Nvidia said Monday its Groq 3 LPX rack has entered full production, the first commercial output from the chipmaker’s largest acquisition to date. The system will be installed at neocloud provider Nebius alongside the company’s new Vera central processors and Rubin graphics processors, with the combined deployment expected to come online before year end.
The asset deal and the hardware
Nvidia paid $20 billion in December for assets from Groq, a startup that built its architecture around 500 megabytes of on-die SRAM to sidestep memory bottlenecks during the decode phase of inference. Samsung fabricates the Groq 3 die, while TSMC continues to make Nvidia’s GPUs. Each LPX rack packages 256 of those chips. Nvidia cites an Artificial Analysis benchmark showing 3,400 tokens per second per rack.
Where it fits in the stack
Senior director Dion Harris described the rack as a complement to, not a replacement for, general-purpose GPUs. Low-latency silicon handles the decode step where responsiveness matters most, coding assistants, agent workflows, any service tier that charges a premium for speed. Cloud operators can price those tokens higher, Harris said, because the latency-sensitive slice of demand is willing to pay for it.
The allocation signal
At the March unveiling, chief executive Jensen Huang indicated that roughly one-fourth of the data-center capacity earmarked for coding workloads would be filled with Groq racks. The remainder of the facility, he said, would run entirely on Vera Rubin systems. Huang also projected $1 trillion in cumulative revenue from the Blackwell and Vera Rubin families combined through 2027.
Competitive pressure
AMD moved earlier this year to pair its rack-scale systems with Cerebras chips, which recently listed publicly. OpenAI’s new Ultrafast mode, powered by Cerebras, advertises 750 tokens per second. Nvidia’s argument is that its rack delivers roughly four and a half times that throughput while sitting inside the same Vera Rubin infrastructure it already sells.
Earnings watch
Nvidia reports quarterly results on Wednesday. Investors will look for any update on Vera Rubin ramp rates, Groq attachment rates at Nebius and other early customers, and whether the $1 trillion revenue trajectory still holds given the capital intensity of running two parallel silicon roadmaps.