The US military used a government version of xAI’s Grok model in workflows that enabled forces to deploy 2,000 munitions during 96 hours of operations against Iran, according to a sworn declaration by Pentagon artificial-intelligence chief Cameron Stanley. The filing provided an unusual official account of generative AI’s role in military targeting and intelligence.
Stanley, the Defense Department’s chief digital and artificial intelligence officer, said the Grok Gov Model was integrated with the Maven Smart System. He described Maven workflows as supporting targeting, intelligence, readiness and recruitment, and presented the Iran operation as evidence of increased operational efficiency. The declaration did not specify the dates on which Grok was used or identify individual targets selected with its assistance.
The details emerged not through a military briefing but through an environmental lawsuit in federal court in northern Mississippi. The NAACP sued xAI and its subsidiary MZX Tech on April 26, 2026, alleging that the companies illegally operated 27 methane-gas turbines in Southaven, Mississippi. The turbines supplied power associated with Colossus 2, a supercomputer facility used to train and improve Grok models.
The civil-rights organisation alleged violations of the Clean Air Act and sought an order stopping unpermitted turbine operations, requiring pollution controls and imposing financial penalties. It argued that surrounding Black communities were bearing harmful pollution associated with the data centre. Methane turbines can produce nitrogen oxides, pollutants associated with ozone formation and respiratory harm.
Stanley’s declaration supported the Trump administration’s intervention on xAI’s side. He argued that interrupting Colossus 2’s power supply could impair the Pentagon’s ability to perform national-security missions and keep pace with adversaries. In a crisis, he said, demand for Grok Gov inference could increase sharply, making available computing capacity strategically important. The Justice Department urged the court to dismiss the suit on national-security grounds.
The filing also linked an AI model’s battlefield use to the physical infrastructure and environmental costs required to operate it. Stanley compared data-centre inference capacity to other long-term defence resources, presenting it as foundational rather than merely commercial. The NAACP’s case challenged whether that role excused compliance with pollution law.
The declaration did not explain the degree of human review involved in targeting or whether the model recommended, prioritised or merely processed information about particular strikes. It therefore confirmed operational use and scale without resolving questions about accountability, accuracy or decision authority. Those omissions are important because deploying 2,000 munitions is not the same as autonomously choosing 2,000 targets, and the supplied evidence supports only the former. The filing left those operational safeguards unexplained.
