Nvidia’s $20 B Groq “license-plus-talent” deal vaults it into AI inference

Chip giants have long bought or licensed niche rivals to leapfrog into emergent workloads—Intel’s $16.7 B Altera FPGA buy in 2015 is a close parallel—yet few have done so through a quasi-acquisition license designed to evade regulators. The Groq pact reflects two powerful currents: 1) the economic center of AI is shifting from one-off model training to perpetual inference cycles, akin to the 1990s move from mainframe builds to client-server transaction volume; 2) scarce semiconductor talent is now treated as a strategic resource on par with fabs or IP. If Nvidia can weave Groq’s low-latency architecture into its GPU ecosystem, it could entrench a vertically integrated AI monopoly for a generation—much as IBM’s 360 platform set computing standards for decades after 1964. Conversely, the workaround structure signals how dominant firms may increasingly bypass formal M&A rules, testing antitrust frameworks that were built for the industrial age. Whether regulators adapt, or rivals counter with open-source or custom silicon, will shape who controls the computational infrastructure that underpins the next 100 years of digital productivity.