 25 models through REST API with Apache Arrow, Python SDK installable via pip, ENTSO-E integration
- EPT-2 technical report details as summarized by third-party technical encyclopedia: native hourly steps, 240 hour horizon, 0.083 degree resolution or about nine kilometres at the equator; deterministic and perturbation ensemble variants; core model initialized at 00, 06, 12 and 18 UTC; product documentation additionally listing EPT-2 RR at 24 updates daily, EPT2-HRRR at about five kilometres, EPT-2e extended range to 60 days, and EPT-2.1 members Helios and Europa
- Company history and funding as summarized by third-party technical encyclopedia: founded 2022 in Zurich with teams in Berlin and Cape Town; founders Andreas Brenner and Marvin Gabler; October 2022 pre-seed of 2.5 million euros led by Promus Ventures with Session.vc and angels Siraj Khaliq and Mehdi Ghissassi; February 2024 seed of 16 million dollars co-led by 468 Capital and Green Generation Fund; June 2025 Series A of 11 million dollars, roughly 10 million euros; about 27 million dollars cumulative as of mid-2025; training data exceeding five petabytes from more than 120 sources and about 10,000 stations on a company basis
- Ananda Impact Ventures, Why We Invested in Jua: 10 million euro Series A co-led with Future Energy Ventures with 468 Capital, Promus Ventures and others participating; positioning as a purely AI-driven earth intelligence platform; a Foundation Physics Model learning fundamental weather and climate rules from real data; EPT-2 showing significant accuracy gains on energy-relevant variables
- Third-party funding database, SVTR: June 2025 Series A of 11 million dollars with Promus Ventures, 468 Capital, Future Energy Ventures and Ananda Impact Ventures; about 27 million dollars raised in total; founded 2022; description of EPT-2.0 outperforming standard weather forecasting systems and models from DeepMind, NVIDIA and Microsoft in independent evaluations
- Third-party company profile, aiaviationhub: EPT-2 native global hourly forecasts to 240 hours at 0.083 degrees or about nine kilometres; EPT-2e as a 30-member probabilistic ensemble; EPT-2 RR updating 24 times daily through satellite assimilation; EPT2-HRRR at 5.5 kilometres for Europe hourly; training data including satellite imagery, weather stations, topography and ERA5 reanalysis; platform integrating ECMWF AIFS, ECMWF IFS, NOAA GFS, Microsoft Aurora and Google GraphCast; Brenner as CEO from 2022 to June 2025 with Gabler taking over at the Series A close
- Company case figures from an August 2025 interview as summarized by third parties: a Belgian customer with a small wind portfolio reducing balancing costs by about 20 percent; larger portfolios reporting imbalance cost reductions in the seven to eight figure range; company-disclosed single cases without independent verification
- Product capability description in a third-party AI tool directory, neuronfeed: EPT-2 as a foundation model for physical reality that can be fine-tuned for other physical domains such as airfoils and shock waves; Athena accepting objectives together with a world model, simulating consequences, calling tools and resolving tasks; customers including energy traders, utilities, hedge funds and large enterprises; research published at ICLR and NeurIPS
- Competitive and market context from public industry sources: ECMWF HRES and the 50-member ENS as global medium-range benchmarks running two to four times daily; Google DeepMind GraphCast and GenCast, Microsoft Aurora, NVIDIA FourCastNet and Huawei Pangu-Weather as the first tier of AI weather models; DTN and Meteomatics as energy weather data incumbents
- Jua website company page: self-description as a research team in Zurich with peer-reviewed work at NeurIPS and ICLR; positioning that weather is what it models and energy is who it builds for; Jua.ai AG registered in Zurich, Switzerland
This report is compiled from public sources for informational purposes only and is not investment advice. The company is private and most financial and customer data are self-disclosed; performance conclusions follow company technical reports without independent third-party reproduction, and valuation-related inferences are labeled as uncertain.
Related papers
Related reports66
- 51World Deep Dive: A Listed Digital Twin Company Where 51Sim Becomes the First Growth Engine While 51Earth Nearly Vanishes, and Two Irreconcilable Financing Naming Conventions Must Both Be Kept
- AMI Labs: A USD 1.03B Seed Bet on World Models, LeCun's Non-Generative Path and Its Blank Scorecard
- Anshi Pacific (PERA) Deep Dive: A Beijing Domestic-CAE Leader Founded in 2003 Whose PERA SIM General Simulation, AIROM Reduced-Order Models, AgentX Multi-Agent and Jingzhi iGPT/Whale Span from Industrial Simulation to Physics AI
- AspenTech Deep Dive: The Process-Simulation Incumbent Acquired by Emerson and Its Hybrid Industrial-AI Route
- Autodesk Deep Dive: Neural CAD Foundation Models and a $200M World Labs Bet on Physical-World AI, with NAVASTO Adding CFD Surrogate Capability
- Bentley Systems Deep Dive: iTwin Digital Twin and Generative AI Empower Full-Lifecycle Infrastructure Engineering
- BeyondMath Deep Dive: A Foundation Physics AI on the Neural-Operator Route, a $19M Consortium Budget, and Three Misread Facts
- Cadence Deep Dive: System-Level Multiphysics EDA and the Millennium M2000 Supercomputer Anchor Physical AI, with AuraStack Agentic AI and Fidelity GPU CFD as Key Levers
- COMSOL Deep Dive: Four Decades of Multiphysics Solvers, Entering Physical AI via DNN Surrogate-Model Training and All-Physics GPU Acceleration
- Dassault Systèmes Deep Dive: 3DEXPERIENCE and SIMULIA Virtual Twin Physics Behavior Anchor Industrial AI, with LEO Virtual Companion and GPU Explicit Dynamics as Physical-AI Levers
- DP Technology: From the Gordon Bell Prize to National-Level Unicorn Status and the Door of a Hong Kong Listing, the AI-for-Science Leader's Technical Delivery and Commercial Blank
- Emmi AI: From Linz Spin-out to a EUR 330M Mistral Acquisition in 17 Months
- Fainite Deep Dive: A Zurich Physics-AI Simulation Startup Whose Caltech, ETH Zurich, Cambridge and Google Team Uses Physics-Aware Deep Learning to Cut Simulation from Tens of Hours to Seconds, With Only Venture Kick Grants and No Priced Round
- Fysics AI: The PhysX Founder's Second Act, a Domestic Differentiable Physics Engine and Its Ecosystem Play
- Frontop (301313) Deep Dive: From Digital Exhibition Halls to a Physical-AI Data Foundation - FTE Engine, FunSim Simulation and the Gap Between Positioning and Revenue
- Godela Deep Dive: A YC X25 Physics-AI Simulation Startup Whose Ex-Apple, Google and Intel Team Uses Physics-Constrained Surrogate Models to Cut Simulation from Days to Seconds, at Seed Stage with Pilot-Only Access
- HollyWave (688682) Deep Dive: RDSim AI Electromagnetic Simulation and Engineering Agents Close the Simulation-Measurement-Product Loop
- IAT (Alt) Deep Dive: China's First Listed Independent Automotive Design Company Pivots to AI-native R&D with the Yuanfang, Yufeng and Taiyi 3.0 Product System and an NVIDIA Omniverse Partnership, Reporting RMB 975M Revenue and a Widened RMB 250M Net Loss in 2025
- Inductiva Deep Dive: A Closed Case in Open-Solver Orchestration, Eight Premises That Must Be Corrected, and Where Co-Founder and CTO Hugo Penedones Went Next
- Intesim (INTESIM) Deep Dive: A Home-grown Chinese CAE Leader Founded in 2009 with Self-developed Solvers for Structure, Fluid, Electromagnetics, Thermal and Acoustics and Multiphysics Coupling, a ~RMB 100M B Round from CNII and Shenzhen Capital, and 2025 Revenue and Bookings Both above RMB 100M
- iSoftStone (301236) Deep Dive: A ChiNext-Listed IT Services Giant Whose TianShu Platform Builds a Three-Layer Physical-AI Stack with Self-Developed TwinHub Digital Twin, FSim Industrial Simulation, CAE Multiphysics and Phyxis Embodied Control
- Keysight (with ESI Group) Deep Dive: Virtual Prototyping and Hybrid AI Bridge Measurement and Simulation, with Natural-Language Assistants in EDA
- Lightwheel Deep Dive: A Physical AI Data and Evaluation Infrastructure Startup Founded by a Former NVIDIA Simulation Lead, Seven Rounds Totalling Roughly CNY 2.65 Billion at a Valuation Above USD 2 Billion, Seats on Both the Newton and EgoVerse International Standards Committees, With Two Incompatible Accounts of Q1 Orders
- Luminary Cloud: The Physics AI Model Factory Backed by NVentures
- MathWorks Deep Dive: MATLAB/Simulink and Simscape Multiphysics Power Physical AI, with R2026a Copilots and an Agentic Toolkit for Edge Intelligence
- Maniformer (Mifeng) Deep Dive: An Agibot-Spun Physical-AI Data Infrastructure Firm Whose Bodyless MEgo Capture Hardware Is in Mass Production and Which Closed Three Rounds in Half a Year Led by China Telecom
- Mistral AI: How Physics AI Moved From Language Models to EUV Lithography and Nuclear Power
- Monolith AI: The British Engineering-AI Firm That Cut Physical Test Counts Instead of Accelerating Simulation, Now Inside CoreWeave
- Monumo Deep Dive: A Cambridge Deeptech Company Redesigning Electric Motors with the Anser AI Engine, Tens of Millions of Multiphysics Simulations per Day, GBP 10.5M Raised in Total, and a Series A Being Raised
- NAVASTO: Moving Solver Insight Onto the Designer's Desk, and How Autodesk Bought That Entry Point
- Navier: Handing CFD Workflows to Agents, a Seed-Stage Bet on AI for Hardware Engineering
- Neural Concept: An EPFL Spin-out Betting on Neural Surrogates, with $138M Raised
- NP Co. (NP Company) Deep Dive: A 2025-founded Paris Inria-born AI-native simulation startup that raised a EUR 6M pre-seed led by Partech with Mistral AI co-founders as angels, claiming a pre-trained physics foundation model at 1,000x speedup — most claims unverified
- NVIDIA Deep Dive: Building the Full-Stack Physical-AI Infrastructure with PhysicsNeMo, Warp, Cosmos World Foundation Models and Omniverse
- Physical Intelligence Deep Dive: A $2.1B Bet on the GPT Moment for Robotics, With pi0.7 Self-Reporting 60-80 Percent Success on Unseen Tasks and Neither Pricing nor Revenue Disclosed for the Pi API
- Physical Superintelligence Deep Dive: A $58M Seed Bet on Virtual Physicists, With Emmy Targeting AI Data Center Multiphysics Optimization as Its First Commercial Proof Point
- PhysicsX: Pretrained Geometry and Physics for Industrial Engineering at a $2.4B Mark
- Prometheus: A $41B Mark, $18.2B Raised, and Not Yet a Product
- PTC Deep Dive: Creo/Onshape and the ThingWorx Digital Thread Power Physical AI, with NVIDIA Isaac Sim Bridging Robotics Design to Simulation
- Qianjue Technology Deep Dive: How a Body-Agnostic Embodied Brain Pairing the Predictive World Model Polibrain with Brain-Inspired Partitioning and Polynomial Representation Reached a Hundred Thousand Cleaning and Service Robots
- Quanscient Deep Dive: Cloud-Native and Quantum-Powered Multiphysics Simulation as the Data Engine for AI-Native Hardware Engineering
- Rescale Deep Dive: An HPC Operating System and Desktop Engineering Workbench in the Cloud, a $115M Series D, and Five Misread Facts
- Schrödinger Deep Dive: Physics-Based FEP+ Simulation and Machine Learning Power Molecular Discovery, with the 2026 Bunsen Agentic Co-Scientist
- Secondmind Deep Dive: A Cambridge Engineering-AI Firm Whose Bayesian Active Learning Cuts CAE Simulation and Test-Bench Runs by up to 90%, with Roughly USD 73M Raised and Mazda Both Customer and Investor
- Shichuang Technology Deep Dive: SupreCAST Cloud-Native Material-Processing Simulation and Industrial Agents Advance the CAE+AI Strategy
- ShonCloud (Shunyun) Deep Dive: Meshless Particle-Method CFD for EV and Nuclear Fluid Multiphysics
- Shuzhi Maijin Deep Dive: A Shenzhen Fully Self-Developed SNN-plus-LBM CFD Startup Claiming Second-Level Results on Hundreds-of-Millions of Grids on Consumer Hardware, With a Tens-of-Millions-RMB Angel Round From a Ministry-of-Finance-Affiliated Fund and an FPGA Migration Plan
- Siemens (incl. Altair) Deep Dive: The 2025 Altair Acquisition Unites Simcenter and Altair, with PhysicsAI Geometric Deep Learning and PhysicsAI Generate as the Physical-AI Core
- Silurian Deep Dive: A Generative Forecasting Transformer Earth-System Foundation Model for Weather Risk on Energy and Infrastructure Assets
- SimScale Deep Dive: A Cloud CAE Platform Built on Core-Hour Subscriptions, EUR 52M of Equity Kept Separate from EUR 25M of Debt, and Five Misread Facts
- Skild AI Deep Dive: A General Robot Foundation Model Incubated at CMU's Robotics Institute, With a USD 1.4B Series C at a USD 14B Valuation, a Conflicting Series B Size, and an ARR Figure That Is a Run-Rate Rather Than Revenue
- Songying Deep Dive: A Huawei-Cloud-Alumni Startup Building a Domestic Physical AI Operating System, With ORCA OS Positioning Against NVIDIA Omniverse, Three Incompatible Financing Narratives, and a Verifiable Third-Party Anchor in a CNY 20 Million Minority Investment
- Sudo Tech Deep Dive: Zero-Real-Data Simulation Training, the Fully In-House Sudo R1 Stack, and an Atomic-Skill Commercialisation Strategy from a Startup Valued Above CNY 20 Billion in 17 Months
- Suochen Technology: The Pain of Becoming a Physics-AI Platform from a CAE Vendor
- Synopsys (Ansys) Deep Dive: The 2025 Ansys Acquisition Turns the EDA Leader into a Silicon-to-Systems Engineering Powerhouse, with Ansys SimAI and the NVIDIA Physical-AI Partnership as the Core Physics-AI Levers
- TenFong Deep Dive: An Academician-Led Chinese Industrial Simulation Software Vendor Whose Official Timeline and Database Records Use Incompatible Round Names, With an Official Financing PDF and an Investment Database Both on File
- Tianfu (Nanjing Tianfu Software) Deep Dive: A 2011-founded Chinese Intelligent-CAE Veteran That Closed a Several-Hundred-Million-RMB Series C, Pivoted to an Industrial World Model in 2024, With Unreconcilable Financing and Valuation Records and No Disclosed Revenue
- Tongyuan Soft Control Deep Dive: MWORKS (Syslab/Sysplorer) as a Domestic MATLAB/Simulink Alternative, with MWORKS AI and the MOGICK Engineering Agent Advancing Home-Grown Physical AI
- Vinci AI: Turning Physics Simulation from a Checkpoint into a Continuous Constraint
- Westlake Robotics Deep Dive: Westlake University's First Robotics Spin-out Builds a Dual-Pretrain Brain-and-Cerebellum Architecture with WR1 Whole-Body Foundation Model and GAE Behaviour Expert
- World Labs: Fei-Fei Li's Spatial-Intelligence Unicorn, Acquired by AMD for $8.2B
- XVERSE Deep Dive: A Simulation-World Route to Physical AI Built on the XVERSE Model Family, an Edge-Cloud 3D Engine and the Physics-Based MotionGen Model
- SimWE: From In-House Solvers to a Physical-AI Engine, the Domestic CAE Leader and What Comes Next
- Yunjing Zhifang Deep Dive: A Peking-University-Incubated Cloud-Native GPU CAE Company with the Simversus Platform and the YJ-DORGP Small-Sample Intelligent Solver, Reporting 275 Percent Order Growth, 160 Percent Revenue Growth and Breakeven in 2025
- Zhongke Diwiji (Five Ages) Deep Dive: A CAS-Institute-of-Automation and Tsinghua Spin-out Whose FAM Embodied Foundation Model Uses Action Optical Flow for Cross-Embodiment Generalisation, Reaching 97% Success from 3 to 5 Demos, with EraBrain and Jiling Robots Serving Sinopec, Leapmotor and Midea
- ZWSOFT (688083) Deep Dive: Overdrive Kernel and ZW3D Wukong 2027 Underpin an Integrated CAD/CAE/CAM Stack, with a CAx Agent Platform as the Practical Route to Engineering AI
This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.