I only just saw this news recently: at the end of last year, NVIDIA acquired Groq’s chip inference business, and the logic behind this deal is actually quite interesting.

At this year’s GTC conference, Huang Renxun explained for the first time in detail why they went after Groq. Put simply, it’s because they saw the inference market splitting into different segments. Previously, everyone was optimizing inference chips toward just one goal: maximizing throughput. But now, things have changed—different users are willing to pay different prices for different response speeds.

This logic is crucial—if I can provide developers with faster token response times, improving their work efficiency, then they’ll be willing to pay for it. This market for high value and low latency is an opportunity that has only recently emerged. Huang calls it an expansion of the inference market’s Pareto frontier: what used to be all about high-throughput solutions now adds a new track—low-latency, high-price solutions.

And Groq happens to be an expert in this area. Their LPU architecture is known for deterministic low latency, which completely complements NVIDIA GPU’s high-throughput direction. The Groq 3 LPU released in March uses Samsung’s 4nm process, and its inference performance on trillion-parameter models is 35 times higher than Blackwell NVL72—those performance differences are indeed huge.

From the perspective of the product lineup, this acquisition fills a gap in NVIDIA’s coverage of the inference market. Think about it: with the same model, you can set different pricing strategies depending on response time. Throughput may be a bit lower, but the unit price can make up for it. Groq’s addition gives NVIDIA a more complete footprint in the inference market. Strategically speaking, this deal is very clear-cut.