Prateek Joshi

<start of post>

There’s a strange asymmetry in how we think about AI progress.

We talk a lot about bigger models, more data, and more compute. But some of the biggest jumps may come from making the surrounding systems much cheaper and more reliable.

Inference is an obvious example. If running a model becomes 10x cheaper, you don’t just run the same workload for less money. You start using models in places where it was never worth using them before. More agents, more experiments, more personalized software, more continuous monitoring.

And then the usage creates new data, which creates better products, which creates more usage. Maybe the real bottleneck is not intelligence but the cost of asking for intelligence.

This is why I find the AI infra layer so interesting. Chips, networking, memory, inference engines, orchestration, evals, and all the unglamorous systems work underneath the model.

The models get most of the attention. But the infrastructure determines how often we can actually use them.

What does the world look like when intelligence is cheap enough that every company can run thousands of agents continuously?

<end of post>