Did you notice NVIDIA's very interesting strategy? They acquired Groq's inference chip business for $200 billion, and now it makes a lot more sense why they did that.

What caught my attention was Huang Renxun's explanation of the logic behind this acquisition. Basically, the inference market is becoming segmented. Previously, everyone was focused on one thing: increasing throughput. But then, the commercial value of tokens changed significantly, and different users are willing to pay different prices depending on response speed.

It's like this: if I can provide engineers with faster responses, allowing them to work more efficiently, they will be willing to pay more for that. And this demand for low latency is relatively new in the market.

Then comes Groq. Their LPU architecture is known precisely for its low deterministic latency, which perfectly complements NVIDIA's high-throughput GPU approach. When they launched the Groq 3 LPU in 4nm, they showed that inference capacity per megawatt in trillion-parameter models is 35 times higher than the Blackwell NVL72. That's no small feat.

In other words, NVIDIA filled an important gap in its product line. Now they cover both the high-throughput segment and the low-latency, high-value-per-unit segment. Pareto expansion, as some call it. Same model, different prices depending on response time. Lower throughput, but the unit price makes up for it.

This is the strategy: it's not competition, it's complementarity. And it makes a lot of sense considering how the AI market is evolving.