From the perspective of "Inference will explode," @inference_labs@, its position is actually quite clear.

When models no longer gain significant advantage from parameter scale, the real variable becomes:
How many inferences can be performed per unit time.
It's not about answering smarter once, but about whether continuous, high-frequency, parallel thinking is possible.

A small model + multiple Agent structure essentially amplifies inference calls.
Dozens of Agents running simultaneously, verifying each other, breaking down tasks, and then merging results—what's consumed isn't "model capability," but inference channels and throughput capacity.

Inference Labs isn't about competing over the model itself but solving a more fundamental problem:
When inference becomes something like request traffic, who will support this road?

This is similar to how, after CPU clock speeds hit bottlenecks in early years, the industry shifted focus to memory, buses, and parallel computing.
Computing power isn't unimportant anymore; how it's scheduled and scaled up is more critical.

So, Inference Labs is more about paving the way for the next phase of AI usage.
It's not just a model answering questions,
but an entire inference system operating continuously.

Models won't grow infinitely, but inference will become increasingly dense.
Once this trend is established, the value of infrastructure is just beginning to be priced.
@KaitoAI @Bybit_Web3