SAN JOSE, California – It’s finally here, the chip that Nvidia CEO Jensen Huang claimed would “surprise people” at GTC was revealed: Nvidia Groq 3 LPU.

The chip, or a language processing unit (LPU), marks the beginning of Nvidia’s work with Groq, following its pseudo-acquisition come licensing deal with the chip designer.

Set to feature in liquid-cooled racks dubbed LPXs, some 256 LPUs interconnected through 640 Tb/s of scale-up bandwidth and 128-gigabytes (GB) of on-chip static random access memory (SRAM) will look to power ultra-low-latency AI inference workloads. They’ll be available in the second half of 2026.

Nvidia is positioning the LPU not as a replacement for its GPU line, far from it. Instead, it’s an augmentation string of its existing stack.

Ian Buck, the company’s data center head, contrasted: “The GPU with its 288-gigabytes for memory, compared to only 500-megabytes of SRAM. The LPU is only one-five-hundredth of the amount of capacity per chip, but the bandwidth is exceptional: 22 terabytes to 150 terabytes per second.”

“Nvidia is working to combine these two processors to do the decoding operations of a GPU and do it as one,” Buck added.

Nvidia is believed to have spent around $20 billion on its deal with Groq, which was announced on Christmas Eve 2025. The tech giant licensed the startup’s technology and snapped up its leadership team, including founder and former Google TPU designer Jonathan Ross.

Despite the deal, Groq is an independent business headed up by finance lead Simon Edwards with no impact on its GroqCloud inference platform.

Prior reports suggest that AI21 Labs may potentially follow in Groq’s wake, with a supposed acquisition in the range of $2 billion to $3 billion floated at the turn of the year.