Argonne National Laboratory is expanding its AI inference infrastructure as demand from scientific research grows.

A lot of attention has gone into the computing power needed to train AI models, and for good reason. Training the largest models requires enormous amounts of hardware and a substantial amount of money. But training is only part of the story. Once a model is ready, it still needs computing resources every time someone uses it. That stage is called inference, and it is becoming a bigger consideration as AI finds its way into scientific research.

A scientist might use AI to work through experimental data or prepare a simulation. On its own, that might not take much computing power. But AI agents are beginning to take on more complicated work, sometimes making hundreds or thousands of calls to AI models along the way. Every one of those requests needs computing resources, and the costs can add up quickly.

Argonne National Laboratory is working to address this problem. It has announced a major expansion of its ALCF Inference Service, which was first launched in May 2026 using the lab's Sophia and Metis systems. The expansion also supports the Department of Energy's Genesis Mission, an effort in which AI is expected to play a much larger role in scientific research and development.

Developed in collaboration with HPE and Nvidia, the new Tara system significantly expands the computing resources available to the ALCF Inference Service, supporting more advanced AI models and a growing range of scientific applications.

Murat Keçeli, a computational scientist at Argonne, has seen the problem firsthand. He helped develop ChemGraph, which uses AI agents to help researchers plan and run chemistry simulations. The challenge is that these agents can make many calls to AI models, even for a single task.

Keçeli says they can consume far more tokens than a typical chatbot conversation, which can make commercial AI models expensive to use. With Argonne's inference service, researchers have another option that does not require paying a commercial provider for every request.

"Since tool calls can be frequent and might involve large inputs and outputs, token consumption can be much higher than conversation style usage and cost a lot for frontier AI models," Keçeli said. "Therefore, the inference service is very useful for agentic AI applications to minimize the cost."

Many research teams are not in a position to buy expensive GPUs or hire staff to manage AI systems. Even well-funded research institutions have limited money and technical support to go around. Argonne's service takes some of that burden off their hands, so researchers can focus on using AI in their experiments and simulations rather than spending time figuring out how to run the models in the first place.

To meet growing demand, Argonne is adding significant computing power. The largest addition is Tara, a system developed with HPE using its Cray EX4000 architecture. Tara has 672 nodes equipped with 2,688 Nvidia GH200 Grace Hopper Superchips. According to Argonne, Tara alone provides roughly 100 times the GPU memory previously available to the service.

The service also adds Minerva, a much smaller system with eight nodes and 64 Nvidia B200 GPUs, giving researchers another platform for running AI models.

Nvidia is not the only hardware supplier involved. Argonne has also expanded Metis, its SambaNova-based system, from two to six SambaRacks, bringing it to 96 SambaNova SN40L processors. These processors are built for high-throughput, low-latency inference. The mix of hardware matters because different AI models have different computing needs, and what works well for one may not be the best choice for another.

Researchers can use the service in two ways: through a web interface or by connecting models to their own applications via OpenAI-compatible APIs. Argonne uses Globus for authentication, and inference requests can be distributed across federated computing clusters.

The service is not limited to Argonne scientists. Researchers at other DOE laboratories can request access, including scientists at Oak Ridge, Los Alamos, Lawrence Livermore, and Brookhaven, and some have already used it.

Beyond the hardware, the service offers a broad selection of models. It provides more than 50 models, including Google's Gemma, Meta's Llama, and Argonne's own AuroraGPT.

Scientists at Argonne have used AI inference to analyze experimental data and to help predict plasma disruptions in fusion experiments.

Argonne has not shared updated usage figures since the expansion, so it is unclear how much demand the service is seeing or whether researchers must wait for resources. As AI agents take on more scientific tasks, another question arises: how do scientists ensure the results are actually correct? This is especially important when AI is involved in several stages of the work.