Energy
EPRI Powers Energy Research With Local AI
Objective
EPRI is a leading independent, nonprofit energy R&D organization, collaborating with more than 450 companies across 45 countries since 1972 to ensure clean, safe, reliable, and affordable access to electricity across the globe. As electrification, industrial onshoring, and data center growth drive new load and reliability requirements, utilities are making generation and interconnection decisions against forecast ranges wide enough to change which investments make sense. Tasked with advancing science and technology for the global energy industry, EPRI's researchers took on a problem facing the whole sector: how to deploy advanced AI locally, without sacrificing performance, data sovereignty, or compliance. NVIDIA DGX SparkTM /Dell Pro Max with GB10 and NVIDIA DGX Station TM brought coherent unified memory to the deskside: enough for Spark to hold a 120-billion-parameter model and a million-token document cache resident at once, and for Station to stage a 50-member ensemble forecast in a single pass. EPRI used them to build Power Chat, a secure AI document research assistant for electric utility organization that runs entirely within controlled environments, and to accelerate the AI weather prediction workflows critical to grid operations and planning.
Customer
EPRI
Partner
Dell Technologies
Topic
Generative AI/LLMs
Simulation / Modeling / Design
Products
Key Takeaways
Run a 120B-Class Model Deskside.
- 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.
240X Faster Than CPU-Only
- 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, giving grid planners critical extra time for scenario evaluation.
On-Prem AI Without Compromise
- Both workflows run entirely on premises and are air-gap capable, with no external network dependency. Small R&D teams prototype, fine-tune, and evaluate domain-specific AI models on deskside hardware—maintaining full data sovereignty and regulatory compliance.
Why Are Energy Organizations Moving AI On-Premises?
EPRI's research is used across utility companies that operate critical infrastructure, and much of the material that research depends on documents how that infrastructure is designed, operated, and maintained. That sensitivity extends across the whole AI system: the source documents, the memory context, and the responses generated from them all have to remain inside the firewall. Those conditions shaped how EPRI approached on-premises AI research for the sector:
- Data movement and exposure: Operational and research data often cannot leave controlled environments due to regulatory and governance requirements.
- Security and governance: Strict isolation, auditability, and compliance are non-negotiable in the energy industry—AI systems must meet these standards by design.
- Cost predictability: As agentic workflows scale, owned infrastructure keeps AI spend fixed and known, leaving budget decisions with the utility rather than with consumption.
- Access and scheduling friction: Shared compute can be scheduled, but experimentation cannot. When every attempt waits for a slot, the scenarios a researcher can test shrink to what the schedule allows, and questions worth exploring go unasked.
The energy industry is seeing growing interest in AI-driven decision support—from grid reliability planning to weather-dependent operations. EPRI needed systems that could run large, reasoning-intensive models entirely on-premises without sacrificing performance. The NVIDIA Grace™ Blackwell architecture provided that at two scales: DGX Spark for individual development and testing, DGX Station for heavier workloads, both at the deskside and neither requiring dedicated data center infrastructure.
“Testing across DGX Spark and DGX Station allowed EPRI IT to see how each system fit in our infrastructure strategy. DGX Spark gives our teams an accessible platform to prototype ideas quickly and fine‑tune open‑weight models, creating the headroom we need for rapid exploration. DGX Station then lets us carry those prototypes forward into demanding AI weather prediction and downscaling workloads, where faster, larger ensemble runs give grid planners and operators more time to evaluate scenarios and make critical decisions in real time.”
Matt Corbett
Director, Information Technology Enterprise Infrastructure & Networking, EPRI
Use Case 1: Power Chat—Local AI for Technical Document Analysis on DGX Spark
What It Does
EPRI developed Power Chat, a document research agent for electric utility companies and R&D organizations. Running entirely on DGX Spark, it allows researchers to have natural language conversations with technical documents, logs, and reports, with no any data leaving controlled environments.
Users select technical relevant documents that are loaded into the model's memory via cache-augmented generation, making dense material available for near-instant conversational exploration. Power Chat uses gpt-oss-120b, an open-weight mixture-of-experts large language model (LLM), served through the NVIDIA vLLM NGC Container with prefix caching. Model weights, document cache, and inference all reside on the local system, so Power Chat remains fully functional on air-gapped and connectivity-limited networks.
Click image to enlarge
Why DGX Spark
EPRI runs Power Chat on the Dell Pro Max with GB10, powered by the Grace Blackwell Superchip. Its 128 GB of coherent unified memory eliminates the CPU-GPU data transfer bottleneck, and it is what makes the workload fit at all—essential for loading a 120B-class model, maintaining a large KV cache, and running the OS, all on a single deskside system. NVIDIA DGX OS provides a fully integrated, stable AI software stack out of the box, reducing setup complexity and enabling rapid prototyping.
The result is a private, cost-effective AI assistant that keeps a one-million-token KV cache for document recall across sessions, complies with data protection and export control policies, and gives small research teams direct access to state-of-the-art AI compute on demand, with no queue, no shared cluster, no data risk.
Use Case 2: AI Weather Forecasting for Transmission Planning on DGX Station
What It Does
EPRI uses DGX Station to improve the weather forecasting that transmission planning and operations depend on, combining AI weather prediction models, including Google DeepMind's GenCast and NVIDIA's FourCastNet, with NVIDIA CorrDiff for dynamically downscaling gridded climate and weather data. EPRI also used NVIDIA CBottle, a diffusion model that generates atmospheric states at kilometer resolution, to produce high-resolution outputs from GenCast and compare results.
CorrDiff uses generative AI to convert coarse gridded 25 km weather data to 2–3 km resolution, a capability that traditionally required expensive high-resolution numerical simulations. Running ensemble forecasts locally on DGX Station gives grid planners and operators more time to evaluate scenarios and make critical real-time decisions.
“Testing across DGX Spark and DGX Station allowed EPRI IT to see how each system fit in our infrastructure strategy. DGX Spark gives our teams an accessible platform to prototype ideas quickly and fine‑tune open‑weight models, creating the headroom we need for rapid exploration. DGX Station then lets us carry those prototypes forward into demanding AI weather prediction and downscaling workloads, where faster, larger ensemble runs give grid planners and operators more time to evaluate scenarios and make critical decisions in real time.”
Matt Corbett
Director, Information Technology Enterprise Infrastructure & Networking, EPRI
Why DGX Station
NVIDIA DGX Station is built on the NVIDIA GB300 Grace Blackwell Ultra Superchip with an NVIDIA RTX PROTM Blackwell GPU, delivering 20 PFLOPS of FP4 compute and 748 GB of coherent memory across the Grace CPU and NVIDIA Blackwell Ultra GPU. The unified memory architecture eliminates host-to-device transfer overhead when staging ensemble members, allowing EPRI to run a 50-member ensemble forecast in a single pass, work that previously required a multi-GPU NVIDIA HGX™ node or scheduled time on a shared CPU cluster. CorrDiff and the wider NVIDIA Earth-2 platform supply the downscaling models, and the compact deskside form factor means small teams can access this capability without using dedicated data center space.
Results at a Glance
Power Chat on DGX Spark (GB10)
| Metric | Result |
|---|---|
| Boot + model load + document encoding | ~15 minutes (one-time warm-up) |
| Single-user throughput (post warm-up) | 35 tokens/sec (streaming) |
| Multi-user throughput (10 concurrent users) | ~15 tokens/sec per user | 150 tokens/sec aggregate |
AI Weather Prediction on DGX Station (GB300)
GenCast model—global weather forecast at 1-degree resolution. Single ensemble member, 1 timestep:
| Platform | Time — 1 Ensemble Member |
|---|---|
| CPU-only server (64 cores / 128 threads) | 964 sec (~16 min) |
| Dell Pro Max GB10 (DGX Spark) | 71 sec |
| HGX (1x H100) | 44 sec |
| DGX Station GB300 + Best in class | 39 sec (~25x faster than CPU-only) |
50-member ensemble forecast, 1 timestep:
| Platform | Time — 50 Ensemble Members |
|---|---|
| CPU-only server | Estimated 10+ hours |
| Dell Pro Max GB10 (DGX Spark) | 1,846 sec (~31 min) |
| HGX (1x H100) | 229 sec (~3.8 min) |
| DGX Station GB300 + Best in class | 150 sec (~2.5 min)* |
The CPU-to-GPU speedup is a key enabler: What took 16 minutes per ensemble member on CPU—or 10+ hours for a full 50-member forecast—now completes in seconds to minutes on DGX Station. For grid planners and operators managing critical energy infrastructure, this time compression is the difference between reacting to weather events and getting ahead of them.
What’s Next for EPRI
EPRI is exploring several areas for continued development in collaboration with NVIDIA:
- Fine-tune domain-specific models for energy industry applications, including grid reliability analysis and regulatory compliance review.
- Scale AI weather prediction by running larger, more frequent ensemble forecasts on DGX Station to provide grid operators with more granular scenario planning data
- Continue collaborating with NVIDIA on Earth-2 platform advancements, including potential evaluation of NVIDIA cuOpt™ on DGX Station.
- Deploy NVIDIA NemoClaw™ on DGX Station, which installs the NVIDIA OpenShell™ secure runtime for policy-based controls over files, networks, credentials, and tools, to build a secure, always-on agent for battery failure and fire event intelligence, automating OSINT monitoring, global incident detection, and structured reporting workflows.
NVIDIA local AI systems share a consistent architecture from deskside to data center, so workloads that outgrow a single system can scale into a Dell AI Factory with NVIDIA deployment without replatforming.
Ready to Try This?
Wondering where your organization would even begin? EPRI's research white paper lays out a step-by-step path to your first on-prem AI use case—so you can stop scoping and start building.
Related Customer Stories
