I rebuilt a position in Netflix. A price in the 70s is very attractive.

I continued buying Microsoft and TSMC.

I also added a small amount of Micron. If it falls below 850 later, I will increase the size of my additions.

I think Meta’s business route correction is necessary. Previously, the company created its own metaverse project, spending a quarter of its quarterly profit on it, only for everything to end up going nowhere. Then it moved into AI large language models, but fell from a top-three position to a level that is now barely worth counting. After that, it acquired several companies, especially Manus. Leaving aside how capable this company actually is, the fact that the deal became involved in government disputes and was then forced to be cancelled is already troublesome enough.

Therefore, based on multiple failures over the past several years, I think it is reasonable for the company to make a strategic contraction. After all, Meta’s core business is still advertising on social platforms, and Meta’s monopoly position remains very solid. From the current perspective, its need to apply AI is not that urgent. So for now, I personally view this only as a business adjustment by Meta itself. Before more detailed information and market data come out, there is no need to over-interpret it. Of course, after stocks have risen a lot, everyone needs some reason to lock in profits. That is completely normal.

As for whether the new Fed chair will raise interest rates, because of Trump’s reckless actions, it has now become impossible to judge. Although oil prices surged and then plunged, the disruption to supply chains has already happened. At the same time, we are also seeing a broad wave of price increases across memory and semiconductors. Consumer electronics will almost certainly become more expensive in the second half of the year, so I personally cannot determine whether inflation will continue to rise in the second half.

If inflation does rise, and the Federal Reserve really raises interest rates, then I would instead see this as proof that America’s system of checks and balances and separation of powers still exists, and that it can still restrain Trump. So even if stocks fall in the short term, in the long run this would absolutely be a good thing for the U.S. economy. I would be even more determined to add to my positions at lower prices.

If all these technocrats simply flatter Trump, go along with him, revise the data, or cut rates regardless of inflation, then they would have become nothing more than the president’s lapdogs. Even if the stock market rises in the short term, that would be the real misfortune for the U.S. economy. It would not be a good thing.

I have been a bit busy recently, so I still have not finished the Micron article. I will delay it for a few more days.

重新建仓奈飞，70多的价格很不错

继续买入微软，台积电。

稍微加了一点美光，如果之后能低于850，会增加加仓额度。

Meta的业务路线修正我认为是应该的，因为之前自己搞了个元宇宙，每个季度花掉四分之一的利润最后全打了水漂。之后又是做AI大模型，也从前三的水平跌落到数不上号。 然后又是收购了几家公司，特别是Manus这家公司，不说这家公司到底水平如何，单单是又卷入了政府纠纷，被迫再把交易作废，也是够头疼的。因此公司基于这么多年的多个失败，进行战略收缩是应该的，毕竟公司的主业是社交平台的广告业务，而Meta的垄断地位是很稳固的，所以目前来看对于AI的应用也并没有那么急迫。所以个人目前认为只局限于Meta单个公司的业务调整，在没有更多详细信息和市场数据出来前，并不需要过度解读。 当然涨的多了，大家都需要些理由落袋为安，很正常。

关于新主席加息与否的问题，因为川普胡搞，导致已经无法判断了，虽然油价暴涨后又暴跌了，但对于供应链的紊乱已经造成，以及本身又遇到了存储和半导体全面涨价潮，电子消费品下半年肯定还是要涨价的，所以我个人无法判断出下半年通胀是否会继续增加。

如果增加，然后美联储真的加息，那我反而认为这说明了美国的权力制衡和分立的制度依旧存在，依旧能对川普具有制衡力，所以即便短期股票会跌，但长期来看这绝对是有利于美国经济的好事儿，我会更加坚定的在低点去加仓。如果这些技术官僚全都趋炎附势地去附和川普，去修改数据或不管通胀就去降息，那就成了总统手下的走狗，即便短期股市涨的好，但那才是美国经济的不幸，并不是好事儿。

最近有点忙，美光的文章还没写完。再拖几天。

在1133美元又加了美光，目前美光的持仓占比已经达到50%。

357加仓微软，微软目前只有1.5%的仓位，目标3—5%。

别的都没有跌很多，不加仓。目前甲骨文占比13%。英伟达18%。台积电4%，高通不到1%，可忽略不计。

Microsoft’s current valuation is already very low, and I remain long-term bullish on its data center business. Micron’s earnings were very strong, and the company has signed multiple long-term contracts. I believe its current single-digit forward P/E is cheap.

I added more Micron at $1,133. Micron now accounts for 50% of my portfolio.

I added more Microsoft at $357. Microsoft currently accounts for only 1.5% of my portfolio, with a target allocation of 3–5%.

The other holdings have not fallen much, so I am not adding to them. Oracle currently accounts for 13% of my portfolio, NVIDIA 18%, TSMC 4%, and Qualcomm less than 1%, which is negligible.

Micron’s earnings were very strong, far exceeding expectations.

Revenue was $41.45 billion, up 74% quarter over quarter and 346% year over year. EPS was $25.1, up 106% quarter over quarter and more than 20% above market expectations, while gross margin reached 85%. Most importantly, in order to reduce cyclical risk, the company signed 14 five-year long-term SCA agreements.

The guidance for next quarter was also excellent: the company expects fourth-quarter revenue to reach a record $50 billion, plus or minus $1 billion.

Gross margin is expected to be around 86%, with operating expenses of $1.65 billion. EPS is expected to reach $31.

In addition, management believes the supply issue still cannot be resolved next year, so I personally think the current price is suitable for adding more. A more detailed analysis article and target price will be written later when I have time.

I previously posted that I had fully exited QCOM. Yesterday, new information came out showing that the company already has data center orders. As I said on June 1, I would add back based on the latest progress of its chips, so for now I am opening a small watch position first.

ORCL has fallen below the reasonable price level of $160 mentioned in my previous article and has now dropped to $150, so I am adding some more. I believe Microsoft and Oracle are currently stocks with the same concept, so choosing one of the two is enough.

美光业绩非常好，远超预期。

营收 414.5 亿美元，环比增长 74%，同比增长 346%，EPS 25.1 美元，环比增长 106%，超出市场预期 20%+，且毛利率达到85%。最重要的是，为了减弱周期性的风险，公司签署了14份5年的长期SCA协议。

下季度指引也非常优秀：预计第四财季营收将达到创纪录的 500 亿美元，上下浮动 10 亿美元。

毛利率约为 86%，营业费用为 16.5 亿美元。预计每股收益将达到31 美元。

并且公司管理层认为明年供应问题依旧无法解决，所以个人认为现在价格可以再加仓。具体的分析文章和目标价格之后会有空的时候写出来。

Qcom之前发过全清仓了，昨天透露出新的消息，公司已经有数据中心的订单了，按照之前6.1号所说的会根据芯片最新进展再加仓，因此目前先开个观察仓。

甲骨文跌破了之前文章里的合理点位160，目前已经跌到了150，因此再加一些。我认为微软和甲骨文目前是一个概念的股票，二选一即可。

Current positions:

Cash：held through SGOV: 20%

MU: 40%
This remains the stock I am most bullish on this year, and I continue to be optimistic about it. Even though the stock has risen significantly since I started buying it at USD 160 per share last September, I have not reduced my position. For the memory sector, DRAM is also an option.

NVDA: 20%
Because Micron had risen too much, I reduced part of my Nvidia position on June 4 in order to raise the cash position to 20%.

ORCL 15%
I remain bullish on Oracle over the long term, and I am keeping this allocation unchanged.

TSM: 4%
It remains the dominant monopoly leader. I continue to be bullish on it over the long term. In addition, as its factory construction in Japan and the United States has progressed smoothly, the geopolitical risk caused by excessive concentration in Taiwan is gradually being reduced.

MSFT: Less than 1%
At the current stage, I view Microsoft and Oracle as the same type of stock. Both are enterprise software companies with exposure to cloud services and data center demand.

The following are screenshots of the portfolio allocation changes that I previously posted in the Discord group between June 1 and June 4.

1. This screenshot shows the portfolio allocation I posted in the group a few days earlier, on June 1:

Texts in the screen shot:

“In the first half of this year, my account generated very substantial gains. Micron, which I started buying last September at around USD 160 per share with an average cost of around USD 250, has now risen fourfold to around USD 1,000 per share.

I remain bullish on Micron. However, because the position has already become very large and my cash allocation is currently not high enough, I decided to sell the lower-priority stocks and raise cash to 15%.

At the moment, I am only keeping the following four stocks:

Micron: 38%

Nvidia: 25%

Oracle: 18%

TSMC: 4%

Tesla needs a separate explanation. Personally, I believe that when the market is already at such elevated levels, the companies worth keeping should be those with real earnings support and valuations that are not excessively stretched.

For Tesla, although I am still very bullish on autonomous driving, it is still in the testing stage. In addition, the company’s current profitability is weak, and its valuation is too high relative to its earnings. If the market pulls back, Tesla may see a large decline. Therefore, I chose to sell it for now. It is not too late to buy it back later when the price becomes more reasonable, or when autonomous driving makes further progress. Also, over the past two years, Tesla has already doubled from my cost of around USD 200 to above USD 400, so it can still be considered a successful investment.

Qualcomm has already generated a 40% profit. Compared with the other companies I hold, it is still more of a future-expectation story. The progress of its AI chips may only become visible by the end of this year or next year, so I also chose to sell it first.

This year’s gains have far exceeded my expectations. Therefore, my current goal is to preserve cash so that I can hold large positions in Micron and Nvidia with greater peace of mind. For now, I have converted the other positions into cash.”

2. This screenshot shows that on June 4, I decided to raise my cash position to 20%.

   

Current View

Personally, I believe the market is now overheated. Almost every semiconductor-related stock has risen sharply. Even some small companies in very niche segments, such as AMKR, PDFS, MKSI, AEIS, and KEYS, have already moved up significantly. Established companies with monopoly-like positions, such as Texas Instruments and ASML, have also reached extremely high valuations, especially as they are starting to raise prices.

My stock purchases are usually made during sharp sell-offs. The only exception is a situation similar to March, when Micron was already trading around USD 400, but based on my analysis, I believed it could reach at least above USD 800 by year-end. Only when I see a high-probability chance of a stock doubling would I be willing to buy at that level.

However, recently I have researched a number of relatively less-followed stocks, and I can no longer find that kind of opportunity. Therefore, I am maintaining my June 4 view: keep 20% in cash. If there is no major pullback, and if the stocks I follow do not fall below the add-on levels I mentioned in my previous articles, I will not add to my positions again.

At the moment, the fundamentals of most semiconductor and AI companies are still not a major problem. The problem is simply that valuations are too high. For that reason, I believe keeping 20% in cash is enough. With cash reserved, there is no need to watch the market every day. I can rest, enjoy life, and take holidays.

If the market does not pull back to the prices I consider attractive, even if I have to wait several months, I will still not add to my positions. Of course, if there is no pullback for several months, and the companies later release another round of strong earnings, their valuations will naturally come down. If valuations fall to a sufficiently reasonable level, I would start buying again.

This year’s gains have already been more than enough. What I need from here is patience, and to wait for the right opportunities.

In addition, I have already made substantial gains this year, so I am not worried about a market decline. Even if the broader market were to pull back by 20% to 30%, it might still not be enough to erase this year’s profits. In that case, what is there to be worried about?

If the market really does fall that much, it would instead become another extremely exciting opportunity.

之前在群里发过自己的仓位，在这里还没发过，在此更新一下。

现在持有标的仓位：

现金（买了sgov）： 20%

美光：40% （今年最看好的股票，依旧看好，即便从去年九月160美元每股买入到现在增长很多，也没减过仓。存储类也可以买入DRAM。）

英伟达：20% （因为美光增长过多，因此在6.4号的时候为了达到20%现金比例，而减仓不少）

甲骨文：15% （长期看好，比例不动）

台积电：4% （依旧是垄断龙头，长期看好，且因为在日本和美国建厂顺利，过于集中于台湾导致的地缘风险在降低）

微软： 不到1% （和甲骨文目前是一类股票，都是企业软件+云服务和数据中心概念）

以下是之前6.1-6.4在Discord群中发的仓位变化截图：

1. 这张截图是前几天6.1号的时候发在群里的仓位：

“今年半年时间账户盈利非常多，美光由去年160美元每股买入，250的成本，如今已经翻了四倍到1000美元每股。 目前依旧看好，但因为仓位已经占了非常多，且现金仓位目前不够多，因此将优先级低的股票都卖出，留出15%的现金。

目前只留以下四支股票：

现在38% 美光，

25%英伟达，

18%甲骨文 ，

4% 台积电

需要特别解释的是特斯拉，个人认为在市场如此高位下，要保留的公司都是业绩有实质性的支撑的，且估值没有过度高估的。

而特斯拉，虽然仍然非常看好自动驾驶，但目前还在测试阶段，其次目前公司盈利能力弱，估值相较于盈利过高，如果回调会回调幅度大，所以暂时卖出，在价格合适或者自动驾驶有更进一步进展的时候再追不迟。而且最近两年已经200成本翻倍到400多，也算是成功的投资。

高通已经盈利40%，且相较于其他公司一样是仍在画饼，要明年或年底才能看到AI芯片的进展，因此也先卖出。

今年盈利远超预期，所以目标就是保留现金，为了安心持有大仓位的美光，英伟达。 其它的都暂时换为现金。”

1. 这在6.4号决定将现金仓位提高至20%

目前判断：

个人认为现在市场过于火热，所有和半导体相关的股票，哪怕是一些非常细分领域的小股票如AMKR,PDFS，MKSI， AEIS，KEYS等等都已经涨上了天，像是德州仪器，阿斯麦这种有垄断地位的老牌股更是因为要开始涨价，估值也高得离谱了。

我股票的买入基本都是在大幅下跌期间，除非和在三月份那样，（当时美光当时价格到了400）但我分析得出年底起码能到800以上，在有大概率翻倍的潜力下，我才会买入。

但最近我研究了多支比较冷门的股票，都找不到这种机会了，因此我维持6.4号的判断，保留20%现金，且如果之后没有大幅回调，没低于我前几篇文章里所发的个股的加仓点位的话，不会再加仓了。

雪中滚雪球

近半年回顾；及如有回调，下半年加仓点位

一、上半年的操作和逻辑…

Read more

a month ago · 25 likes · 2 comments · 雪中滚雪球

目前半导体和AI公司的基本面基本都没有大问题，只是估值太高，因此我认为留20%现金够了。留出现金，就不需要天天看股票，可以好好休息，娱乐，休假了。

假如一直不回调到我认为的价格，哪怕是要等几个月，我也不会再加仓的，当然如果几个月都不回调，等到新的优异的财报发布，估值自然会降下去，如能降到足够合理，也是会再开始买入的。今年已经盈利足够多了，之后需要的就是耐心，等待好的机会。

并且今年都赚了这么多，我也不怕下跌，即便大盘回调20%-30%都不一定能抹掉今年的盈利，那还有什么可怕的呢。如果能下跌这么多，反而又是令人兴奋的绝佳机会。

In March this year, when the stock market correction was at its most intense, I noticed that the valuations of many semiconductor stocks had already become highly attractive. Even though the earnings performance of these companies generally far exceeded market expectations, their share prices still saw a meaningful pullback due to the foolish and pointless war initiated by Trump. It was precisely during this period that I chose to increase my positions substantially.

Looking back, the most correct decision I made this year was to continue buying Micron during the market panic. As I wrote previously, from last September to this March, I had already been continuously adding to my Micron position in the range of USD 160 to USD 350 per share, raising its weighting in my portfolio to around 7%. However, when Micron’s earnings remained strong, its fundamentals had not deteriorated, and yet its share price fell from above USD 400 to around USD 300 due to the impact of the war, I believed this was indeed a rare opportunity and decisively continued to increase my position.

Of course, this was also because my own analysis suggested that Micron’s fair value could reach around USD 1,000 by year-end, which gave me greater conviction. For details, please refer to my past article in Chinese: March 18, 2026 — Micron Valuation.

At the same time, I also made a clear asset allocation decision: to exit stocks with weaker relevance to AI and further concentrate capital into AI-related names.

Below are the trading records I posted in real time in the group at that time. Since there is no review or censorship issue here, screenshots of future add-on purchases will be posted in real time on Substack.

Because of its excellent earnings performance, Micron had already risen to USD 400 in February this year. At that time, I also made a small purchase of NVO. As can be seen, Nvidia was only around USD 170 at the time, TSMC was only around USD 330, and Oracle had already fallen from USD 300 to USD 156. I also kept increasing the size of my add-on purchases in Oracle all the way down from around USD 300 to USD 156.

As shown in the chart below, by late February, I had already judged that Oracle below USD 170 was a cheap valuation level at which I could increase my purchases more aggressively. I also bought AMD at around USD 200. To be honest, I had not conducted particularly detailed research on the company at that time, but the price looked simply too cheap, so I initiated a position. The stock later rose to above USD 500.

I also bought TSMC at a little over USD 300, which subsequently climbed to above USD 400. In addition, I substantially increased my position in Oracle below USD 150, averaging down the cost basis of a position that had been in loss for half a year to around USD 180. As a result, when Oracle later rose to around USD 250 last week and I partially reduced the position, the trade also generated a meaningful profit.

I also exited the non-AI stocks I had previously bought and further concentrated my capital into AI-related names. Even though my PayPal position was down 35%, I still chose to close it and reallocate the capital toward AI stocks.

After Micron released its earnings report, I saw that despite such outstanding results, the stock continued to fall back to around USD 300 due to the market’s fear over the war. I therefore continued to add to my position.

Trump then made another round of reckless and meaningless remarks, triggering another sell-off and once again creating an attractive opportunity to add to positions. I continued buying TSMC and Micron.

I continued to be bullish on the memory sector. My view was not limited to Micron alone; I also bought DRAM.

However, after entering May, the situation began to change. Almost every AI semiconductor company that had not yet rallied started to surge on a daily basis. A clear example was that I had taken small positions in TTMI and Qualcomm at the time, and both stocks rose by around 40% to 50% in less than a month.

Therefore, after U.S. equities had moved to elevated levels in May, I found that there were no longer any genuinely cheap opportunities left. On May 11, I posted the chart below in the group, stopped adding to positions, and began reducing exposure.

On June 1 and June 4, I posted my portfolio allocation in the Discord group and clearly pointed out that U.S. equities had already reached elevated levels. I said that it was necessary to sell part of the portfolio and raise the cash position to 20%, as shown in the charts below.

At that time, I could already clearly feel that the market was becoming overheated. Stocks were rising on almost any piece of news, and many investors who had missed the add-on opportunities in March and April were beginning to FOMO back into the market. My profit on Micron alone had already exceeded 350%, so I believed it was time to raise the cash allocation and wait for the pullback that would inevitably come at some point, even though I did not know exactly when.

Looking back at last year’s operations, I did indeed make many mistakes. For example, I bought Oracle at too high a price. I also bought PayPal and DUOL simply because they looked cheap, but both positions ended up falling by more than 30% to 50%. However, I still strictly controlled the position sizes at around 5% to 7%, which limited the losses. More importantly, during this year’s sharp correction, I decisively sold the non-AI stocks and rotated the capital into AI-related names, which ultimately generated substantial gains.

Of course, many readers have also asked why I only kept 20% in cash. Since the market fell so much last Friday, why not keep 50% in cash? The reason is simple. After entering June, I already believed that the stock market was overheated and that the probability of a correction had increased significantly. But I did not know exactly when the correction would happen, or whether it would happen at all. In addition, the latest earnings results of the companies I own still remained far above expectations. Therefore, I judged that having more than 20% in cash was sufficient. Given the gains I had already accumulated, I was not afraid of a pullback in profits. Even if my profits were to retrace by 30%, I would still have a level of return that I consider satisfactory.

II. Potential Add-on Levels in the Second Half if a Major Pullback Occurs

Looking at the broader market, I personally believe that if the Nasdaq were to pull back to its 60-day moving average, it would become highly attractive to me. Of course, this is because my current equity exposure is still around 80%, and it does not necessarily apply to other investors’ situations. Moreover, whether the market can actually correct that much is something no one can know in advance.

In addition, if the individual stocks I remain bullish on fall below the following prices, I would also consider adding to my positions.

Micron: USD 650

Oracle: USD 180 （I have already added to the position once on June 16. Personally, I allocated the remaining 20% cash position according to the weighting of the stocks I currently hold, and then divided that amount into six separate add-on purchases. The screenshot below shows the add-on transaction.）

Microsoft: USD 390 （Also added to the position once）

Nvidia: USD 190

TSMC: USD 350

Tesla: USD 350

Google: USD 330
I have recently started paying closer attention to Google, largely because of Berkshire Hathaway’s significant increase in its position.

At present, although there is still a threat of rate hikes caused by Trump’s reckless policy actions, the earnings performance of leading companies remains very strong. If there is significant short-term volatility, there may be opportunities to pick up some of the above names at attractive prices.

The companies I currently like, and the prices at which I think it is relatively reasonable to add positions, remain unchanged:

My current position level is still 80%.

Micron: $650

Microsoft: $390

Nvidia: $190

TSMC: $350

Tesla: $350

Oracle: $180
(I have already added once. The screenshot was posted on Substack in real time.)

Google: $330

This article mainly shares some of my thoughts on SpaceX’s IPO and the company’s future prospects. Those who have not subscribed can subscribe for free, and future articles will be sent by email.

First, I want to make one thing clear: I have always avoided buying newly IPO’d companies, and I also avoid buying companies that are still unable to make a profit. There are several reasons for this.

First, a company can choose when to go public. Most companies will choose to IPO when market sentiment is strong, overall valuations are high, and investors are optimistic. This allows the company and the investment banks to raise more money, and it also makes it easier to price the offering higher. Therefore, from a probability perspective, the valuation at IPO is often not cheap, and the market’s expectations for the company are usually already overdrawn in advance. Under this kind of situation, it is very difficult for me to buy at the early stage of listing and hold for the long term.

Second, for a company that is not yet profitable, valuation itself is very difficult. Without stable profits, it is hard to use traditional methods such as PE, profit margins, or free cash flow to determine whether the company is cheap or expensive. In the end, it easily becomes a game of stories, imagination, and market sentiment. Under my personal investment framework, I usually do not buy this kind of company.

Now, back to SpaceX.

This article is mainly about my personal subjective view. On the one hand, SpaceX is still losing money. On the other hand, Musk’s personal ability, influence, charisma, and ability to sell big visions are all extremely strong. This means the company’s share price will be very difficult to judge purely through valuation. Therefore, I also cannot give SpaceX a clear valuation. What follows is more of an analysis from the perspectives of the company’s development, Musk’s change in attitude, business prospects, and risks.

1. Musk’s changing attitude toward a SpaceX IPO

Before discussing whether SpaceX is worth buying, I think it is necessary to revisit the company’s development process, as well as Musk’s changing attitude toward taking SpaceX public.

I asked AI to organize Musk’s past public views and timeline regarding a SpaceX IPO. The rough timeline is as follows.

From 2013 to 2021, Musk was clearly opposed to taking SpaceX as a whole public.

In 2013, Musk sent an email to SpaceX employees about “Going Public”. The core message was very clear: he was worried about SpaceX going public before the “Mars transportation system” had matured. This is because the fundamental goal of SpaceX is not simply to make money, but to build the technology that allows human beings to survive on Mars. If going public would reduce the likelihood of achieving that goal, then SpaceX should not go public.

His reasoning mainly had four parts.

First, the public market would turn the company into one driven by quarterly results. For a company like SpaceX, launch delays, rocket failures, repeated changes in technical routes, and projects being pushed back or restarted are all part of the business itself. But if it becomes a public company, every accident and every failure could bring a sharp fall in the share price, short-selling pressure, media pressure, and investor pressure. In his 2013 email, Musk also mentioned that the stock price of a public company could fluctuate violently due to execution problems or macro factors, distracting employees with share price movements instead of allowing them to focus on the product and engineering itself.

Second, going public would bring additional compliance costs and legal risks. He specifically mentioned that public companies have to deal with regulatory requirements such as Sarbanes-Oxley, as well as class-action lawsuits and insider trading restrictions. In other words, he was not simply unwilling to share the returns. He believed that going public would drag management and employees into a large amount of work unrelated to engineering or product development.

Third, he believed that Tesla and SolarCity going public at the time was actually a “had no choice” decision, not an ideal choice. In his 2013 email, he said Tesla and SolarCity went public because their capital structures and financing needs had become complex, requiring large amounts of equity and debt financing. But if SpaceX could still solve its funding and liquidity needs through private financing and employee liquidity windows, then there was no need to bear the pressure of the public market.

Fourth, he wanted to give employees “public-company-like liquidity” while avoiding the short-term pressure of the public market. In other words, early SpaceX chose to remain private while providing some liquidity for employees and early investors through internal liquidity windows, private financing rounds, and valuation increases.

By 2020 to 2021, Musk began to soften his stance on a separate Starlink IPO, but this still did not mean taking SpaceX as a whole public.

Around 2020 and 2021, he mentioned several times that Starlink might IPO in the future, but only after cash flow and revenue growth became sufficiently smooth and predictable. In other words, Musk was not completely opposed to going public. What he opposed was going public when the business was not mature, cash flow was not predictable, and technical risks remained very high.

In May 2025, Musk was still publicly saying that SpaceX was “not in a rush to go public”. But only a short period of time passed between that statement and his later confirmation of the IPO plan.

On December 10, 2025, Musk posted a short message on X confirming that space reporter Eric Berger’s report about SpaceX planning a 2026 IPO was accurate. This was the first time in many years that he formally softened his stance on taking SpaceX as a whole public. CNBC later also reported that Musk personally confirmed that the report about “SpaceX going public in 2026” was accurate, and he responded by saying, “As usual, Eric is correct.”

In February 2026, SpaceX completed the acquisition of xAI at an extremely high valuation, combining Musk’s AI business with his space business. This was a key move in reshaping the company’s story before the IPO.

On April 1, 2026, SpaceX confidentially filed its S-1 prospectus with the SEC, officially starting the IPO process. The target valuation was around $1.75 trillion to $2 trillion, with a planned fundraising size of $75 billion.

On May 20, 2026, SpaceX publicly released its IPO prospectus, which covered financial data, Starlink progress, and future AI and Mars-related plans.

In June 2026, SpaceX began its roadshow and planned to list on Nasdaq at $135 per share under the ticker “SPCX”, with a valuation of around $1.77 trillion.

From “confirming it would go public” to “officially going public”, the whole process took only about half a year. From confidentially filing with the SEC to starting the roadshow, it took only around two months. The pace was extremely fast. At the same time, Anthropic and OpenAI were also planning IPOs in the near future. I do not know whether this is just a coincidence, or whether Musk also wanted to push SpaceX to the market as quickly as possible while AI sentiment is at its hottest.

I have been following this company for the past five or six years. Musk’s long-standing narrative has been that Tesla had no choice but to go public because it needed financing; Tesla had to make money, and its commercial success would support SpaceX, a company closer to idealism, less focused on short-term profit, and even aimed at opening up a new option for human survival.

I have always agreed with this view. A company like SpaceX, which carries ideals and long-term goals, and which often experiences launch failures, project delays, and technical routes being overturned and rebuilt, is not really suitable for a group of investors who only care about short-term profits and do not truly understand aerospace engineering or long-term goals to stand behind it and point fingers.

This kind of company should stubbornly follow Musk’s own ideas and move in the direction he believes is right, exploring the universe and a new world.

But in the past six months to a year, his thinking has taken a very rapid turn. In my view, this is a fairly rushed pivot. He quickly inserted several big narratives into SpaceX that could logically be connected to future profits, such as xAI and space-based computing, and then used this wave of AI enthusiasm to turn SpaceX into one of the largest IPOs in history by valuation.

Of course, in this IPO, Musk still retains the overwhelming majority of control through the company’s share structure. Investors basically can only put in money, but will find it very difficult to truly participate in company decision-making. In other words, although SpaceX is going public, it is not a public company in the ordinary sense.

2. Why I remain bullish on SpaceX in the long term

Although I remain cautious about this IPO, I still believe SpaceX is a company that I will very likely consider buying in the future.

First, this company is truly pushing forward something great and meaningful. It is advancing human access to space, lowering the cost of space transportation, building a global satellite communication network, and even exploring whether human beings may one day migrate to other planets. To put it plainly, if any one of these big visions can be realized well, it would be enough to go down in history and bring enormous economic value.

Second, at present, SpaceX has almost monopolized the space transportation business at the corporate level, and in many areas it has already surpassed the national-level aerospace capabilities of countries such as China and the United States. Reusable rockets, launch frequency, cost control, and engineering iteration speed are all part of SpaceX’s moat.

Third, Starlink itself is already a very successful and excellent business. It is no longer just a concept. It already has real users, real revenue, and real application scenarios. Whether it is communications in remote areas, aviation and maritime connectivity, battlefield communications, or future integration with ground communication networks, Starlink has huge potential.

So I am not bearish on SpaceX. On the contrary, I believe it is a company I will very likely consider buying in the future. But the greater the potential of a company, the more important it is not to blindly chase it when sentiment is at its hottest. Especially when the company is not yet profitable and the valuation is already very high, it is even more important to observe first.

3. Several issues that need to be watched carefully

First, is SpaceX going public a decision Musk made after deep consideration, or is it another impulsive move?

Musk is a technological genius. He is also a person with extraordinary execution ability and imagination. But at the same time, he is someone who loves attention, cannot stand being quiet for too long, and often begins looking for the next goal and the next big headline before a business has fully matured.

Twitter is a typical example. He acquired Twitter at a clearly inflated price. Later, after taking over the company, he found that it was not worth that price at all, but by the time he wanted to regret it, it was already too late. Even now, I still have not seen him truly take Twitter, now X, to a new level.

At the same time, while he was busy messing around with Twitter, Tesla’s new vehicles repeatedly faced production delays, the low-cost model kept being postponed, the Cybertruck had high production costs and many small problems, and sales still have not been impressive. The electric truck has also been delayed again and again, and the new battery he talked about for so long has still not been fully delivered.

More importantly, while these problems remained unresolved, he suddenly turned toward politics, believing that he could single-handedly change government efficiency. He supported Trump, participated in government layoffs, and in the end I did not see any real results. Instead, he and Trump parted ways, attacked each other publicly, and damaged his own public image and reputation.

So I believe Musk is a technological genius, but he is not a mature, stable, and predictable CEO in the traditional sense. His upside is extremely high, but his uncertainty is also extremely high.

Therefore, is this SpaceX IPO a strategic choice made after deep consideration, or is it another rapid pivot, rapid hype-building, and rapid financing move during the AI boom? I think this requires caution. It is better to observe first.

Second, SpaceX’s short-term profit outlook is still unclear.

In order to connect a rocket company to the AI boom, Musk reshaped SpaceX’s narrative in a very short period of time.

From a timeline perspective:

In late October 2025, Musk for the first time made a relatively clear public statement that SpaceX would work on orbital data centers.

In January 2026, he only then upgraded space-based computing into one of the company’s important goals.

By the time of the IPO in 2026, AI computing had already become an important part of the SpaceX listing story.

I am not against the direction of space-based computing. From a technical and logical perspective, it does have a chance of working. Space has more stable solar energy resources. Ground-based data centers face constraints in electricity, land, water resources, and approvals. If SpaceX can greatly reduce launch costs through Starship, and combine that with Starlink’s satellite network and laser links, then in theory it does have an opportunity to combine communications, computing power, and AI infrastructure.

But the question is: when will this business actually become profitable? How large is the profit potential? How much money needs to be invested upfront? How many years will it take to see returns? These questions are currently very unclear, and there is no comparable case to refer to.

More importantly, I strongly disagree with Musk’s decision to merge xAI into SpaceX. My understanding is that he did this to significantly increase SpaceX’s valuation by inserting an AI business into the company, allowing SpaceX to ride the AI concept. Otherwise, if SpaceX only had Starlink, space transportation, and space-based computing that is still at the PowerPoint stage, it would be very difficult to support such a high valuation today.

Personally, I think Grok is already a lagging large language model. It is also possible that Musk believes xAI needs more money in order to catch up with its competitors, so he merged it into SpaceX and used SpaceX’s IPO to raise more capital. But I do not think competition in large language models is something that can be won simply by having money. This has already become one of the most competitive fields in the world, and xAI has no clear advantage. OpenAI, Anthropic, and Google have already invested enormous amounts of capital in this field and have built advantages in data, products, ecosystems, and talent.

xAI will likely consume a large amount of SpaceX’s capital in the future. But if its only core product is a large language model, it will be very difficult for it to bring meaningful profits to SpaceX. In comparison, I believe SpaceX should focus more on Starlink, space transportation, and Musk’s newly painted big vision — space-based computing. These are the areas where SpaceX truly has a chance to form a monopoly and build a core competence that it has and others do not. It should not be competing head-on with OpenAI, Google, and Anthropic in large language models.

Third, the universe has unlimited potential, but for now there is still no real “spice”.

I have always felt that SpaceX is somewhat like the Dutch East India Company during the Age of Exploration. It is facing a huge, unknown new world full of imagination. Just like in the Age of Exploration, the reason European countries and merchants were willing to keep exploring, taking risks, and investing was because there were enormous profits behind it.

Back then, the company was trying to break the monopoly that Muslim powers had over land-based trade routes, take the sea route to obtain spices, and make huge profits from it.

At that time, shareholders who invested in the Dutch East India Company saw the stock double in the second and third years. For the next hundred years, it paid dividends of 20% or more every year, and the company operated for around 200 years before being taken over by the state. It was truly the kind of investment where an ordinary person could buy the stock, become wealthy for life through the stock price increase, and support several generations of descendants.

But the problem is that, so far, there is nothing in space that has been proven to be similar to “spice”. Mining resources on other planets is not a commercial issue that can be seriously considered for at least the next five years. Resources on Earth are still sufficient. Human beings do not yet have a strong economic motivation to mine on other planets. Unless we discover some kind of mineral on another planet that does not exist on Earth and has extremely high use value, then if it is merely the same type of resources we already have on Earth, such as oil or gold, Earth’s existing reserves are enough for a long time. Our generation does not need to worry about this, and there is no real need to go to another planet for it.

So, for SpaceX, what can support its valuation in the short to medium term is still Starlink, space transportation, government and military orders, and possibly space-based computing in the future. Mars, extraterrestrial mining, and a multi-planet civilization are all great visions, but at present they are very difficult to turn into profits that can be valued, predicted, and realized. It will be very difficult for our generation to see that happen.

If the stock market experiences a major correction, these grand ideals will be pulled back to the surface of the Earth together with the share price by gravity. Of course, that may also become a good buying opportunity.

Fourth, although space-based computing makes sense logically, the financing pressure could be enormous.

Space-based computing is Musk’s newly painted big vision, but this vision is not completely illogical. Technically, it does have a chance of working, and it also fits the AI era’s massive demand for electricity and computing power.

The question is whether the profit potential is large enough to support SpaceX’s continuous investment. How long will it take to become profitable? How much money will need to be invested before profitability? Will the funds raised in this IPO be enough? Will the company need to issue more shares, borrow money, or issue bonds in the future to support the research and construction of space-based computing?

Look at Oracle. It is only building data centers on Earth, and it already needs to invest enormous capital and take on huge debt pressure. If data centers are built in space, the costs and uncertainties will only be higher.

In space, heat dissipation, maintenance, radiation, satellite lifespan, launch costs, disposal cycles, network latency, and limitations on task types will all bring uncertainty. My personal ability is limited, and I cannot determine whether space-based computing will eventually form a sufficiently large commercial closed loop. Therefore, on this issue, I choose to observe first.

4. Conclusion

SpaceX is undoubtedly a great company. It carries humanity’s imagination about exploring the universe, and it may even change history. It is somewhat like a new-era Dutch East India Company, with the potential to form a monopoly in an unknown, vast, and largely undeveloped field — even the potential to monopolize space. That is also why it can reach a valuation of around $1.7 trillion while still losing money.

But being great does not mean it is worth buying right now. A great company, if bought at the wrong price, can still bring a long period of pain.

I remain cautious about SpaceX for three main reasons.

First, I am skeptical about Musk’s relatively rushed decision to take the company public. I am not sure whether this is a deeply considered strategic arrangement or another rapid pivot during the AI boom.

Second, I am not optimistic about Grok and xAI being merged into SpaceX. I do not think large language models are SpaceX’s strongest area. Instead, xAI may consume a large amount of the company’s capital.

Third, I remain cautious about the profitability and profit timeline of space-based computing. It is attractive from a technical logic perspective, but it may still be a long way from forming predictable profits.

Because the company is not yet profitable and there are so many uncertainties, SpaceX is currently destined to be very difficult to analyze through traditional valuation methods. It is even harder to analyze than Tesla. Whether it surges or collapses after listing is something no one can accurately predict.

My personal view is very simple: wait a few months, or even half a year to a year, and then reassess. I will continue to follow it closely, because this is indeed a company that is very worth studying and may also be worth investing in. But at the same time, it is also a company full of risk, full of uncertainty, and extremely dependent on Musk’s personal judgment.

For a company like this, I do not want to enter when market sentiment is at its hottest and valuation is at its highest. I would rather wait until the financial data and business progress become clearer before seriously considering whether to buy.

For me, steadily capturing Micron and DRAM’s relatively more certain upside over the next six months to one year is far more rational and much safer than buying SpaceX now and betting on a twofold or threefold return.

我目前看好的公司，以及我认为相对合理的加仓价格保持不变：

我个人目前仓位依旧是80%。

美光：650美元

微软：390美元

英伟达：190美元

台积电：350美元

特斯拉：350美元

甲骨文：180美元 （已加仓一次，截图实时发在substack了）

谷歌：330美元

本文主要写一下对于SpaceX上市以及公司未来前景的一些看法。没有订阅的可以免费订阅一下，之后新的文章将会邮箱通知。

首先要说明的是，我个人一直以来不买刚IPO的公司，也不买仍然无法盈利的公司，主要有以下几点原因。

第一，公司选择IPO的时间，本身就是可以主动安排的。绝大多数公司都会选择在市场热度较高、整体估值偏高、投资者情绪比较乐观的时候上市。这样公司和投行才能拿到更多钱，发行价格也更容易打高。因此，从概率上说，IPO时的市值往往并不便宜，市场对公司的期待也容易被提前透支。这种情况下，我很难在上市初期就买入并长期持有。

第二，对于还没有盈利的公司，估值本身很难做。没有稳定利润，就无法用PE、利润率、自由现金流等传统方法判断公司到底是便宜还是贵。最后很容易变成只看故事、只看想象空间、只看市场情绪。对于我个人的投资体系来说，这种公司我通常不会买。

回到SpaceX。

这篇文章主要是写我个人的主观看法。因为SpaceX一方面仍处在亏损状态，另一方面马斯克的个人能力、影响力、魅力，以及画饼能力都非常强，这会导致公司股价很难单纯通过估值来判断。所以我也无法对SpaceX给出一个明确估值，下面更多是从公司发展、马斯克态度变化、业务前景和风险几个角度来分析。

一、马斯克对SpaceX上市态度的变化

我认为，在讨论SpaceX是否值得买之前，很有必要重新回顾一下这家公司的发展过程，以及马斯克对于SpaceX上市这件事的态度变化。

我让AI整理了一下马斯克过去对于SpaceX上市的公开态度和时间线，大致如下。

2013年至2021年，马斯克一直明确反对SpaceX整体上市。

2013年，马斯克曾给SpaceX员工发过一封关于“Going Public”的邮件，核心观点非常清楚：他担心SpaceX在“火星运输系统”成熟之前上市。因为SpaceX的根本目标不是简单赚钱，而是建立让人类可以在火星生存的技术。如果上市会降低这个目标实现的可能性，那就不应该上市。

他的理由主要有四点。

第一，公开市场会让公司变成季度业绩驱动。SpaceX这种公司，发射延迟、火箭异常、技术路线反复、项目推倒重来，本来就是业务的一部分。但如果成为上市公司，每一次事故、每一次失败，都可能带来股价暴跌、做空压力、媒体压力和投资者压力。马斯克在2013年的邮件中也提到，公众公司股价会因为执行问题或宏观因素剧烈波动，从而让员工被股价分散注意力，而不是专注于产品和工程本身。

第二，上市会带来额外的合规成本和法律风险。他特别提到，上市公司需要面对Sarbanes-Oxley等监管要求，也要面对集体诉讼、内幕交易限制等问题。换句话说，他并不是单纯不想分享收益，而是认为上市会把管理层和员工拖进大量非工程、非产品的事务中。

第三，他认为Tesla和SolarCity当年上市，其实是“没办法”的选择，而不是理想选择。2013年的邮件中，他说Tesla和SolarCity之所以公开上市，是因为资本结构和融资需求已经变得复杂，需要大量股权和债务融资。但SpaceX如果还能通过私募融资、员工流动性窗口来解决资金和流动性问题，就没有必要承受公开市场的压力。

第四，他希望给员工“类似上市公司的流动性”，但避免公开市场的短期压力。也就是说，早期SpaceX采取的是：保持私人公司身份，同时通过内部流动性窗口、私募融资和估值上调，让员工和早期投资人获得一定流动性。

到了2020年至2021年，马斯克开始对Starlink单独IPO松口，但仍然不是SpaceX整体上市。

他在2020、2021年前后多次提到，Starlink未来有可能IPO，但前提是现金流和收入增长变得足够平滑、足够可预测。也就是说，马斯克并不是完全反对上市，而是反对在业务不成熟、现金流不可预测、技术风险仍然很大的时候上市。

2025年5月，马斯克还公开表示SpaceX“不急着上市”。但这句话距离他后来正式确认IPO计划，只过去了很短时间。

2025年12月10日，马斯克在X上发了一条简短推文，确认太空记者Eric Berger关于SpaceX计划2026年IPO的报道属实。这是他多年来第一次对SpaceX整体上市正式松口。CNBC随后也报道，马斯克亲口证实“SpaceX 2026年上市”的报道是准确的，并回应称“像往常一样，Eric是正确的”。

2026年2月，SpaceX以极高估值完成对xAI的收购，将马斯克的AI业务与太空业务整合到一起。这是IPO前夕重塑公司叙事的关键动作。

2026年4月1日，SpaceX向SEC秘密提交S-1招股文件，IPO进程正式启动，目标估值约1.75万亿至2万亿美元，拟募资750亿美元。

2026年5月20日，SpaceX正式公开IPO招股书，内容涵盖财务数据、Starlink进展，以及未来AI和火星业务布局。

2026年6月，SpaceX启动路演，并计划以每股135美元在纳斯达克挂牌交易，股票代码为SPCX，估值约1.77万亿美元。

从“确认要上市”到“正式上市”，整个过程只用了大约半年。从秘密提交SEC文件到启动路演，也只用了大约两个月，节奏非常快。而且同期Anthropic和OpenAI也计划在近期IPO，我不清楚这只是巧合，还是马斯克也希望趁着AI最火热的时候，尽快把SpaceX推向市场。

我最近五六年一直关注这家公司。马斯克长期以来的叙事是：特斯拉是为了融资而不得不上市，特斯拉要赚钱，要用商业上的成功去养活SpaceX这种更接近理想主义、更不以短期盈利为目的、甚至是在为人类生存开拓新选择的公司。

我一直也很认同他的这个看法。SpaceX这种承载理想和远大目标的公司，而且经常会经历发射失败、项目延期、技术路线推倒重来的公司，本来就不太适合让一堆只想着短期赚钱、并不真正理解航天工程和长期目标的投资者站在背后指手画脚。

这种公司本来就应该固执地坚持马斯克一个人的想法，朝着他认为正确的方向去探索宇宙和新世界。

但最近半年到一年，他的想法出现了非常快速的转弯。在我看来，这是一次相当仓促的转向：他快速给SpaceX塞进了几个逻辑上能够赚钱的大饼，比如xAI、太空算力，然后结合这一轮AI热潮，把SpaceX推成了史上市值最高的IPO之一。

当然，这次IPO中，马斯克通过股权架构设计，依旧保留了绝大多数控制权。投资者基本只能投钱，而很难真正参与公司决策。也就是说，SpaceX虽然上市了，但它并不是一家普通意义上的公众公司。

二、为什么长期看好SpaceX

虽然我对这次上市保持谨慎，但我认为SpaceX依旧是一家未来很大概率值得买入的公司。

首先，这家公司确实在推进一件伟大且有意义的事情。它是在推进人类进入太空、降低太空运输成本、建设全球卫星通信网络，甚至探索未来人类是否有可能移居其他星球。说白了任何一个饼能画好，都是能载入史册的，也能带来很大的利益。

其次，目前来看，SpaceX几乎是在公司层面垄断了太空运输业务，而且在很多方面已经远超了中美这种国家层面的航天能力。可回收火箭、发射频率、成本控制、工程迭代速度，这些都是SpaceX的护城河。

再次，Starlink本身已经是一个非常成功且优秀的业务。它不再只是一个概念，而是已经拥有真实用户、真实收入和真实应用场景。无论是偏远地区通信、航空航海、战场通信，还是未来和地面通信网络的结合，Starlink都有很大潜力。

所以，我并不是不看好SpaceX。相反，我认为它是我未来很大概率会考虑买入的公司。但越是潜力巨大的公司，越不能在最热的时候无脑追高。尤其是当它还没有盈利、估值又已经很高的时候，更应该先观察。

三、需要警惕的几个问题

第一，SpaceX上市到底是马斯克深思熟虑后的决定，还是又一次一时兴起？

马斯克是一个技术天才，也是一个极具行动力和想象力的人。但他同样是一个爱出风头、不甘寂寞，并且经常在一个业务还没有完全成熟时，就开始寻找下一个目标和下一个大新闻的人。

Twitter就是一个典型例子。他当年以明显偏高的价格收购Twitter，后来接手后发现这家公司根本不值那个价，想反悔时已经为时已晚。到现在，我也没有看到他真正把Twitter，也就是现在的X，带到一个新的高度。

与此同时，在他折腾Twitter的过程中，特斯拉的新车接连难产，低价车型一直延期，Cybertruck造价高昂、小问题不断，销量至今也不理想，电动卡车持续延期，新电池也一直没有完全兑现。

更重要的是，在这些问题还没有解决时，他又突然转向政治，认为自己可以凭借一己之力改变政府效率问题。他站台特朗普，参与政府裁员，最后并没有看到什么实际结果，反而和特朗普分道扬镳，互爆猛料，导致自己的公众形象和名誉受损。

所以我认为，马斯克是一个技术天才，但他不是一个传统意义上成熟、稳定、可预测的CEO。他的上限极高，但不确定性也极高。

因此，这次SpaceX上市到底是深思熟虑后的战略选择，还是他在AI热潮中又一次快速转向、快速造势、快速融资？我认为需要保持警惕，先观察再说。

第二，SpaceX短期的盈利前景并不清晰。

马斯克为了让火箭公司搭上AI热潮，在很短时间内重塑了SpaceX的叙事。

从时间上看：

2025年10月底，马斯克第一次比较明确地公开表示SpaceX会做轨道数据中心。

2026年1月，他才把太空算力升级成公司的重要目标之一。

到了2026年IPO期间，AI算力就已经变成SpaceX上市故事中的重要组成部分。

我并不反对太空算力这个方向。从技术和逻辑上说，它确实有可能成立。太空中有更稳定的太阳能资源，地面数据中心面临电力、土地、水资源和审批限制，而SpaceX如果能通过Starship大幅降低发射成本，再结合Starlink卫星网络和激光链路，理论上确实有机会把通信、算力和AI基础设施结合起来。

但问题在于，这个业务到底什么时候能真正盈利？盈利空间到底有多大？前期要投入多少钱？需要几年才能看到回报？这些问题目前都非常不清楚，而且没有任何可参考案例。

更重要的是，我非常不赞同马斯克把xAI并入SpaceX。我的理解是，他是为了大幅提升SpaceX的市值，把AI业务塞进来，从而让公司搭上AI概念。否则，如果只有Starlink业务、太空运输业务，以及还停留在PPT阶段的太空算力，很难支撑现在这么高的市值。

我个人认为，Grok目前已经是一个落后的大模型。也可能正是因为马斯克认为xAI需要更多钱，才能让模型重新追上竞争对手，所以才把它并入SpaceX，通过SpaceX上市筹集更多资金。但我认为，大模型竞争并不是单纯有钱就一定能赢，这个领域已经成了竞争最激烈的地方，而xAI没有任何优势。OpenAI、Anthropic、Google、都已经在这个赛道上投入了巨额资金，并且形成了数据、产品、生态和人才优势。

xAI未来会消耗SpaceX大量资金，但如果它只有大模型这一款核心产品，很难给SpaceX带来可观利润。相比之下，我认为SpaceX应该更专注于Starlink、太空运输，以及马斯克新画的这个大饼——太空算力。因为这些才是SpaceX真正有机会形成垄断、有自己有，但别人无的核心竞争力的领域，而不是去和OpenAI、Google、Anthropic正面竞争大模型。

第三，宇宙潜力无限，但目前还没有真正的“香料”。

我一直觉得，SpaceX很像大航海时代的荷兰东印度公司。它面对的是一个巨大、未知、充满想象空间的新世界。就像当年的大航海时代，欧洲国家和商人之所以愿意持续探索、冒险、投入，是因为背后有足够丰厚的利润。

当年公司是打破穆斯林国家对陆上贸易的垄断，走海路去获得香料，并从中赚取巨大收益。

而当年投资荷兰东印度公司的股东们，第二，第三年股票就翻倍，且之后长达一百年，每年分红20%或以上，然后公司经营了200年才被国家收购。的确是随便一个人买了，就可以靠股票的增幅富裕一辈子，养活几代人。

但问题是，目前宇宙里还没有被证明存在类似“香料”的东西。去外星球开采矿物，至少在未来五年内都不是一个能认真考虑的商业问题。地球上的资源依旧足够，人类并没有强烈的经济动机去外星球挖矿。除非我们在外星球发现了某种地球上没有，且使用价值极高的矿物，否则仅仅是和地球一样的资源，比如说石油，黄金这些，地球上的存量足够使用很久，我们这几代人都不需要 担心，完全没有去外星球的必要。

所以，对SpaceX来说，真正能在中短期支撑估值的，仍然是Starlink、太空运输、政府和军方订单，以及未来可能出现的太空算力。火星、外星采矿、人类多星球文明，这些都很伟大，但它们目前很难变成可估值、可预测、可兑现的利润。我们这一代人都很难看得到。

如果遇到股市大幅度回调，这些远大的理想都会和股价一起被地球的引力吸到地表。当然那时候可能会是好的买入机会。

第四，太空算力虽然逻辑成立，但融资压力可能非常大。

太空算力是马斯克新画出来的大饼，但这个大饼并不是完全没逻辑。它在技术方向上确实有可能成立，也符合AI时代对于电力和算力的巨大需求。

问题是，利润空间是否足够支撑SpaceX持续投入？多久能盈利？在盈利之前还要投入多少资金？这次IPO募资是否足够？未来公司是否还要继续增发股票、借债、发债，去支撑太空算力的研发和建设？

看看甲骨文，只是在地球上建设数据中心，就已经需要投入庞大资金、承担巨大债务压力。那么如果是在太空中建设数据中心，成本和不确定性只会更高。

在太空里，散热、维护、辐射、卫星寿命、发射成本、报废周期、网络延迟、任务类型限制，都会带来不确定性。我个人能力有限，无法判断太空算力最终到底能不能形成足够大的商业闭环。因此，在这个问题上，我选择先观察。

四、总结

SpaceX毫无疑问是一家伟大的公司。它承载的是人类对宇宙探索的想象，甚至有可能改变历史。它有点像新时代的荷兰东印度公司，具备在一个未知、广阔、几乎没有被充分开发的领域中形成垄断宇宙的潜力。也正因为如此，它可以在亏损状态下就达到1.7万亿美元左右的市值。

但伟大不等于现在就值得买。伟大的公司，如果在错误的价格买入，也可能带来很长时间的痛苦。

我目前对SpaceX保持谨慎，主要有三个原因。

第一，我对马斯克这次相对仓促的上市决策保持怀疑，不确定这到底是深思熟虑后的战略安排，还是他在AI热潮中的又一次快速转向。

第二，我不看好Grok和xAI被并入SpaceX。我认为大模型并不是SpaceX最有优势的领域，xAI反而可能消耗公司大量资金。

第三，我对太空算力的盈利能力和盈利周期保持谨慎。它在技术逻辑上有吸引力，但距离真正形成可预测利润，可能还有很长距离。

因为公司尚未盈利，又存在以上众多不确定性，所以SpaceX目前注定很难进行传统估值分析，甚至比特斯拉还要更难分析。它上市后是暴涨还是暴跌，没有人能准确预测。

我个人的看法很简单：先等几个月，甚至半年到一年后再看。我会一直保持关注，因为这确实是一家非常值得研究、也很可能值得投资的公司。但它同时也是一家充满风险、充满不确定性、并且极度依赖马斯克个人判断的公司。

对于这种公司，我不想在市场情绪最热、估值最高的时候进去。我更愿意等财务数据和业务进展更清楚一些，再认虑是否买入。

对我来说，稳稳赚取美光未来半年到一年内相对确定的上涨空间，远比现在买入SpaceX，去赌它未来两倍、三倍的涨幅，更加理性，也更加稳妥。

今年 3 月，在股市回调最剧烈的时候，我发现不少芯片股的估值已经非常有吸引力。即便相关公司的业绩表现普遍远超市场预期，股价却因为川普发动的愚蠢且无意义的战争而出现明显回调。正是在这一阶段，我选择了大幅加仓。

回头看，今年最正确的决定，就是在市场恐慌时继续买入美光。之前写过，从去年 9 月到今年 3 月，我已经在每股 160 美元至 350 美元区间持续加仓美光，并将其仓位提高到约 7%左右。但当美光业绩依然强劲、基本面并未恶化，股价却因战争影响从 400 多美元回落至 300 美元附近时，我认为这确实是一次难得的机会，果断继续提高持仓。

当然这也因为自己分析得出其合理价位，年底可达1000左右，因此更有底气。详情可见文章：2026.3.18 美光估值

与此同时，我也做出了明确的资产配置决定：清理掉与 AI 相关性弱的股票，将资金进一步集中到 AI 标的上。

以下为当时实时发在群里的交易记录。这边没有审核的问题，所以以后加仓截图会实时发到substack上。

因为业绩极佳，美光在今年二月份已经涨到了400，当时还小仓位买入了NVO，可以看到当时的英伟达才170，台积电才330，且甲骨文也已经从300降到了156。而我也是从300左右一路不断加大加仓力度，加到了156。

如下图，在二月底的时候，我判断170以下的甲骨文就是可以加大买入的便宜价格。并在200买入了AMD，说实话，这个公司没仔细研究过，但当时就是看价格太便宜了，就开个仓，之后涨到了500多。买了300多的台积电，之后也涨到了400多，以及在150以下大幅加仓ORCL，将亏了半年的甲骨文的成本平摊到了180左右，从而之后涨到上周250，进行部分减仓的时候也赚了不少。

清掉之前买的非AI股票，即便Paypal亏损35%也清掉，往AI股票上集中。

待美光财报出来后，发现即便如此优异的财报，但因为市场对战争的恐慌情绪，继续下跌回300左右，因此继续加仓。

川普还出来讲了废话，导致下跌，又给了加仓的良机。继续买台积电和美光。

继续看好存储，不只是美光，也买入DRAM。

但进入五月后，情况开始发生改变，只要是还没涨的AI芯片公司，都开始天天暴涨。很明显的是，当时小仓位买入了TTMI和高通，但一个月不到都涨了40%-50%。

因此在五月，美股高位以后，我发现已经没有便宜的标的了，我在五月十一号在群中发了下图，停止加仓，开始减仓。

在六月一号和四号的时候，在Discord群中发过当时的仓位比例，并明确指出美股已经高位，需要卖出部分股票，将现金仓位提升到20%，如下图。当时已经明确感觉到美股的热度过高了，无论什么消息都会涨，有很多人因为错失了三四月份的加仓而开始FOMO, 只是美光就盈利超过350%，所以就认为应该提高现金比例，等不知道什么时候必然会到来的回调了。

回顾去年的操作，的确犯了很多错误，比如甲骨文买入的价格过高了；以及因为捡便宜，买入了paypal,和duol, 而两个都亏损了超过30%-50%，不过依旧严格控制了仓位在5-7%，减少了亏损，且在今年在大回调的时候，果断卖出了非AI股票，换到了AI股上，从而获利颇丰。

当然也有很多读者问为什么就留20%现金，上周五跌那么多，为什么不留50%？原因很简单，我进入六月以后就认为股市过热了，回调的概率大大增加，但我也不知道哪天会回调，是否会回调，而且我买的公司最新的业绩依旧保持远超预期的状态，因此我判断20%以上现金就够用了，而且已经这么多了，我并不害怕利润回撤，即便回撤30%，我依旧有让我感到满意的盈利。

二、下半年如有大幅回调，加仓点位。

如果看大盘的话， 个人认为如果能回调到纳指的60日均线，会对我有较大的吸引力，当然这是因为我个人目前仓位依旧有80%的原因，并不代表其他人的情况。而且是否能否回调这么多，没人知道。

其次，如果个别看好的股票能跌破以下价格，也会考虑加仓。

美光：650美元

微软：390美元

英伟达：190美元

台积电：350美元

特斯拉：350美元

甲骨文：180美元

谷歌：330美元 （因为伯克希尔大幅加仓的缘故，最近对谷歌开始关注）

目前虽然有因为川普乱搞导致的加息的威胁，但头部公司的业绩都非常好，短期如有大幅波动，可能会有捡到以上便宜价格的机会。

This article mainly aims to discuss my view on the earnings cyclicality of memory companies, a concern that is widely shared by the market.

Since this is my first article on Substack, I will first briefly review my previous trades and analysis on Micron stock. After all, words alone are not enough.

I began to turn bullish on and buy Micron at the end of September last year, when the stock price was around $160. I then continued to add to my position. When my average cost reached $250, I stopped buying further. Since October 2025, I have been posting articles and transaction records about these trades and judgments in my Discord group and on my WeChat public account, “雪中滚雪球”. The two screenshots below can also serve as reference.

Later, in March this year, when Micron was trading in the $300–400 range, I published a more detailed stock analysis and predicted that its share price could reach $900–1,000 before the end of this year. That prediction has now also been realized. For details, please refer to the article: “2026.3.18 Micron Valuation.”

Main Text:

I believe the cyclicality of the memory industry is changing. It is not that cyclicality has completely disappeared, but rather that the magnitude of cyclical fluctuations is likely to weaken.

First, we need to understand why the memory industry has historically experienced such large cyclical swings. Every three to five years, the industry tends to go through a cycle of demand surging, capacity becoming excessive, prices collapsing, and then gradually returning to normal.

I believe there are several reasons for this.

The technological progress of traditional memory has not been strong enough to stimulate users.

Although traditional memory has always been undergoing technological iteration, this kind of iteration has not been strongly perceived by ordinary consumers or many enterprise customers. Annual upgrades have rarely been able to stimulate consumers or large companies to immediately buy new products to replace the products they already own, in the same way that GPUs or CPUs often can.

There are several reasons behind this.

First, upgrades in traditional memory are more often reflected in larger capacity, faster speed, and lower cost. However, in many cases, these improvements do not directly change the user experience.

Second, for ordinary consumers, as long as the storage in their phones or computers is sufficient, the motivation to upgrade is not particularly strong. Most people will not immediately replace their phone or computer just because storage has become 10% or 20% faster.

Third, for enterprise customers, traditional memory is not an asset that directly creates profit. Therefore, customers care more about price, stability, and supply capability, rather than paying a significant premium for performance improvement. Nor will they buy new products to replace existing ones simply because the speed is 10% faster. Data migration itself also carries costs.

For example, after buying smartphones for so many years, in which year did a storage upgrade or faster read speed make you more eager to replace your phone? Over the past decade or more, I have always paid attention to PC building and played computer games. There was only one time when I felt a real need to upgrade: when solid-state drives, compared with mechanical hard drives, greatly improved game loading times and system performance. Apart from that, the improvements were never that urgent.

Therefore, the progress of traditional memory in the past was not obvious enough to users, and it was not sufficient to continuously create strong new demand.

Traditional memory is prone to price competition.

The technical barrier of traditional memory is actually not low, but the problem is that it has a strong commodity-like nature.

There are also several reasons behind this.

First, as long as specifications, performance, and stability are similar, customers often care more about price than the brand itself.

Second, products such as DRAM and NAND are relatively standardized and do not have obvious differentiation. The difficulty for customers to switch between different suppliers is not as high as it is for CPUs and GPUs.

Third, when industry supply increases, companies often cut prices to clear inventory and win orders, which eventually leads to price wars.

Therefore, the memory industry is an industry with strong commodity characteristics. Once supply becomes excessive, prices can fall very quickly.

Memory capacity is slow to adjust, which easily leads to oversupply.

It takes a long time to increase memory production capacity, and once capacity has been added, it is not easy to adjust. Generally, the construction of a factory and the release of capacity take three to five years, while companies cannot accurately predict market demand several years into the future.

Moreover, once a factory’s production line is running, it is very difficult to stop. Because the memory industry requires massive investment, and equipment depreciation and operating costs are high, companies often continue producing even when prices fall. Otherwise, the losses may be even greater.

Therefore, it is easy to see a situation where the market is tight at the moment, so everyone expands capacity. But several years later, when the factories are completed and capacity is released, demand has already changed, ultimately leading to oversupply. To clear inventory, companies can only sell at lower prices, which causes prices to fall further.

I think a similar example is China’s hog cycle. In China, hog producers include a large number of individual farmers as well as companies, so there are too many sellers, and they cannot make unified adjustments in response to market demand in a timely manner. At the same time, pigs also have a relatively long growth cycle. When the supply of pigs in the market is too low, prices surge, and sellers believe there is profit to be made, so everyone starts raising pigs. But when the pigs are ready for slaughter, this again creates a severe oversupply, eventually causing prices to collapse, and nobody makes money.

Of course, the memory industry is certainly much more complex than hog farming, but the logic is similar: demand changes quickly, supply adjusts slowly, the products themselves do not differ much, and companies cannot collectively control capacity. As a result, cyclical fluctuations are amplified.

After discussing the cyclical problems of memory in the past, why do I think things may improve this time?

First, let us look at a positive example from the same semiconductor industry. Why is it that GPUs and CPUs are also part of the semiconductor industry, yet they have not frequently experienced the kind of extreme price collapses and sharp earnings volatility seen in the traditional memory industry?

I believe there are several reasons.

CPUs and GPUs have higher technical barriers, stronger product differentiation, and fewer competitors.

CPUs and GPUs are both high-barrier industries, so there are only a few truly competitive global players, such as Nvidia, Intel, and AMD.

This brings several consequences.

First, competition is more stable. When there are fewer players, it is less likely that many companies will expand capacity disorderly at the same time and collectively engage in price wars.

Second, product differentiation is stronger. Chips from different companies differ greatly in performance, architecture, software ecosystem, and customer stickiness. They are not as easily compared purely by price as traditional memory. For example, Nvidia’s CUDA ecosystem today is a good example. Once you start using it, even if Nvidia’s GPUs are more expensive, and even if many people shout “AMD yes,” it is still difficult for you to switch to AMD cards.

Third, leading companies are more likely to have pricing power. The more monopolistic an industry is, the higher its technical barriers are, and the stronger its product differentiation is, the easier it is to control prices and avoid falling into pure price competition.

Technological iteration in CPUs and GPUs can more directly create demand.

Since I first came into contact with GPUs, over the past fifteen or sixteen years, performance improvements in Nvidia graphics cards have often been very obvious, and there have frequently been products with major performance leaps.

There are three main reasons behind this.

First, new GPUs can directly improve game graphics, frame rates, and user experience, so consumers can clearly feel the value of upgrading.

Second, GPUs are not only consumer products, but also productivity tools. For design, rendering, scientific computing, AI training, and inference, performance improvement can directly bring efficiency improvement.

Third, in the AI era, GPU performance improvements directly affect training and inference costs, and also determine whether customers can obtain stronger computing power at a lower cost. Therefore, customers are willing to continuously buy better GPUs, because they are not merely “a little faster,” but can potentially bring stronger competitiveness and higher profits.

Traditional memory in the past had difficulty achieving this. It was also improving, but this improvement was not direct enough, not strong enough to stimulate demand, and not enough to make customers willing to pay a significant premium for better performance.

So what has changed in memory today?

I believe that HBM, which is specifically used in AI data centers, is a product that changes the cyclicality of memory. It does not mean that the memory industry will have no cycles at all, but it may change several of the core problems of the traditional memory industry that I mentioned above, thereby making cyclical fluctuations much smaller.

HBM has high technical barriers and a better competitive structure.

HBM is not ordinary DRAM. It is not something that can be produced simply by expanding capacity. It involves many steps, including packaging, as well as close collaboration with GPU manufacturers.

This brings several changes.

First, very few companies can produce high-end HBM. At present, the companies truly capable of mass-producing high-end HBM are mainly the three giants: Samsung, SK Hynix, and Micron.

Second, HBM is much more difficult to produce. It is not something new competitors can enter simply because they want to. It requires not only technology, but also yield, capacity, customer certification, and long-term cooperation capabilities.

Third, HBM is more deeply tied to GPU manufacturers. For example, HBM used for Nvidia must go through strict and long-term validation before it can pass qualification. Therefore, whether a company can enter the supply chain of customers such as Nvidia and AMD is itself a very high barrier.

Therefore, the competitive structure of HBM is better than that of traditional memory. There are fewer competitors, and the pressure from price wars should also be relatively smaller.

HBM’s technological iteration can directly stimulate demand.

Because HBM is used in AI data centers, its bandwidth, power consumption, and performance directly affect the overall performance of AI GPUs. The faster it iterates, the more obvious the improvement to the computing system becomes, and the more obvious the improvement to AI model training and inference efficiency becomes.

There are several main reasons.

First, AI GPUs have extremely high requirements for memory bandwidth. The stronger the performance of HBM, the less likely GPU computing power is to be limited by memory bottlenecks.

Second, AI data centers care deeply about energy efficiency. If HBM can control power consumption while delivering higher performance, it can help customers reduce long-term operating costs.

Third, for AI companies, stronger HBM is not simply a hardware upgrade. It can directly affect model training efficiency, inference efficiency, and overall computing cost.

This will strongly incentivize manufacturers to produce higher-performance HBM, and it will also make buyers willing to purchase the best HBM.

If a new generation of HBM can deliver higher performance, lower power consumption, and greater efficiency, then customers will be willing to pay a higher price.

This is somewhat similar to what Jensen Huang said: “The more you buy, the more you save.” For ordinary consumers, spending more money on faster memory may not feel that obvious. But for AI data centers, improvements in performance and efficiency can be directly converted into cost advantages and competitive advantages.

HBM demand comes from AI data centers, not ordinary consumer electronics.

In the past, much of the demand for traditional memory came from smartphones, PCs, and ordinary servers. Once consumption in these industries weakened, demand would decline quickly, customers would begin destocking, and memory prices would eventually collapse.

Therefore, the demand foundation for HBM is different from traditional memory’s past dependence on smartphone and PC cycles. The cyclicality of a B2B business is smaller than that of a B2C business.

The three giants now also place more emphasis on capacity control than before.

After experiencing multiple memory cycles, Samsung, SK Hynix, and Micron should all understand very clearly that if everyone blindly expands capacity, the industry will only repeat the same outcome of oversupply and price collapse.

There are also several changes here.

First, companies now place greater emphasis on high-end capacity, rather than simply expanding ordinary DRAM and NAND capacity.

Second, HBM capacity expansion itself is more complex. It is not simply a matter of buying equipment and building factories, after which effective supply can immediately be released. Yield and customer certification both take time.

Third, the industry has now added tools such as three- to five-year long-term agreements, making capacity and orders more controllable. In the past, the memory industry often expanded capacity based on short-term prices and demand, which easily led to large boom-and-bust swings. But if more future capacity is pre-bound to major customers, with long-term orders, customer certification, and clear demand, then supply and demand should become more stable than before.

Of course, I am not saying they will not expand capacity at all. As long as HBM profits are high enough, companies will certainly still invest in capacity expansion. But the difference this time is that expansion is more focused on high-end HBM and advanced processes, rather than simply increasing ordinary DRAM and NAND capacity.

After the emergence of HBM for AI, I believe it has indeed changed the three core problems that caused the cyclicality of traditional memory.

Therefore, I believe the memory industry will not be free of cycles from now on, but the magnitude of the cycle will be greatly reduced. Especially for companies such as Micron, SK Hynix, and Samsung, which can truly produce high-end HBM, if the share of HBM continues to rise in the future, their earnings volatility may no longer be as severe as it was in past traditional memory cycles.

That is why I started buying last year, and why I focused on Micron, which can manufacture HBM, as well as a DRAM ETF that includes SK Hynix and Samsung. One point I should add is that I believe Micron is the better choice because it is an American company.

In my analysis in March this year, I already used a PE valuation. At that time, many institutions were still criticizing the use of PE valuation, only to switch to using PE valuation themselves one or two months later, which I find rather ridiculous.

As for Micron, it currently accounts for more than 35% of my portfolio. I believe that if this round of decline takes the stock below $600, I will add to my position again.

I am too lazy to write any disclaimer here. Other people’s views are always only references. Everyone should be responsible for their own decisions.

This is my first article here. I hope everyone will share it widely.

本文主要想谈一下我对于市场普遍担忧的存储公司业绩周期性的看法。

因为这是我在 Substack 上的第一篇文章，所以先简单回顾一下此前我对美光股票的交易和分析，毕竟口说无凭。

我是从去年 9 月底开始看好并买入美光的，当时股价约为 160 美元，之后一路加仓。当持仓成本达到 250 美元后，我便停止了继续买入。这些交易和判断，从 2025 年 10 月开始，我一直在 Discord 群以及微信公众号“雪中滚雪球”上发布过文章和交易记录，下面两张截图也可以作为参考。

之后在今年 3 月份，当美光股价处于 300—400 美元区间时，我发布了更具体的股票分析，并预测其股价在今年年底前有望达到 900—1000 美元。如今这一预测也已经实现。详情可见文章：2026.3.18 美光估值。

正文：

我认为存储行业的周期性正在改变，并不是完全没有周期性了，而是周期波动的幅度将会减弱。

首先我们要了解一下存储行业为什么会周期波动这么大，每过三五年就会经历一次需求暴增、产能过剩、价格暴跌，然后又恢复正常。

我认为有以下几点原因：

1. 传统存储的技术进步，对用户的刺激不够强。

以前传统存储虽然也一直在技术迭代，但这种技术迭代对普通消费者和很多企业客户来说，感知并没有那么强。每年的升级，很难像 GPU、CPU 那样刺激消费者或者大公司立刻去购买新款产品，去替代已经买到手的旧款。

这背后主要有几个原因。

第一，传统存储的升级更多是容量变大、速度变快、成本下降，但这些提升很多时候并不会直接改变用户体验。

第二，对于普通消费者来说，手机、电脑里的存储只要够用，升级动力就不会特别强。大多数人不会因为存储快了 10%、20%，就立刻换手机或者换电脑。

第三，对于企业客户来说，传统存储不是直接创造利润的资产。所以客户更在意价格、稳定性和供货能力，而不是为了性能提升支付明显溢价，也不会速度快了10%就去再买新的来替换，而且数据的迁移一样花费成本。

就比如你买了这么多年手机，哪年因为存储升级、读取速度更快了，所以让你增加了换手机的欲望？这十几年来，我一直关注装机，玩电脑游戏，也只有一次因为固态硬盘相较于机械硬盘能让游戏读取、系统运行速度大大提升，而激发了我更新的需求。其它时候的提升都没有那么迫切。

所以传统存储以前的进步对用户来说不够明显，不足以持续创造强烈的新需求。

2. 传统存储容易陷入价格竞争。

传统存储的技术门槛其实并不低，但问题在于它的商品属性比较强。

这背后也有几个原因。

第一，只要规格、性能、稳定性差不多，客户很多时候更看重价格，而不是品牌本身。

第二，DRAM、NAND 这类产品相对标准化，没什么明显差异，客户在不同供应商之间切换的难度没有 CPU、GPU 那么高。

第三，当行业供给变多的时候，企业为了消化库存，就很容易通过降价来抢订单，最后形成价格战。

所以存储行业是一个商品属性很强的行业。一旦供给过剩，价格就会跌得非常快。

3. 存储产能调整慢，容易造成供过于求。

存储的产能提升需要较长的时间，且一旦提升就不易再调整。一般一个工厂的建造和产能释放需要3-5年时间，而公司是无法准确预测几年之后市场需求的。

而且工厂的产线开起来之后就很难停。因为存储行业投资巨大，设备折旧和运营成本很高，即使价格下跌，企业很多时候也会继续生产，否则损失更大。

所以很容易出现当下市场紧缺，大家都扩产，但几年后当工厂建好、产能释放出来时，需求却已经变了，最后导致供大于求。企业为了消化库存，只能降价出售，最终导致价格进一步下跌。

我认为有一个类似的例子就是国内的猪周期。因为我国养猪主体包含了众多个体户以及公司，所以卖家过多，无法针对市场需求及时做出统一调整。而且猪生长的周期也较长，当市场上猪的供给过少时，价格暴涨，卖家就觉得有利可图，全都去养猪；当猪出栏的时候，却又造成了供给严重过剩，最终价格暴跌，大家都赚不到钱。

当然，存储行业肯定比养猪复杂得多，但逻辑相似：需求变化快，供给调整慢，而且每家的猪肉没什么大的差异，大家又无法统一控制产能，所以周期波动就会被放大。

说完以前存储的周期问题，那为什么我认为这次会有所改观？

首先看一下同样是芯片行业的正面案例，为什么这么多年来 GPU、CPU 同样是芯片行业，但它们没有像传统存储行业那样，经常出现那么极端的价格暴跌和盈利大幅波动？

我认为有以下几点原因：

1. CPU 和 GPU 的技术门槛更高，产品差异化更强，竞争对手更少。

CPU 和 GPU 都是高技术门槛行业，所以全球真正有竞争力的玩家就那么几家，比如英伟达、英特尔和 AMD。

这会带来几个结果。

第一，竞争更稳定。玩家少，就不容易出现很多企业同时无序扩产、一起打价格战的情况。

第二，产品差异化更强。不同公司的芯片性能、架构、软件生态、客户粘性差异很大，不像传统存储那样容易被简单比价。就比如现在英伟达的CUDA生态，你一旦用了，即便英伟达的显卡更贵，再多人喊AMD yes, 你也难换到A卡上。

第三，龙头公司更容易掌握定价权。越垄断、越有技术壁垒、产品差异化越强的行业，越容易控制价格，也越不容易陷入纯粹的价格战。

2. CPU 和 GPU 的技术迭代，更容易直接创造需求。

从我接触 GPU 开始，这十五六年来，英伟达显卡的性能提升很多时候都非常明显，而且经常会有性能跨越比较大的产品。

这背后主要有三个原因。

第一，新 GPU 能直接提升游戏画面、帧率和用户体验，所以消费者能明显感受到升级价值。

第二，GPU 不只是消费品，也是生产力工具。对于设计、渲染、科学计算、AI 训练和推理来说，性能提升可以直接带来效率提升。

第三，到了 AI 时代，GPU 的性能提升会直接影响训练和推理成本，也会影响客户能不能用更低成本获得更强算力。所以客户愿意持续购买更好的 GPU，因为它不只是“更快一点”，而是可能直接带来更强的竞争力和更高的利润。

而以前的传统存储很难做到这一点。它也在进步，但这种进步不够直接、不够刺激需求，也不够让客户愿意为了性能提升支付明显更高的价格。

那么现在的存储有什么变化？

我认为专门用于 AI 数据中心的 HBM，是改变存储周期性的产品。它不是让存储行业完全没有周期，而是有可能改变我上面所说的传统存储行业几个核心问题，从而让周期波动变的小的多。

1. HBM 技术门槛高，竞争格局更好。

HBM 不是普通 DRAM，不是简单扩产就能做出来。它涉封装等及众多步骤，以及和 GPU 厂商的配合。

这带来几个变化。

第一，能做高端 HBM 的企业很少。目前真正有能力大规模生产高端 HBM 的，主要还是三星、海力士和美光这三家巨头。

第二，HBM 的生产难度更高，不是新竞争对手想进来就能进来。它不仅要有技术，还要有良率、产能、客户认证和长期配合能力。

第三，HBM 和 GPU 厂商绑定更深。就比如目前用于英伟达的HBM必须要经过严格的长期验证才能经过检验。所以能不能进入英伟达、AMD 等客户的供应链，本身就是一道很高的门槛。

所以 HBM 的竞争格局比传统存储更好，竞争对手数量更少，价格战的压力也会相对小一些。

2. HBM 的技术迭代，能直接刺激需求。

因为 HBM 是用于 AI 数据中心的，它的带宽、功耗和性能会直接影响 AI GPU 的整体表现。你迭代越快，对于算力系统的提升就越明显，对于 AI 模型训练和推理效率的提升也越明显。

主要有几个原因。

第一，AI GPU 对内存带宽要求极高。HBM 性能越强，GPU 的算力越不容易被内存瓶颈限制。

第二，AI 数据中心非常看重能效。HBM 如果能在更高性能下控制功耗，就能帮助客户降低长期运营成本。

第三，对于 AI 公司来说，更强的 HBM 不是简单的硬件升级，而是能直接影响模型训练效率、推理效率和整体算力成本。

这会大大激励厂家去生产更高性能的 HBM，也会让买家愿意购买最好的 HBM。

新款 HBM 如果能带来更高性能、更低功耗、更高效率，那客户就愿意付更高的价格。

这就和老黄说的“你买得越多，省得越多”有点类似。对于普通消费者来说，多花钱买更快的存储可能没那么明显；但对于 AI 数据中心来说，性能和效率的提升是可以直接转化成成本优势和竞争优势的。

3. HBM 的需求来自 AI 数据中心，而不是普通消费电子。

以前传统存储的需求，很多来自手机、PC、普通服务器。这些行业一旦消费变弱，需求就会快速下滑，客户也会开始去库存，最后导致存储价格暴跌。

所以 HBM 的需求基础，和过去传统存储依赖手机、PC 周期是不一样的。做To B生意的周期性波动是小于To C的。

4. 三家巨头现在也比以前更重视产能控制。

经历过多轮存储周期之后，三星、海力士和美光应该都很清楚，如果所有人都无脑扩产，最后只会重蹈产能过剩、价格暴跌的覆辙。

这里面也有几个变化。

第一，企业现在更重视高端产能，而不是简单扩大普通 DRAM 和 NAND 产能。

第二，HBM 的扩产本身更复杂，不是简单买设备、建厂就能立刻释放有效供给，良率和客户认证都需要时间。

第三，现在又增加了三年到五年长约这种手段，使得产能和订单都更可控。以前存储行业很多时候是根据短期价格和需求来扩产，容易造成大起大落；但如果未来更多产能提前和大客户绑定，有长期订单、有客户认证、有明确需求，那么供需关系就会比以前更稳定。

当然，我不是说它们完全不会扩产。只要 HBM 利润足够高，企业肯定还是会投资扩产。但这次和以前不一样的是，扩产更多集中在高端 HBM 和先进制程上，而不是简单扩大普通 DRAM 和 NAND 产能。

用于 AI 的 HBM 出现后，我认为确实改变了导致传统存储周期性的那三个核心问题。

所以我认为，存储行业并不是从此没有周期了，而是周期的幅度会大大减弱。尤其是对于美光、海力士、三星这些真正能做高端 HBM 的公司来说，如果未来 HBM 占比越来越高，它们的盈利波动可能就不会像过去传统存储周期那样剧烈。

因此我在去年就开始买入，且重点买了能制造HMB的美光和包含海力士、三星的 DRAM ETF。需要说明的一点是，我认为美光更优，因为它是美国公司。

而在今年三月分析里我就用了 PE 估值，众多机构在当时还在批判使用PE估值，但又在一两个月后也改为使用PE估值，也是挺可笑的。

对于美光，目前已经占我35%以上仓位，我认为如果这轮下跌能低于600美元美股，我会再次加仓。

在这里我也懒得写什么免责声明了，别人的东西永远只是参考，为自己的决策负责。

这是这里的第一篇，希望大家多多分享。

Share 雪中滚雪球

之后将会逐渐停止在微信公众号上发布有关具体美股股票和公司的分析文章，买入点位，股票交易截图，看法等等；看清趋势，去顺应而不是逆行，没有意义。

也考虑过流量更大的Youtube, 但剪辑视频耗费时间，况且盈利并不是主要目的，舍弃自己更重要的休息娱乐时间是得不偿失的，所以目前依旧更倾向以文字为主。之后会主要在Substack发布分析文章和交易记录等，喜欢的朋友可以输入自己的邮箱subscribe一下。

如果想查看我之前的文章和观点（特斯拉，美光，英伟达，台积电等等），也可以在微信公众号搜索“雪中滚雪球”阅读。部分仍有参考价值的分析文章，我会重新整理、编辑后发布到 Substack。

这个平台的设计理念挺有意思，通过更原始的邮件订阅，而不是用社交媒体的算法和推荐，让作者和读者更直接的沟通。

想加入群的朋友可以通过这个链接加入Discord的美股交流群，discord.gg/Jpu9CmwrYy。但需要注意的是，群众每人的想法都是与我不同的，如果自己没有主见，容易受到他人意见的干扰，不建议加入，我的分析都会发在这里，看文章即可。