At Hot Chips 2026 Nvidia announced the start of full production for the Groq 3 LPX chip, the first device to emerge from its acquisition of Groq last year. Nvidia paid $20 billion for Groq, a company previously valued at under $7 billion.
What the Groq 3 LPX does
The Groq 3 LPX is optimized for a single task: accelerating token generation. According to Nvidia’s demonstration, a 5,000-token response that previously required 50 seconds can now be produced in 1.5 seconds. That example corresponds to a 35× reduction in token-generation cost if costs are primarily tied to execution time.
It is important to note that the chip does not change model intelligence: Groq 3 LPX does not alter model architectures or capabilities, it only substantially reduces response latency.
Practical implications
Recent AI development has trended toward larger, more capable models, but those models often respond more slowly because each additional reasoning step adds seconds to latency. For multi-step agents this can mean minutes of waiting.
Groq 3 LPX effectively removes much of that waiting: a 35× reduction in token-generation cost and dramatically shorter latencies can make products viable that were previously too expensive or too slow to operate profitably. Lower infrastructure costs may therefore be as influential for AI adoption as improvements in model quality.
Strategic context for Nvidia
At the time of the acquisition, many analysts considered the $20 billion price high relative to Groq’s prior valuation of under $7 billion. The launch of Groq 3 LPX suggests the purchase may yield significant infrastructure advantages for Nvidia, since a single hardware generation can materially lower the operating costs of AI services.
Summary
Nvidia’s Groq 3 LPX, now in full production, focuses on speeding token generation and — in the demonstrated example — cuts token-generation costs by 35×. While it does not increase model intelligence, the chip reduces execution costs and latency in ways that could broaden which AI products are economically feasible.
