Jua AI

Jua AI: A Foundation Model for Physical Reality

Jua AI is an industrial intelligence company building a foundation model for the physical world, starting with atmospheric prediction and extending to broader physical systems. Unlike language models trained on text, Jua’s approach centers on learning the governing physics of reality from observational data—particularly the complex, chaotic dynamics of Earth’s atmosphere—to enable accurate, long-term forecasting and decision-support across high-stakes industries.

The core innovation is EPT-2, Jua’s atmospheric foundation model, which the company states achieves state-of-the-art performance on weather prediction benchmarks. According to Jua’s published benchmark comparisons, EPT-2 scores 100 on normalized aggregate skill (combining RMSE, ACC, and CRPS) on a 2024 held-out test set, outperforming incumbent models such as ECMWF’s IFS (68), NVIDIA’s FourCastNet (73), Microsoft’s Aurora (79), and Google DeepMind’s GenCast (84). This performance is highlighted as evidence that a model trained on atmospheric physics can generalize to other physical domains.

Built atop EPT-2 is Athena, Jua’s AI agent designed to resolve objectives inside physical reality. Given a goal, a world model, and a set of tools, Athena simulates consequences, invokes tools, and iteratively improves both its actions and the underlying model. This compounding loop enables the agent to become more capable as the world model improves, and vice versa. Athena is already in production use among energy traders, utilities, and hedge funds for three primary objectives: minimizing atmospheric forecast error (to beat ECMWF in energy and shipping markets), generating alpha in prediction markets by pricing contracts more accurately than the market, and accelerating internal research by automating experiment setup, data retrieval, and result analysis on Jua’s GPU cluster—while keeping humans in the loop.

Jua emphasizes that its work began with the atmosphere because it represents one of the most challenging continuous-physics datasets ever recorded: a high-dimensional, partially observed, chaotic fluid system monitored at kilometer resolution for decades. The company argues that if a model can learn atmospheric physics from data, it can learn the physics of almost any governed system—opening pathways to applications in turbomachinery, thermal systems, materials science, and beyond.

The company targets industrial clients in sectors where weather and physical dynamics directly impact operations and profitability. For renewable energy, Jua’s forecasts aim to improve wind and solar energy production prospects. In agriculture, the technology supports crop forecasting and risk mitigation for drought and flooding. Insurance clients use it to better assess and reduce losses from extreme weather. Logistics providers leverage it to optimize supply chains and avoid weather-related delays. Jua notes that its models are used by over 100 GW of energy capacity worldwide, citing clients such as TotalEnergies, Shell, Enel, Statkraft, RWE, EDF, Hydro-Québec, Adani Energy, Vitol, Origin Energy, and ESB—though it does not specify the nature or scale of each engagement.

Jua’s team combines expertise in artificial intelligence, physics, and data science, with deep experience in weather systems. The company states it has raised substantial funding from investors who believe in its approach to physical AI, though it does not disclose funding amounts, investor names, or valuation details. All research is peer-reviewed, with work presented at ICLR and NeurIPS.

While Jua’s technology shows promise in benchmark-driven atmospheric forecasting and agent-based decision support, it is primarily positioned for large-scale industrial and financial users with access to complex data pipelines and domain-specific objectives. Smaller organizations, individual developers, or those seeking general-purpose AI tools may find the platform less accessible due to its enterprise focus, lack of public APIs or self-serve options in the provided materials, and emphasis on custom deployment via demo and sales engagement. The company does not advertise a free tier or self-serve pricing in the sourced text.

For more information, visit the official website: Jua AI.