Written by: Olivier Lam, Physical AI Team, Jua.ai AG | Last updated: July 13, 2026
Key Takeaways for Energy Trading Teams
- DTN Weather Intelligence is a subscription analytics layer on top of traditional NWP, with forecasts typically updating 2–4 times per day.
- Jua for Energy, powered by the EPT-2 foundation model and Athena AI agent, beats ECMWF HRES on every lead time for key energy variables such as wind, temperature, and solar radiation.
- EPT-2 supports up to 24 daily refreshes at far lower inference cost (~$0.20–$15 per simulation) than traditional NWP runs that cost €1,000–€20,000.
- The Jua platform replaces fragmented SaaS workflows with one AI-agent workspace that auto-generates briefings, alerts, and backtests from natural-language queries in about 90 seconds.
- Book a demo with Jua to see EPT-2 benchmarked directly against your current forecast provider.
How DTN Fits Into the Weather Value Chain
DTN, originally short for Data Transmission Network, now operates as a subscription-based weather analytics and intelligence provider for energy, agriculture, and transportation. In practice, DTN Weather Intelligence acts as a processing and distribution layer. It ingests raw NWP outputs from ECMWF, NOAA GFS, and regional models, then turns them into sector-specific dashboards, alerts, and advisory reports.
In January 2026, DTN launched DTN Weather Hub, a platform that combines hyper-local forecasts, asset-level alerts, and configurable dashboards for utilities, energy, aviation, and transportation. This launch reflects growing pressure on legacy SaaS weather vendors to modernize their user interfaces and delivery surfaces.
DTN’s core architecture still sits on top of traditional NWP. The company does not own or operate a proprietary forecasting model. Update frequency follows the underlying NWP cycle, which usually delivers 2–4 global forecast updates per day. Between those runs, numbers remain static. Workflow integration stays manual. Traders download outputs, cross-reference dashboards, and assemble a coherent view of the day before markets open.
The platform offers no transparent, independently verifiable accuracy benchmarks against competing models. It also lacks an AI agent that can turn a natural-language question into a briefing or backtest. For energy traders, meteorologists, and quant teams in 2026, the practical question becomes whether a foundation-model-plus-agent architecture now delivers better accuracy, higher frequency, and tighter workflow integration at lower cost.
How Much Does DTN Cost for Enterprises?
DTN Weather Intelligence pricing is not publicly listed. Enterprise SaaS weather contracts in energy usually follow annual subscriptions with tiers by geography, variable coverage, and API volume. Government procurement data for similar commercial cloud weather platforms shows multi-year contract values that provide a reference range for enterprise-grade services.
The more useful comparison sits at the infrastructure level. A single traditional NWP simulation, which underlies DTN’s outputs, consumes about 8,400 kWh of compute and costs €1,000–€20,000 to run on HPC infrastructure. That cost structure caps update frequency at 2–4 runs per day. The energy industry has lived with this constraint for roughly forty years.
By contrast, a single EPT-2 inference runs on one GPU in minutes, using about 0.25 kWh and costing roughly $0.20–$15 per simulation. EPT-2 trained on 8 × H100 GPUs over 10 days, while Microsoft Aurora used 32 × A100 GPUs over 18 days. The inference cost gap versus traditional NWP is roughly four orders of magnitude. This economics enables up to 24 daily refreshes and turns the DTN update ceiling into a structural limitation rather than a commercial one.
For a 1 GW wind portfolio, a four-percentage-point improvement in forecast accuracy saves about €1.5 million per year in hedging and imbalance costs. A 1 GW solar portfolio at the same accuracy gain saves about €3 million per year. These savings scale linearly across multi-GW portfolios, so the cost gap between EPT-2 and traditional NWP becomes decisive at enterprise scale.
How Reliable Is DTN Compared With EPT-2?
DTN Weather Intelligence inherits the reliability profile of the NWP models it distributes, mainly ECMWF HRES and GFS. ECMWF HRES has set the deterministic NWP benchmark for four decades and remains the standard reference for traders repricing risk around heating demand, renewable output, and system tightness.
In 2026, that benchmark no longer represents the accuracy ceiling.
EPT-2 outperforms ECMWF HRES on every lead time from 0 to 240 hours for 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation. These results use more than 10,000 real ground stations on open-source StationBench, with no post-processing or station fine-tuning, as documented in arXiv:2507.09703. EPT-2 also beats Microsoft Aurora on 10 m wind, 100 m wind, and 2 m temperature across the full 0–240 hour range. Aurora does not produce surface solar radiation output.
EPT-2e, the ensemble version of EPT-2, outperforms the 50-member ECMWF ENS mean on both RMSE (root mean square error) and CRPS (continuous ranked probability score, a probabilistic skill metric) at almost every lead time, according to arXiv:2410.15076. EPT-2e achieves this with only 10 members.
DTN’s reliability remains bounded by the models it resells. Jua for Energy’s reliability is documented in peer-reviewed technical reports, benchmarked against ground truth, and can be verified by any prospect in under 30 seconds on the live benchmarking surface.
How Accurate Is AI Weather Forecasting in 2026?
AI weather forecasting accuracy in 2026 depends heavily on model architecture. Many AI weather models, including Microsoft Aurora and Google DeepMind GraphCast, train on a fixed 6-hour grid and generate forecasts by rolling forward in 6-hour steps. Each autoregressive step adds error, and that accumulation becomes visible at longer lead times.
EPT-2 uses native any-Δt forecasting. The model trains to predict at arbitrary time steps instead of rolling forward in fixed increments. EPT-2 does not roll. This design produces a physics-constrained forecast at any requested lead time without the error build-up that characterizes fixed-step autoregressive systems.
The physics constraint drives the reliability difference. EPT is a spatiotemporal transformer foundation model trained on observational physics. It learns the conservation laws for mass, momentum, and energy directly from data, in a latent representation that integrates forward in time. Outputs respect those conservation laws by construction. This architecture allows traders to trust EPT-2 in workflows where physically inconsistent forecasts create real P&L risk.
In Europe’s weather-driven energy markets, traders now use AI tools not only to predict weather but also to forecast changes in the forecast itself. That use case demands high update frequency and deep ensembles. EPT-2 RR (rapid refresh) delivers up to 24 updates per day, and EPT-2e provides the ensemble depth required for probabilistic positioning.
How Energy Traders Use DTN vs Jua Day to Day
The typical morning routine for an energy trader using DTN Weather Intelligence follows a familiar pattern. Between 7 and 9 a.m., the trader downloads overnight ECMWF and GFS runs, processes them through in-house pipelines, consults an internal meteorology team or external consultancy, and stitches together a view from several dashboards and terminal screens. Markets often move before that view is complete.
Between the 2–4 daily NWP runs, forecasts remain stale. When ECMWF or GFS revises an output mid-cycle, the change arrives silently. The trader notices only after someone else has already traded on it. Model divergence, which represents a trading opportunity, often stays hidden until it appears in prices.
Jua for Energy replaces this fragmented stack with a single workspace. Specific workflow changes include:
- Day-ahead and intraday briefings auto-refresh on every new model run. They cover model consensus across 25+ models, model deltas since the previous run, convergence tracking, and price implications in written form.
- Divergence alerts trigger as soon as two or more models disagree on a key variable. Correction alerts trigger as soon as a model revises its own output. Traders can filter both by zone and PSR (Production Source Resource) type.
- Power forecasts for solar, wind onshore, wind offshore, total wind, total renewables, load, and residual load refresh every 15 minutes for actual generation and extend 20 days ahead on the fundamental model across Germany, Great Britain, France, the Netherlands, and Belgium.
- Athena, the AI agent built for energy trading, converts a natural-language question into a briefing, benchmark, backtest, or custom widget in about 90 seconds. Athena reads market context and models participant behavior to turn raw physics predictions into trading-relevant outputs.
Customers such as Axpo, TotalEnergies, Statkraft, EnBW, EDF, and Hydro-Québec already run daily trading decisions on the platform across five continents.
DTN Weather Intelligence vs Jua: Side-by-Side Comparison
The table below compares DTN Weather Intelligence and Jua for Energy on the dimensions that matter most for energy trading procurement. All Jua figures come from peer-reviewed sources or operational specifications.
| Capability | DTN Weather Intelligence | Jua for Energy (EPT-2 + Athena) |
|---|---|---|
| Deterministic accuracy vs. ECMWF HRES | Resells HRES outputs, with no independent benchmark published | EPT-2 outperforms HRES on every lead time (0–240 h) for 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation |
| Ensemble / probabilistic forecasting | Passes through ECMWF ENS outputs, with no proprietary ensemble | EPT-2e (10 members) beats the 50-member ECMWF ENS mean on RMSE and CRPS at virtually every lead time |
| Update frequency | 2–4 times per day, constrained by NWP cycles | Up to 24 times per day with EPT-2 RR, plus actual-generation power forecasts every 15 minutes |
| Inference cost per simulation | Derived from NWP, about 8,400 kWh and €1,000–€20,000 per run on HPC | About 0.25 kWh and $0.20–$15 per simulation on a single GPU |
| Native any-Δt forecasting | No, uses fixed NWP time steps | Yes, EPT-2 forecasts at arbitrary lead times without rolling 6-hour increments |
| AI agent (natural-language analyst) | No | Athena delivers briefings, benchmarks, backtests, and custom widgets in about 90 seconds per query |
| Spatial resolution (product) | Depends on the underlying NWP source | Up to about 5 km resolution (EPT2-HRRR, Europe) |
| Transparent cross-model benchmarking | Not available | More than 25 models on one platform, any region and variable, with results in under 30 seconds |
Frequently Asked Questions
How DTN Differs From an AI Weather Model
DTN Weather Intelligence operates as a data aggregation and analytics layer rather than a forecasting model. It ingests outputs from NWP systems, mainly ECMWF and NOAA GFS, and repackages them into dashboards, alerts, and advisory products for energy, agriculture, and transportation. DTN does not run its own atmospheric model, does not generate independent forecasts, and does not publish accuracy benchmarks against other providers.
An AI weather model such as EPT-2 belongs to a different category. EPT-2 is a general physics foundation model, a spatiotemporal transformer trained on observational physics that learns conservation laws for mass, momentum, and energy directly from data. It produces its own deterministic and ensemble forecasts, benchmarked against more than 10,000 ground stations, with results published in peer-reviewed reports on arXiv (2507.09703 and 2410.15076). This distinction matters for procurement. DTN’s accuracy ceiling matches the NWP model it resells, while EPT-2’s accuracy is independently verifiable and currently exceeds ECMWF HRES on every lead time across the four variables that drive energy P&L.
How Accurate AI Weather Forecasting Is for Energy Trading
Accuracy in AI weather forecasting for energy trading varies by model design and evaluation method. EPT-2, Jua’s deterministic flagship, outperforms ECMWF HRES, the forty-year benchmark for energy trading, on every lead time from 0 to 240 hours for 10 m wind, 100 m wind, 2 m temperature, and surface solar radiation. EPT-2e, the ensemble variant, beats the 50-member ECMWF ENS mean on both RMSE and CRPS at almost every lead time with only 10 members. Both evaluations use more than 10,000 ground stations on open-source StationBench, without post-processing or station fine-tuning.
For trading, the variables that matter most are wind at hub height (10 m and 100 m), surface solar radiation, and 2 m temperature. These variables drive renewable generation, load, and gas demand. EPT-2 leads on all four. AI forecasting systems have reduced wind and solar generation prediction errors by 30–50% versus traditional methods, enabling higher renewable penetration without harming grid stability and delivering measurable value through lower curtailment and better storage dispatch. These accuracy improvements translate directly to bottom-line savings, and the €1.5M–€3M annual impact described earlier compounds across multi-GW portfolios and multi-year contracts.
Main Alternatives to DTN Weather Intelligence
Energy trading teams evaluating DTN Weather Intelligence alternatives in 2026 typically see three categories. The first category is raw NWP subscriptions such as ECMWF HRES, GFS, and DWD ICON. These provide the model outputs that DTN resells but require in-house pipelines, manual benchmarking, and analyst time to turn into a tradeable view.
The second category is AI weather research outputs such as Microsoft Aurora, Google DeepMind GraphCast, and ECMWF AIFS. These models can improve accuracy over traditional NWP for some variables but arrive as raw files without ensembles, productised refresh schedules, or workflow tooling.
The third category is productised AI platforms such as Jua for Energy. These platforms combine a physics foundation model (EPT-2), a probabilistic ensemble (EPT-2e), up to 24 daily refreshes, and an AI agent (Athena) in one workspace with a REST API and Python SDK.
Productisation separates the second and third categories. Aurora and GraphCast remain research outputs from large AI labs. Jua for Energy operates as a platform where Aurora and GraphCast run as guest models on the same benchmarking surface as EPT-2. The comparison comes built in. Sales cycles at Jua for Energy have shortened to as little as two weeks for trading houses that run the live benchmark and inspect the numbers directly.
Conclusion: Why Energy Desks Move to Foundation Models and Agents
DTN Weather Intelligence still serves as a functional SaaS layer for operators whose workflows depend on processed NWP outputs. In 2026, it no longer defines the frontier for accuracy or frequency. The accuracy advantage documented earlier, EPT-2’s lead over HRES across all lead times and energy-relevant variables, forms the technical base for the platform’s operational edge. EPT-2e’s ensemble performance and the four-order-of-magnitude cost advantage that enables up to 24 daily refreshes make continuous, high-frequency forecasting commercially viable. Athena then turns natural-language questions into briefings, benchmarks, and backtests in about 90 seconds.
Jua operates as a foundation model and agent company, and Jua for Energy is its first applied product. The architecture learns physics, and the domain becomes a variable.
Book a demo to see EPT-2 benchmarked against your current provider on your regions and variables. You can also run live benchmarks directly at athena.jua.ai, with results in under 30 seconds across more than 25 models for any variable you trade.