What happened
Argonne’s inference service now runs on two additional systems, Tara and Minerva. They were among five Argonne computing installations announced by the US Department of Energy and NVIDIA in October 2025, ahead of SC25, and now support the service that Argonne launched in May. The article describes them as capacity for researchers to run larger models and develop more complex autonomous scientific workflows.
Tara was developed with Hewlett Packard Enterprise and is built on the HPE Cray Supercomputing EX4000 architecture. Its 672 compute nodes contain 2,688 NVIDIA GH200 Grace Hopper Superchips and 258 TB of combined GPU memory. The article says that is roughly 100 times the GPU memory 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 lists Minerva as serving models including Mistral Large 4 and Reflection’s Beam.
The expansion also includes Metis, Argonne’s SambaNova-based inference platform. It grew from two to six SambaRacks, for a total of 96 SN40L processors. Unlike Tara and Minerva, Metis uses SambaNova reconfigurable dataflow processors, which are designed to reduce data movement during inference. The service now hosts more than 50 models, including general-purpose language models, scientific foundation models and computer-vision models. It is available to researchers across DOE laboratories and Genesis Mission projects, runs models on premises, and supports integration with development tools such as Codex and Claude Code.
The article places the capacity build-out alongside the Genesis Mission. Rick Stevens, Argonne associate laboratory director for computing, environment and life sciences, and director of the mission’s Transformational AI Models Consortium, said the goal is about 100,000 persistent agents running in DOE infrastructure by 2029. He estimated that a state-of-the-art agent may consume about 15 million tokens per day. Supporting that scale would require computing resources to host the agents as well as separate inference capacity for the models they call. Genesis has established a model-access gateway with a common API for commercial frontier, open and specialized scientific models, allowing agents to switch models without separate integrations for each provider.
Argonne is also developing tools for agents to submit computing jobs and access data on its other systems. The source describes a Genesis project, SYNAPS-I, in which an agent and computer-vision models scanned a microelectronics sample to find a specified region without manual direction at every step. It gives no pricing or general public access terms for the service.
What the archive adds
Oracle’s accelerator expansion is an earlier report about cloud capacity and accelerator options, not this national-lab service. This is a new fact in the broader infrastructure story: Argonne’s service now has quantified GH200/B200 capacity and a catalog serving more than 50 models. The other archive notes concern open-model inference and cloud expansion, rather than this specific deployment.
Trend signal
This supports agent economics moves to completed work only weakly: the source links agent ambitions to large inference and hosting requirements, but gives no cost per verified task or outcome measure. It is better read as evidence that the compute layer for persistent scientific agents is being provisioned before those economics are established.
Why now
The service expansion is occurring alongside the DOE’s Genesis Mission planning, whose stated target is about 100,000 persistent agents by 2029. The article also points to SC26, scheduled for November 15–20 in Chicago, where Argonne’s Rick Stevens is due to discuss the mission and secure research environments for workflows spanning facilities.
Why it matters
For scientific teams, a service that runs models on premises and spans general-purpose, scientific and vision models can become a shared substrate for research workflows. It may reduce the need to build each model connection independently, especially as a common gateway is intended to let agents switch among commercial, open and specialized models. The integration with Codex and Claude Code also indicates that the service is designed to meet researchers in familiar development environments, although the source does not detail the integrations.
The larger operational lesson is that agent capacity is not just a model-selection problem. Hosting persistent agents and serving the models they call are distinct resource demands, and agents that can submit jobs or reach instruments require controls beyond ordinary inference. Argonne’s scale is specific to DOE research, not a ready-made blueprint for every company, but it makes the infrastructure and governance questions visible: how to allocate compute, connect data and facilities, and maintain trustworthy operation when agents act across systems.
Limits and uncertainty
The article reports system specifications and service capabilities but does not provide throughput, latency, utilization, reliability, or comparative performance figures. It does not state who can obtain access beyond researchers across DOE laboratories and Genesis Mission projects, nor give costs, quotas, service-level commitments, or deployment instructions. The 100,000-agent figure is a goal for 2029, not a confirmed deployment, and the estimate of 15 million tokens per agent per day is attributed to Stevens rather than presented as measured usage across agents. The source also does not establish that the described SYNAPS-I demonstration generalizes to other experiments or that agents can yet operate laboratory equipment autonomously at scale.
Reader takeaway
Argonne has materially increased the compute and model capacity behind its scientific inference service, while DOE’s agent ambitions make clear that hosting agents and serving their models require separate infrastructure. The useful signal is the integration direction—on-prem models, a common gateway, and connections to research systems—not a claim that autonomous science is already operating at the stated future scale.