Nvidia's $20 billion deal to license Groq's AI chip technology and hire away its top engineers has drawn an antitrust probe from the US Department of Justice, according to a New York Times report. The arrangement left Groq's inference business technically independent but stripped it of the engineering talent needed to keep developing its own hardware, prompting scrutiny over whether it was structured deliberately to dodge merger review while still handing Nvidia effective control.

Groq built its reputation on SRAM-based accelerators that dramatically sped up AI inference compared with Nvidia's GPUs, and Nvidia has already folded this technology into its LPX server racks, unveiled at GTC in March, which pair 256 Groq-3 chips with Nvidia's Vera Rubin GPUs. Even if the DOJ successfully forced Nvidia to unwind the deal, the article argues it would change little, since Nvidia had already opened its rack and interconnect designs to the wider industry via the Open Compute Project, and rivals such as Cerebras, SambaNova and d-Matrix are pursuing similar hybrid architectures with other partners. Nvidia maintains the deal represents "the American system working as designed" to reward innovation and entrepreneurship.