Argonne National Laboratory has expanded its AI inference service with new computing resources that will allow researchers to run larger AI models and develop more complex autonomous scientific workflows.

Two new systems, Tara and Minerva, were among five Argonne computing installations announced by the DOE and Nvidia in October 2025, ahead of last year’s SC25 conference. Both now support Argonne’s inference service, launched in May.

Developed with Hewlett Packard Enterprise, Tara is built on the HPE Cray Supercomputing EX4000 architecture. Its 672 compute nodes contain 2,688 NVIDIA GH200 Grace Hopper Superchips and a combined 258 TB of GPU memory, roughly 100 times what was previously available to the inference service.

Minerva, developed with World Wide Technology, is an eight-node NVIDIA DGX B200 cluster with 64 Blackwell GPUs and 11.5 TB of high-bandwidth memory. Argonne’s model directory already lists Minerva as serving models including Mistral Large 4 and Reflection’s Beam.

Argonne has also expanded Metis, its SambaNova-based inference platform, from two to six SambaRacks, bringing its total to 96 SN40L processors. Unlike the NVIDIA-based machines, Metis uses SambaNova’s reconfigurable dataflow processors, designed to reduce data movement during inference.

The inference service now hosts more than 50 models, including general-purpose LLMs, scientific foundation models and computer vision models. Available to researchers across DOE laboratories and Genesis Mission projects, the service runs models on premises and supports integration with AI development tools such as Codex and Claude Code. Argonne is also developing tools that will allow agents to submit computing jobs and access data stored on its other systems.

Genesis Mission Plans for 100,000 AI Agents

Those connections are central to the DOE’s Genesis Mission, which seeks to accelerate scientific discovery by giving AI agents access to research data, supercomputers and laboratory instruments.

During a presentation to the Google Cloud Advanced Computing Community on Tuesday, Rick Stevens, Argonne’s associate laboratory director for computing, environment and life sciences and director of the Genesis Mission’s Transformational AI Models Consortium, outlined an ambitious plan to deploy autonomous AI agents across DOE research facilities.

“Our goal by 2029 is to have about 100,000 persistent agents running in the DOE infrastructure,” Stevens said.

Stevens estimated that a state-of-the-art agent can consume approximately 15 million tokens daily. Supporting agents at that scale would require vast computing resources to host them and separate inference capacity for the models they call. He also explained how Genesis recently established a model access gateway that provides a common API for commercial frontier models, open models and specialized scientific models. The approach allows agents to switch between models for different tasks without needing separate integrations with each provider.

Autonomous Research Raises New Infrastructure Questions

Argonne has already demonstrated how its inference service can support autonomous experiments. In one Genesis project, SYNAPS-I, researchers used an AI agent and computer vision models to scan a microelectronics sample and locate a specific region without manually directing each step.

The demonstration offers a vision of how autonomous agents could eventually operate scientific instruments, launch simulations and use experimental results to determine what to investigate next. For HPC facilities, that raises questions about how to manage agents that can access computing resources and laboratory equipment without constant human supervision.

Those questions will be on the agenda at SC26, Nov. 15-20 in Chicago, where Stevens is scheduled to discuss the Genesis Mission. A separate session will examine how to provide secure, trustworthy research environments for AI workflows spanning multiple facilities.