I think there are a handful of converging factors contributing to the recent pull back in AI stocks. Just wanted to throw out some thoughts on one of those factors for the sake of discussion.
The release of Kimi K3 and ensuing panic about its impact on US AI businesses has unfairly hit every layer of the US AI stack. Kimi K3 is a problem for openAI and anthropic without a doubt, although I would argue its impact is less significant than the market seems to think. Yes, it does put some pressure on the labs' margins, but I think there is a trust factor that, warranted or not, will keep especially large corporations locked into US models. There also may be questions about the hyperscalers willingness to serve chinese models (and I could see the US gov potentially intervening in this area) - if hyperscalers don't serve the models, there is just no world where chinese labs can find enough compute independently to meaningfully take significant revenue from anthropic/openAI.
The bigger point I want to make is that in my opinion Kimi K3 is in no way a threat to hyperscalers' business or semiconductors. I firmly believe that jevons paradox is going to be true of AI, and if you believe this, the cost/token going down is not necessarily a headwind to revenue, and may actually be a tailwind. For hyperscalers, as long as there is demand for compute and labs are willing to pay the same margins for compute, their business is unaffected. Just because kimi is more efficient doesn't mean the model provider can't afford to pay hyperscalers. If the end user can get more value / compute, they should in theory pay more for compute, and so this should be accretive to both hyperscaler and model margins. The question is only whether demand for compute increases or decreases.
Of course there is a chain reaction that flows from the labs to hyperscalers down to semiconductors. My point is that for the lower layers, it doesn't matter which lab is in the lead or what model is the most efficient as long as demand for tokens and the price that consumers can be charged for compute remains the same.