Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: June 24, 2026
Key Takeaways for Energy Traders
- EPT-2 outperforms ECMWF HRES on all four energy-critical variables (10 m wind, 100 m wind, 2 m temperature, SSRD) across the full 0–240 hour lead-time range.
- EPT-2e beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time, delivering superior probabilistic skill for energy trading.
- EPT-2 RR updates up to 24 times per day at a fraction of traditional NWP cost (~0.25 kWh and $0.20–$15 per run on a single GPU).
- Native any-Δt forecasting and physics-constrained outputs give EPT-2 an edge over research models like Aurora and GraphCast that roll forward in fixed 6-hour steps.
- Book a demo with Jua to run live benchmarks against your current forecast provider and see the accuracy gains for yourself: Verify these accuracy claims on your own data in under 30 seconds.
2026 AI Weather Model Ranking for Production Use
The table below ranks seven models on the four dimensions that determine production fitness for energy trading. The key finding is straightforward: only EPT-2 and EPT-2e combine superior accuracy on all energy-critical variables with an operational update schedule and GPU-level inference costs. The other AI models remain research outputs without productised ensembles or firm operational commitments.
Deterministic accuracy is expressed as performance relative to ECMWF HRES (the 40-year gold standard) on the four energy-critical variables across the full 0–240 hour lead-time range. Ensemble skill is expressed relative to the 50-member ECMWF ENS mean on RMSE and CRPS. All EPT-2 and EPT-2e figures are anchored to arXiv:2507.09703 and arXiv:2410.15076.
| Model | Det. accuracy vs HRES (10 m wind, 100 m wind, 2 m temp, SSRD) | Ensemble skill vs ECMWF ENS mean (RMSE / CRPS) | Update frequency | Inference cost per run |
|---|---|---|---|---|
| EPT-2 / EPT-2e (Jua) | Beats HRES on all four variables across full 0–240 h range | EPT-2e beats 50-member ENS mean on RMSE and CRPS at virtually every lead time | Up to 24×/day (EPT-2 RR); EPT-2e 4×/day | ~$0.20–$15 (~0.25 kWh) on a single GPU |
| Microsoft Aurora | Loses to EPT-2 on 10 m wind, 100 m wind across full range; loses on 2 m temp up to ~130 h; no SSRD output | No productised ensemble | Typically 4×/day (research cadence; no published operational schedule) | Similar order of magnitude to EPT-2, around 25% slower inference than EPT-2 |
| GFS GraphCast (Google DeepMind) | EPT-1.5 outperforms GraphCast on European wind and temperature | No productised ensemble | Typically 4×/day (research cadence) | Similar GPU-class inference cost |
| ECMWF AIFS | Competitive with HRES on standard variables; no published SSRD head-to-head vs EPT-2 | AIFS-ENS available; member count and CRPS vs ENS not publicly benchmarked against EPT-2e | 4×/day (operational) | GPU-class; exact cost not publicly disclosed |
| ECMWF HRES | The benchmark itself, 40 years of NWP leadership at 9 km resolution | N/A (deterministic model) | 4×/day | ~€1,000–€20,000 (~8,400 kWh) on HPC |
| ECMWF ENS | N/A (probabilistic model) | Gold standard, 50 members; EPT-2e beats its mean on RMSE and CRPS at virtually every lead time | 4×/day | ~€1,000–€20,000 (~8,400 kWh) on HPC |
| NOAA GFS | Below HRES on energy variables; free public baseline | GFS Ensemble Mean available; below ENS skill | 4×/day | Public HPC; no per-run cost to end users |
Run these benchmarks on your own region and variables. Book a demo.
Model Cards: How Each System Performs in Practice
EPT-2 and EPT-2e (Jua)
EPT-2 outperforms ECMWF HRES on every lead time and on all four energy-critical variables — 10 m wind, 100 m wind, 2 m temperature, and SSRD — across the full 0–240 hour range. As documented in arXiv:2507.09703, EPT-2 maintains this accuracy advantage across the entire horizon.
EPT-2e, the ensemble variant, keeps the probabilistic edge described in the key findings and delivers higher skill than the ENS mean across virtually all lead times. Native any-Δt forecasting means EPT-2 predicts at arbitrary time steps rather than rolling forward in fixed 6-hour increments. Error does not compound in the same way as in stepwise models.
EPT-2 RR keeps an hourly refresh cadence referenced earlier, while EPT-2e updates 4 times per day to match traditional operational schedules. EPT-2 delivers hourly global weather updates and outperforms leading AI weather models and traditional numerical baselines across all forecast horizons on RMSE.
Jua’s models can natively forecast up to a 5 km resolution, with up to 1 km resolution available as a product. A single inference run uses a fraction of a kilowatt-hour on commodity GPU hardware. Both models are benchmarked against more than 10,000 real ground stations on open-source StationBench, with no post-processing or station fine-tuning, and results are published in peer-reviewed technical reports on arXiv (2507.09703, 2410.15076).
Microsoft Aurora
Aurora is a large-scale AI weather model from Microsoft Research. Aurora loses to EPT-2 on 10 m wind and 100 m wind across the full 0–240 hour range, and on 2 m temperature up to approximately 130 hours. Aurora produces no SSRD output, which creates a direct gap for solar-generation forecasting.
The model rolls forward in fixed 6-hour steps, so error compounds at longer lead times. No productised ensemble equivalent is available. Aurora is a research output from Microsoft’s AI lab, not a productised operational platform. It runs as a guest model on the Jua platform, available for head-to-head comparison alongside EPT-2.
GFS GraphCast (Google DeepMind)
GraphCast is a graph neural network weather model from Google DeepMind, initialised with NOAA GFS data. EPT-1.5 outperforms GraphCast on European wind and temperature. GraphCast operates at approximately 25 km published resolution with a fixed 6-hour time step.
No productised ensemble is available. Like Aurora, GraphCast is a research output consumed as raw model files. It runs as a guest model on the Jua platform for direct comparison.
ECMWF AIFS
AIFS is ECMWF’s AI-based forecasting system, developed in parallel with the institution’s flagship NWP products. It runs at operational cadence of 4 updates per day and benefits from ECMWF’s data assimilation infrastructure.
AIFS runs natively on the Jua platform alongside EPT-2 and HRES, which enables direct comparison. An ensemble variant, AIFS-ENS, is available, though published head-to-head benchmarks against EPT-2e on RMSE and CRPS have not yet appeared in the open literature.
ECMWF HRES and ENS
ECMWF HRES is the deterministic flagship of the European Centre for Medium-Range Weather Forecasts and serves as the universal benchmark for 40 years of operational NWP at 9 km resolution. ECMWF ENS is the 50-member operational ensemble and the gold standard for probabilistic NWP.
Both run 4 times per day on HPC infrastructure at a cost of approximately €1,000–€20,000 and around 8,400 kWh per simulation. Jua for Energy does not replace ECMWF. Serious customers keep their ECMWF subscription and run Jua for Energy alongside it.
Jua for Energy instead replaces the plumbing around the incumbent feed: the in-house grib pipeline, the manual benchmarking, and the morning-briefing routine.
Compare these models side-by-side on the Jua platform. Book a demo.
Physics Foundation Models Versus Research AI Output
The trustworthiness gap between EPT-2 and earlier research AI models starts at the architecture level. Standard transformers applied naively to atmospheric data can produce outputs that violate conservation laws, such as mass, momentum, and energy, because nothing in the training objective enforces physical consistency.
EPT, the Earth Physics Transformer, is a general spatiotemporal transformer foundation model that learns the governing physics of complex systems directly from observational data. It does this in a latent representation that is integrated forward in time. The conservation constraints live inside that representation, not as a post-hoc patch.
The practical consequence is that EPT outputs are physically constrained by construction. An LLM is unconstrained on the symbolic surface, while a physics foundation model is constrained at the representation. That architectural difference is why meteorologists at regulated utilities, who must defend every forecast to internal risk and regulatory stakeholders, treat EPT-2 differently from research AI outputs.
The trust is backed by external validation. EPT-2 is benchmarked against more than 10,000 real ground stations on open-source StationBench, with no post-processing or station fine-tuning.
Aurora and GraphCast are research outputs from large companies’ AI labs. They do not function as foundation models with agents on top of them. Jua’s category, a general physics foundation model (EPT) paired with an AI agent (Athena), with Jua for Energy as the first applied product, sits one level above the research-output category. The relationship mirrors Anthropic and Claude Code: a horizontal AI platform with a flagship vertical product.
Operational Metrics That Matter for Production Use
Four operational metrics determine whether an AI weather model is usable in a live trading environment rather than only in a research setting.
Update frequency. Traditional NWP delivers 4 global forecasts per 24 hours, a hard constraint imposed by HPC economics. That schedule leaves traders working with stale numbers between runs, which matters when intraday price swings unfold in minutes. EPT-2 RR updates up to 24 times per day. EPT-2e maintains a 4 times per day cadence that aligns with existing operational workflows.
That higher refresh rate only works because of the second metric, inference cost. Inference cost. A single EPT-2 inference runs at ~0.25 kWh and $0.20–$15 on a single GPU, in minutes. A single NWP simulation consumes around 8,400 kWh and costs €1,000–€20,000 on HPC. This cost gap, roughly four orders of magnitude, is what makes frequent updates economically realistic.
Dissemination time. A typical Jua run completes approximately 2.5 hours ahead of competing operational runs at the same cycle. In a market where milliseconds and gigawatts drive profit, dissemination time behaves as a direct P&L variable.
Native any-Δt forecasting. EPT-2 is trained to predict at arbitrary time steps. Aurora and most peers roll forward in fixed 6-hour increments and compound error at each step. EPT-2 does not roll in that way. For intraday energy trading, where the relevant horizon is hours rather than days, this difference is material.
A 1 GW wind portfolio that gains four percentage points of forecast accuracy saves around €1.5 million per year. That figure scales linearly across multi-GW portfolios.
Frequently Asked Questions
What makes an AI weather model “production-ready” for energy trading in 2026?
Four criteria separate production-ready models from research outputs. First, they need verified deterministic accuracy on energy-relevant variables, specifically 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation, benchmarked against ECMWF HRES across the full 0–240 hour lead-time range using ground-truth observations rather than reanalysis.
Second, they must ship a productised ensemble with published probabilistic skill scores, RMSE and CRPS, relative to the ECMWF ENS mean. Third, they require an operational update frequency suitable for intraday trading, at minimum 4 times per day and ideally higher. Fourth, they need an inference cost and dissemination architecture that supports those update rates without HPC infrastructure.
EPT-2 and EPT-2e, delivered through Jua for Energy, meet all four criteria. Aurora and GraphCast do not meet the last three as productised offerings.
How does EPT-2e compare to ECMWF ENS for probabilistic energy forecasting?
EPT-2e is Jua’s ensemble variant of the EPT-2 foundation model. It beats the 50-member ECMWF ENS mean on both RMSE and CRPS at virtually every lead time, as documented in the peer-reviewed technical report arXiv:2507.09703. EPT-2e updates 4 times per day.
For energy traders who need probabilistic forecasts to position around generation uncertainty, such as wind ramps, solar dips, and temperature-driven demand spikes, EPT-2e provides ensemble depth that no other AI weather model currently ships as a productised offering. ECMWF ENS remains the gold standard for probabilistic NWP and is available on the Jua platform alongside EPT-2e for direct comparison.
Can I integrate Jua for Energy forecasts into my own trading models and pipelines?
Yes. Jua for Energy exposes a REST API with Apache Arrow support for large payloads and a Python SDK installable via pip install jua. The API covers more than 25 models, including 10 proprietary AI models from the EPT family plus 15 third-party NWP and AI models such as ECMWF HRES, ENS, AIFS, NOAA GFS, DWD ICON, Aurora, and GraphCast, under a unified schema.
Hindcast data is available across multiple Jua and third-party models for backtesting. A typical integration that takes a quant team a quarter to build elsewhere stands up in days. ENTSO-E grid data is integrated directly for European power-market variables. Documentation is at docs.jua.ai and the developer dashboard at developer.jua.ai.
Is Jua a weather AI company?
No. Jua is a foundation model and agent company. EPT is a general physics foundation model, domain-agnostic by architecture and currently fine-tuned for atmospheric prediction. Athena is an AI agent, currently instrumented with the Jua for Energy tool surface. Jua for Energy is the first applied product.
The relationship mirrors Anthropic and Claude Code, a horizontal AI platform with a flagship vertical product. The atmosphere is the first physical system EPT has been fine-tuned for, and energy trading is the first market Athena has been instrumented for. Both will expand to other physical-economy domains such as plasma fusion, aerospace, and materials.
How quickly can I verify EPT-2 accuracy against my current forecast provider?
The live benchmark on the Jua platform returns a head-to-head accuracy comparison in under 30 seconds. A prospect selects a region and variable that matters to their book, selects their current provider alongside EPT-2, and the platform returns the comparison on the spot.
Backtests against years of historical forecasts run in approximately 5 minutes via Athena. This speed usually acts as the deal trigger for Jua for Energy customers. Meteorologists who were sceptical of vendor accuracy claims often become internal champions once they run the benchmark themselves.
Conclusion: How to Act on These Benchmarks
The 2026 AI weather model landscape has a clear production leader on energy-relevant variables. The accuracy advantage detailed earlier, EPT-2’s superiority on all four energy-critical variables, holds across the full forecast horizon. EPT-2e’s probabilistic advantage over ENS, described above, completes the accuracy picture.
EPT-2 RR maintains the high-frequency update pattern at low GPU cost, while Aurora and GraphCast remain research outputs without productised ensembles, operational refresh schedules, or the agent layer that turns forecast data into a tradeable briefing in 90 seconds.
Jua is a foundation model and agent company. Jua for Energy is the first applied product, used by Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec, and by quant funds across five continents who pipe Jua into their own models via pip install jua. The evaluation criteria in this article match the ones those customers applied.
The live benchmark remains the fastest way to validate these claims on your own region and variable. Verify EPT-2’s accuracy on your portfolio in under 30 seconds. Book a demo.