The Toronto-headquartered AI firm will build systems that can run on hardware that customers control directly, including on the chip giant’s DGX Spark supercomputers, designed for running AI agents. (The Logic)

Talking point: Cohere claims its AI models can run on less computing power than its rivals’ products require, making them well-suited to applications and clients that can’t or don’t want to use massive data centres or cloud services. The firm didn’t identify the “specific hardware footprints” for which it’s working on new technology with Nvidia. But Cohere executives have previously touted its models for uses such as analyzing data on ships at sea. Governments around the world are trying to increase compute capacity within their borders, so they maintain control of the AI that’s set to be used in critical services and industries. Nvidia hopes to sell the chips that will fill many of those new data centres. Cohere has also capitalized on interest in sovereign AI.