Goldman Sachs: The sell-off triggered by this round of DeepSeek is nearing its end

Four trading days after Deepseek, let's take a look at some new sentiment vs narrative changes;

1/ On Monday, the market experienced a huge shock, and the observations from Goldman Sachs' trading desk were: 1) Everyone was selling off comprehensively to manage risk; 2) No one was catching the falling knife; 3) Leverage/ETFs amplified volatility; 4) There was a clear rotation among sectors (semis flowed out to some defensive sectors, including software);

2/ As of now, the semis sector has narrowed its decline to -5% this week (having rebounded 5% + 1% YTD); on Monday, the market was in a frenzy of risk aversion, and while there are still some gradual reductions in positions, the selling pressure has eased significantly; simply put, those who wanted to sell have mostly sold.

3/ Those who want to buy are also starting to appear; the stabilized state of the market allows everyone time to slowly digest information, and there is no need to panic.

4/ Regarding the information gap on Deepseek, both buy-side and sell-side have nearly exhausted their discussions; yesterday we promoted a Deepseek seminar with 5,000 people from a foreign investment bank; the effect of various AI bulls coming out to stabilize the market is quite evident, for example, the "Jevons Paradox" has already become widely known, and the narrative is starting to resonate; META + MSFT's performance has also helped;

5/ After META's performance, NVDA still seems quite weak; regarding inference, who benefits from inference, NVIDIA's moat, and tighter chip bans, there is still much debate;

Regarding META's performance - AVGO vs NVDA vs Groq (?)

After META's performance, the buy-side has shown a clear higher interest in ASICs; mainly due to the mention of more custom chips (including replacing part of the training computing power) during META's earnings call;

A comment from BofA yesterday;

Our view on Meta's layout in the GPU and custom chip sectors is positive but needs to maintain an objective perspective (i.e., NVIDIA still has bullish logic). Of Meta's total capital expenditure of $65 billion for fiscal year 2025, the investment related to "custom chips" is only about $2-3 billion The proportion of Meta's capital expenditure of $60-65 billion remains very limited.

I haven't fully figured out this issue yet; in the short term, although ASICs have higher certainty (the shift from training to inference + continuous good news from various ramp-ups, such as CLS also hinting that OAI/AVGO may reach the scale of GOOGLE's business + no export control impacts like H20 + no capacity issues like BW); however, in the long term, if the market continues to pursue an open-source route, then more innovation leading to more architectural changes seems to suggest that a GPU cluster that is easier to scale up may be more flexible than ASICs?

As for Groq, which many people discussed last night, it doesn't seem to have changed much for now; there are many expert summaries about Groq TPU, and everyone has researched it quite a bit recently. JPM had an expert summary yesterday, and I'll organize it and share it later;

Morgan Stanley has raised its capex expenditure forecast;

Adjustments have been made for 25/26;

They also provided a breakdown for META:

Risk Warning and Disclaimer

The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial conditions, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investing based on this is at one's own risk