Groq signs MOU with US Department of Energy on inference for the Genesis Mission

Groq has signed a memorandum of understanding with the US Department of Energy (DOE) under the Genesis Mission. The document is dated December 18, 2025. It contains no amounts, no deadlines and no supply commitments: it is a framework for exchanging information and "potential cooperation." But for a company that builds its own LPU chips and serves more than 2.8 million developers through GroqCloud, appearing in a federal program carries weight in itself. Chief Revenue Officer Ian Andrews attended the Genesis Mission event at the White House.

What the MOU actually says

The parties listed four areas. The first is low-latency inference for scientific discovery, including agentic scenarios with a closed experimental loop. The model proposes a hypothesis, the instrument produces data, and the model adjusts the next step. The shorter the pause between those steps, the more iterations fit into a day.

The second is energy-efficient inference and resilient domestic supply chains. Here the document explicitly names an evaluation of LPU-based architectures. The third is joint work on benchmarks, practices and standards for deterministic and energy-efficient inference. The fourth is selecting applied tasks for the Genesis Mission.

Key parameters of the agreement:

• Document type: MOU, meaning a non-commercial framework with no financial terms.
• Date: December 18, 2025.
• Program: the US Department of Energy's Genesis Mission.
• Technology base: LPU and the GroqCloud platform.
• Metrics for standardization: reproducibility, latency, throughput, energy consumption.
• Groq's audience: more than 2.8 million developers and large Fortune 500 companies.

Why the emphasis on determinism

The benchmark item is the most interesting one. Scientific workloads require reproducibility: the same run should produce a comparable result and predictable timing. Groq has long sold the idea of deterministic execution, where latency is known in advance rather than fluctuating with the load from neighboring tasks. If DOE evaluations adopt this metric, the advantage goes to whoever designed their hardware for it.

That is the calculation. Measuring inference only in tokens per second and price per million tokens suits GPU makers. Add energy per response and latency variance, and the picture shifts. For now this is an intention of the parties, not a methodology, but whoever helps write the methodology usually wins under it.

A statement from Jonathan Ross, Groq's CEO and founder, sets the tone: "There isn't enough compute for AI to benefit everyone. We are working with the White House and the Department of Energy to change that." It is a statement of purpose, not of volumes: it does not say how much capacity or in which centers.

Political context and limitations

Groq presents the agreement as part of the "American AI stack" that the Genesis Mission is meant to define, scale and ultimately export to partners. In July the company backed the AI Action Plan. The position is clear: a vendor with its own silicon, its own software and its own cloud fits neatly into the narrative of national control over the supply chain.

The limitations are obvious. An MOU is not a contract. There is no information on which DOE labs will get access, which systems the pilots will run on, or who pays. There are also no comparative measurements of LPU against GPU on scientific tasks.

What to watch next:

• whether pilot projects appear at national laboratories with published results;
• whether joint benchmarks are released, and in what form: open or closed;
• whether the MOU turns into procurement or grant funding;
• how competitors, above all GPU vendors, react to an attempt to shift the metrics toward energy and determinism.

For teams building agentic scientific pipelines, the practical takeaway is simple: when evaluating inference, factor in not only the token price but also latency variance and energy per loop. If DOE standards move in this direction, procurement requirements will start to echo them fairly quickly.