Groq, the AI‑chip startup that Nvidia licensed in December, recently raised $650 million in a new round. The move surprises observers because the company never merged with Nvidia; its corporate entity stayed independent, continuing to run its own datacenters and inference API, which handles high‑speed inference for smaller models at scale to serve clients.

Groq’s chips use an all‑SRAM design that favors fast token throughput over cost efficiency. The company’s LPUv1 units, now seven years old, power its existing four datacenters. Newer LPUv3 models are sold by Nvidia to any cloud provider, eroding Groq’s unique speed advantage while retaining a niche for ultra‑fast inference on models like GPT OSS 120B today.

Hardware is only one side of the story; Groq’s real asset lies in its four fully operational datacenters, each already loaded with LPUv1 chips. Compared to public rivals CoreWeave and Nebius—both valued near $50 billion yet owning far fewer sites—Groq’s facility count alone suggests a valuation in the billions, albeit at a lower cost base.

With Nvidia’s licensing deal and a steady pipeline of small‑model inference workloads, Groq can leverage its low‑cost, high‑throughput architecture to capture market share where speed outweighs raw scale. The company’s ability to scale its datacenters quickly will determine whether its 650‑million‑dollar infusion translates into lasting competitive advantage for developers seeking latency‑critical applications in cloud services today.