Nvidia's upcoming LPX rack, which features 256 Groq LPUs, will have a power envelope of up to 160kW.

The system will be based on Nvidia's MGX reference architecture, and has the same power consumption as the Vera Rubin NVL72, which is required next to an LPX.

"Each tray has eight LPUs, a host CPU, an FPGA, and a Bluefield 4 SuperNic," Nvidia's senior leader, accelerated computing and inference products, Stuart Pitts, told DCD.

"So 32 CPUs in a rack, 32 FPGAs, 32 Bluefield 4s, and then 256 LPUs, fully liquid cooled on the standard MGX architecture."

Pitts joined Nvidia as part of the December $20bn licensing deal and acquihire of Groq, which developed the LPU. He declined to disclose the CPU's specs, with the chip believed to be an Intel Xeon rather than Nvidia's own Arm CPUs.

Ensuring the rack has the same power draw as Vera Rubin was "very important," Nvidia's data center head, Ian Buck, told DCD.

"We wanted to make sure that the LPX rack could go into a data center that was designed, built, and ready to go for Vera Rubin."

Both are expected to launch around the same time in the latter half of 2026.