Mandeep Singh

We have heard commitments from OpenAI, which everyone has started to question now in terms of the 26 gigawatts that they have been talking about, but clearly the numbers are big in terms of the gigawatts that are in the pipeline in terms of the capacity that companies want to add.

You know, the efficiency part when it comes to token per watt will remain a key metric because you won't be able to bring all that power online in the next 12 months, and it's going to be a lot more staggered than people can imagine.

Yeah.

Well, their notion of model layer as an intelligence layer is to open source a lot of their models.

And that's where they have kind of taken a different approach.

The models are...

available for any company to use.

In fact, they want to build an ecosystem outside of China where Europeans or other companies adopt their open source models.

And from that perspective, they're trying to distill a lot of their models to smaller versions so that they can run on the phones.

I just feel like they have espoused that open source concept and you see that with ByteDance and all these companies and they're trying to get to agentic use case a lot quicker in terms of models being able to book your travel or book your restaurant.

We've seen that how good the WeChat app is in terms of being that super app.

So that's what they're trying to get to with the agentic use cases is deploy a lot more of that using the LLM layer and do that at a very efficient price because at the end of the day, they have espoused cost and efficiency when it comes to running their infrastructure.

So they want that to be reflected at the agent layer as well.

Right, so for our workload specifically, for image and video, H200s are a little bit overkill for us, at least for the Hopper generation.

But when we're actually...

Moving away from the Hopper architecture to Blackwell architecture, we're actually just going directly from H100s to B200s or B300s because that actually gives us more flops per watt or per dollar.

We're actually moving over.

Yeah.

And look, I mean, based on consensus right now, we are talking about at least a 50% increase in capex or meta next year.

And no one, I think, would mind that given everyone's fees, there will be lift in revenue from AI down the line in the near term.