Enterprise Knowledge Search · Real-time Analytics
EPRI needed to run advanced AI for sensitive utility documents and weather-planning workflows locally, because energy-sector research data could not leave controlled environments and shared compute delays limited how quickly researchers and planners could test new scenarios.
Reported impact
EPRI can now run a 120B-class local document assistant at 35 tokens per second for one user or 150 aggregate tokens per second for ten concurrent users, while DGX Station compresses a 50-member weather ensemble from an estimated 10+ CPU hours to 150 seconds, giving grid planners more time to act without sacrificing air-gapped deployment requirements.
EPRI built two local AI workflows on NVIDIA deskside systems so energy researchers could keep performance, sovereignty, and compliance without relying on external networks. Power Chat runs on DGX Spark and Dell Pro Max with GB10 to host a 120B-class open-weight model, a one-million-token document cache, and inference on a single local system. It serves technical documents, logs, and reports through a natural-language interface using a NVIDIA vLLM NGC container with prefix caching, all inside air-gapped or connectivity-limited environments. For planning and grid operations, EPRI uses DGX Station with AI weather models such as GenCast, FourCastNet, CorrDiff, and CBottle to downscale weather data, generate high-resolution ensemble forecasts, and give grid planners more time to evaluate scenarios. Across both workflows, coherent unified memory, local DGX software, and on-prem deployment let EPRI prototype, fine-tune, and operate energy AI without exposing regulated data or waiting on shared infrastructure queues.
Technical documents and logs are loaded into Power Chat’s local cache on DGX Spark for natural-language retrieval, while weather and climate grids feed DGX Station models such as GenCast, FourCastNet, CorrDiff, and CBottle to produce ensemble forecasts and downscaled outputs for grid operators and planners.
Both workflows run entirely on premises and are air-gap capable, with no external network dependency.
NVIDIA DGX Station GB300 generates a full 50-member GenCast ensemble weather forecasts in 150 seconds, compared to an estimated 10+ hours on a CPU server.
Power Chat runs a 120-billion-parameter open-weight LLM entirely on one deskside system, streaming 35 tokens/second to a single researcher and 150 tokens/second aggregate across 10 concurrent users.
The result is a private, cost-effective AI assistant that keeps a one-million-token KV cache for document recall across sessions.
Users select technical relevant documents that are loaded into the model's memory via cache-augmented generation.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
Databricks Sample Sep 17, 2025 Williams Historically, business intelligence was centralized: analysts used tabular perspectives and legacy reporting tools, which became increasingly obsolete as data volumes and demands surged. “We would get requests from… 100+ Employees empowered with near-instant, self-serve analytics Genie · Microsoft Teams · Unity Catalog Databricks Sample Jul 30, 2025 Axpo Group But legacy challenges — including data silos, fragmented tools and manual processes — slowed progress. 30% Time savings for power users Agent Bricks 
Microsoft Sample Dec 27, 2024 Aydem Energy Seasonal challenges, like hot summers or high number of travelers during vacation periods, lead to increased demand while windy winters cause disruptions. 90% of inquiries are fully resolved, boosting agent Azure OpenAI Service · Power Platform 
