Chris Miller

But design only matters if you can get them manufactured.

And that's the challenge that Huawei and the other

Chinese chip designers have faced.

And so I think if you look at the startup ecosystem in China, for example, companies like MoreThreads, which you just mentioned, had an IPO earlier this month, very successful IPO, valuing it at around $50 billion.

That's a lot, but it's only 1% of Nvidia's valuation, precisely because they can't manufacture enough chips to sell them at volume right now.

I think one of the debates that the AI world was having in the first half of this year is whether the scaling laws were still holding.

The idea that you build a bigger AI system, you get a better AI system.

In the last couple of months, we've seen a number of important data points that suggest, yes, it still is true that quality follows size.

Gemini 3, for example, is, I think, a strong data point.

in that direction opus 4.5 from anthropic but the best is actually deep seek which released a new model earlier this month and the paper that accompanied that release explicitly said we deep seek are actually struggling to uh pre-train at the scale we want because we don't have access to enough chips so that's why i think there's so much at stake and the decision by the us to allow h200 sales to china and by the decision by the chinese government uh what volume to actually

Except my guess, and there's a ton of uncertainty here, my guess is actually that the volumes that ship are smaller than many people think, both because of uncertainty on the US side as to what exactly the security concerns will be that need to be addressed.

And the Trump administration has promised some sort of security check.

And also because China's government wants to provide a space for Huawei and other firms to sell to a protected market.

And so there may be less H200 sales than I think some of the headlines suggest.

And if that's true, we might not see as big of a reversal in the balance in computing power as we've got today.

But it's worth tracking very closely because the US advantage in compute has been a core facet of its leadership in AI over the past couple of years.

Yeah, I think that's right.

DeepSeq trying to disassemble NVIDIA servers that were in Southeast Asia, ship them into China and reassemble them is a data point just how far Chinese firms are going to get access to the compute that they realize they need.

And if your AI supply chain requires disassembling, smuggling and reassembling, you know, that's not the kind of supply chain you want to rely on to produce your most sensitive and advanced technology.

I think to start with the second part of the question first on open source, we've now begun to get some reasonable data on whether cheaper or often free open source models from China are winning market share.