This week, the entire US stock market and South Korean stocks were all bloodbathed, especially in semiconductors. Under the US stock technology momentum factor strategy, it fell 40%, setting the fastest and deepest drawdown record in history.

In contrast, US bank stocks’ earnings reports all came in above expectations (suggesting that consumption and credit have not been an issue). ASML and TSM’s earnings also beat expectations, and their guidance was raised.

The core reason behind this drop is still that storage leverage is too high. The high-position “many-to-sell-many”行情 caused by a chain liquidation, then when storage falls, the other sector’s targets can only fall with it, because there is no sector with fundamentals stronger than storage.

This week also had another major downside: the release of kimi3, whose performance went straight to No.1. I think the “kimi” thing needs to be viewed from two dimensions: the training side and the inference side, which are different perspectives.

On the training side, I believe it’s a major downside. Although we don’t yet know how many GPUs and how much cost Kimi used to train kimi3, it’s very likely far cheaper than OpenAI and A (an unknown company). Right now, everyone is guessing that at least part of it involved distilling leading closed-source models. This leads to a question: if it’s possible to quickly narrow the gap with closed-source models through distillation or some method we don’t know, where exactly are the barriers of large models? Does OpenAI and Anthropic’s massive spending on frontier training still have economic benefits? I think kimi 3 mainly changes people’s view of training compute demand.

On the inference side, I think it’s a positive. Even though kimi 3 isn’t much cheaper than closed-source models at the same level—roughly 20% cheaper—lower costs are actually beneficial for expanding AI demand. The more useful and cheaper large models there are, the wider the range of what AI can be used for and the more use cases there will be. Inference demand then increases. Also, kimi 3 has very large parameters; even ordinary H100s may not be able to run it, which further accelerates demand for high-end compute.

Overall, the panic path brought by kimi this time looks like this: KIMI 3 appears—people start questioning OpenAI and Anthropic’s compute demand and gross margin. If OpenAI and Anthropic’s compute demand and gross margin cannot be maintained, they won’t be able to place orders with the “four major cloud” providers (remember, most orders for the four major clouds currently come from these two). If the four major clouds’ backlog can’t be fulfilled, it can’t support capital expenditures—then the AI bubble bursts.

Panic has logic and reasons, but each link in the chain needs facts to prove it. However, one thing can be confirmed: Anthropic’s 80% gross margin cannot be sustained. No matter whether others distill or really have some special method, the fact is that models at the same level but cheaper already exist.

Additionally, even if OAI and A’s compute demand declines, it may be filled in by other models. Training can take shortcuts, but inference is real: how much compute translates into how much production power. Microsoft’s CEO has also been talking about “sovereign AI” recently, and in the future large companies may combine open-source models with their own company needs for post-training, to prevent data from being captured by closed-source models and then becoming increasingly stronger. These are all substantial compute-demand drivers.

Back to the stock market: I think the last two weeks of July basically set the tone for Q3. Based on current market sentiment, even after a deleveraging move, I don’t think there will be a V-shaped rebound, because sentiment is too bad—any bargain-buying rebound funds will just take profit after it rises a bit. Long-term capital will definitely wait until the four major tech companies’ earnings reports confirm results before entering, and only then will sentiment gradually recover.

If the four major clouds continue to raise CapEx in their earnings reports, then this is a rare bottom; if the four major clouds underperform expectations and also cut CapEx (a low-probability event), then everything is over—crypto can fully return in the second half of the year.