Nvidia Bought Groq to Own the Part of AI You Actually Pay For

Nvidia wasn’t satisfied being the landlord of AI training, so it just reportedly bought Groq to collect rent on inference too.

Yes, inference. The unsexy part where your chatbot actually answers you, your company’s “AI assistant” hallucinates your job away, and the bill shows up like it pays taxes.

Groq’s whole thing is making inference cheaper and faster with purpose-built chips and a software stack that doesn’t cry for a GPU every time you ask it to summarize an email. Nvidia’s whole thing is seeing that and going, “Cute. Mine now.”

The PR version will be “accelerating innovation” and “meeting skyrocketing customer demand.”

Training is the flashy gym montage. Inference is the lifetime subscription where the app charges you monthly to tell you you’re still ugly. Nvidia already had the training chokehold via CUDA, supply deals, and basically being the only bouncer letting you into the club. Groq was one of the few credible “we don’t need Nvidia” angles for serving models at scale.

Competitors wanted to pry open Nvidia’s grip with alternative inference hardware, open stacks, and “choice.” Nvidia responded the way every monopoly responds: by buying the exit.

Meanwhile, startups get to celebrate “ecosystem maturity” while quietly realizing their roadmap now depends on Jensen Huang’s mood and a supply chain that treats chips like wartime rations.