I just saw very important news about the acquisition Nvidia made. The company acquired Groq’s inference chip division for $20 billion last December, and this move is starting to make more sense now.

Huang Renxun, CEO of Nvidia, explained in a recent interview the real reason behind this strategic decision. It turns out this isn’t just about a normal productivity boost—Nvidia is targeting a completely new market: low-latency, high-value inference. This market has only recently started to take shape as users begin paying different prices depending on response speed.

The idea is simple but powerful: if you can offer tokens to developers with lower latency, meaning they can work more efficiently, they’ll pay a higher price. Huang described it as expanding the market’s boundaries—adding an entirely new sector instead of focusing only on high productivity.

The first product after the acquisition appeared in March: Groq 3 LPU, made with Samsung’s 4-nanometer technology. The numbers are impressive: inference throughput per megawatt on trillion-parameter models reaches 35 times that of Blackwell NVL72. Groq’s architecture is known for its low, predictable latency—which is exactly what Nvidia’s product lineup was missing.

This is incredibly smart: the same model, but with different prices depending on response time. Even if productivity is lower, the higher price makes up for it. Groq filled the gap that was present in Nvidia’s strategy, and clearly the market has begun to split into different segments.