各位網友，昨天已經傳出美國和伊朗的和談談好了，其實Donald Trump回去白宮是在等和約的消息，不是預備去打伊朗，我看見那些支持Donald Trump的人，說他準備大開殺戒，我覺得非常好笑。昨天放出來的消息，美國和伊朗都說那些原則已經談好了，是可能有些wording有些問題，本來說今天會簽的，現在可能說遲一兩天，那會不會再有變故呢？我覺得有變故的機會都是很小的，我expect幾個小時後開始股市會反彈。
我首先看看其他的factor，布蘭特原油104元，我告訴大家，如果美伊真的達成和約，我最新的看法是油價會大幅度下跌。為什麼呢？其實很簡單，因為在過去那三個月，一直在drawdown庫存的石油，那些庫存的石油drawdown了出來不會立即補回來的，避免影響股市，會慢慢等它跌下來才補回來。也就是說, 一直過去那80至90天沒有賣的石油，全部可以推出來，只是在波斯灣裏面的船，已經有幾億桶的石油，是要很快推出來賣，鬥快賣。而美國北美正在生產的石油繼續推出，短期內石油的價錢會跌，石油的價錢一回跌的時候，通脹的factor就會跌。而新的聯儲局主席一定是推動減息，他做聯儲局主席是因為他答應了Donald Trump這件事，這樣對股市是好消息。
黃金一直站在similar的價錢，4500多元，如果減息的時候是對黃金有利的。白銀站在75元，最近白銀最有趣的就是，現在光伏板不用白銀，我真的被confuse。
比特幣那一下跌下來，又跌回7500多元，比特幣現在是一種高危的產品，變成了危險率很高的投資，等於AI股一樣，所以如果減息也是對比特幣有利。比特幣最不利的東西是已經沒有故事好說，比特幣已經沒有一個故事，那時候的故事就是說，如果ETF那些傳統基金開始進入比特幣，比特幣就發大財了，這件事已經發生了，結果沒有發大財。
好，Tesla一連幾天反彈，由400元一直上升到440元，然後下降到426元，因爲那次上升到440元，是expect中國放full self driving。現在就是說，早兩天有人登了一段文章出來，說在Tesla的網頁，已經把中國放進去，有level 2的full self driving的服務，那個是一個錯誤。一直以來中國都有被放進去，因爲有一套不完整的full self driving在大陸可以使用的，但是這套東西的用途是有限的。Tesla訓練了一套新的出來，那套新的一直都還沒approved，一直在說的是這個問題。
我都說Tesla現在完全看一件事，就是看它的Robotaxi。14.3.3已經出了，那個smart summon好了很多，現在差不多開Tesla的人，全部覺得完全不用開車的，是非常舒服的，你永遠不會回到過去，我覺得那個subscription for full self driving，是會增加得很快的。但是美國的車有限度，在北美128萬輛已經有很高的滲透率，去到三四成的滲透率，其他的就要等hardware 3到月底會出一個full self driving，推出14.3.3的light，那套如果都可以的話就會增加subscription。
現在最大的希望當然就是兩件事，一樣就是中國批准了，中國批准了那個subscription最大，因為中國Tesla的車最多；歐洲西班牙很快批准，立陶宛批准，比利時將會很快批准，其他到了6月底的時候，是不是可以歐洲很多國家批准呢？那些人在歐洲開過Tesla的full self driving，是覺得很震撼的，在歐洲這樣的situation都很成熟。meanwhile我跟大家説説，Waymo很大問題，原來一上去highway是非常危險，如果高速運行，Waymo是全部被withdraw了，現在Waymo是不可以上highway，Waymo是露出了很多缺點。為什麼Waymo露出很多缺點呢？因為我可以告訴你，這個full self driving的長尾，那條尾是長到你猜不到的長，這個是用programming無法解決的問題，長尾多到。
Tesla是怎樣的呢？Elon Musk這個傻子，只是用了11個星期，弄了一條模擬的馬路，在奧斯丁，是所有各種各樣的彎，上斜路、下斜路，他那條路都有的。然後一有長尾的事出現，就是有一樣東西他們未曾experience過的，他們就立即把Tesla車放在他那條路，similar的路段，然後加上simulation，把他變成一萬種variation出來，不停嘗試，又用一個physical的world加simulation。舉個例子，原來轉左然後在樹上面有一隻松鼠掉下來，他一發現這樣東西，以前未曾見過，於是他就立即similar不同高低的樹、不同的松鼠、不同的動物掉下來，model Y、model 3，cybertruck的反應是怎樣，他立即simulate很多的situation出來，用來解決長尾的問題，這個是要這樣的，沒有real life data，做simulation是不行的，其實是不行的，simulation只是幫到的。
另外就是Semi是很受歡迎，現在開始出單，就是Semi接很多單。但是現在有一個比較令我失望的消息就是4680，終於出來的電池密度是不夠好的，比起他當初答應的差很遠，電的密度和持久力和charger沒有顯著的改善，而且也便宜不過，因為你弄了幾年4680，人家在弄磷酸鐵鋰電池，人家也跌了很多價，是便宜不過磷酸鐵鋰電池的，這是為什麼4680電池不能夠令成本得到很大的優勢。但anyway，我每天都在監視Tesla，在5月內，究竟Robotaxi無監督版還可不可以double，他上個月是double了，現在去到38，39架，如果他一過過百架，我覺得商業途徑已經成熟了。
另夕外就是輝達，NVIDIA的問題就是，過去那幾個月我都是不停問這個問題，那些人是相信，對NVIDIA的批評是完全荒謬的。NVIDIA已經price in了它的高速增長，所以它這個是完美的故事，如果它有一點點瑕疵，那些人就覺得不看好。你現在是怎樣price in了它的高速增長，你現在只有20幾倍PE，你price in了它是一個很low growth的東西，它的增長不是80%至90%，是跌了一半，是40%至50%，好像都應該40幾50倍PE，好像是，它keep不到80%至90%，怎樣都keep到40%。這些理由完全荒謬。又説可能危機供應不到中國，現在他的數目沒有供應中國的數目，供應到中國只會加分，不會因為供應不到中國就減分。最大的問題就是，還有Cerebras那件事是荒謬的，Cerebras怎麽會威脅到NVIDIA，當你做一個這麽大的晶片出來，那個難度是貴到不得了，我告訴大家，而且你根本拿不到產能做。這個叫做什麽呢？你做一張這麽大、12吋的晶片，是有80萬個核心，80萬個parallel running，可以80萬一起parallel run，而是用SRAM放在記憶體，這樣的時候它的latency就很低。但是最大的問題就是，你這麽多的算力，SRAM放進去的記憶體是不夠多的，於是這張碟就很多時間在等，因為不夠data去運算，於是他現在想切開，有些常用的data，有些不常用的data，但那些太technical，我不説了。但anyway你根本和台積電做一樣這樣的事情，世界上只有台積電會做，因為光罩是有一定的size，印很多個光罩，然後用stitches將它聯合在一起，變成一個聯合的光罩，明白我說什麽嗎？光罩的size是有限制的，那怎麽弄到一隻這麽大的碟呢？這個叫做光罩stitches，就是將它聯合在一起，只有台積電會做，而你根本和台積電是拿不到大的產能的。
真正要問的輝達的問題就是，在inference和agentic的世界裏面，究竟ASIC晶片，尤其是TPU，是否挑戰到它呢？這個問題我答了很多次，我不想再說。這個是我用一個最簡單的答案來告訴大家，NVIDIA是利用它的economy of scale，和它對產能、全世界產能的控制，它是可以給很多錢下去，預早訂了所有的產能，令到你沒有產能，全世界只有它可以afford這樣做，訂了台積電的CoWoS、HBM，這些全部被它訂了，接下來的全部給它訂了，你根本就拿不到產能，因此它的迭代是快過你很多。因為譬如你流片，我解釋給大家聽，這些比較technical，為什麼Elon Musk這麼生氣呢？他那個AI5，你現在流片了，你猜幾時做出來呢？2027年尾，流完片之後是12至18個月。
NVIDIA不是，它流完片之後幾個月就做出來了，因為它一早就給了錢下訂了，他那隻今年年頭才搞的Rubin Ultra，他才買了Colossus回來，第四季已經量產了，你那隻TPU還要排到明年，你現在發明是十幾個月後，台積電已經變成第二代了，它的迭代之快是令到你們沒辦法追，就是因為它是一間很恐怖的公司。它今年賺大約2100億美金，世界上從來沒有一間公司曾經賺過2100億美金，這個是驚人的數字，它的現金流，它用這個壓死你們。
台積電也是，我應該出一條片解釋一下CoWoS，它只是這個CoWoS封裝已經是壟斷了，因為你現在一定要把很多晶片疊在一起，這個就要靠先進封裝，變成一隻很大的晶片，才減少Latency。現在Cerebras走另一條路，它弄一個很大的晶片出來，然後還有一條路就是EMIB，Intel搭橋，這個我應該做一條專業的分析給大家聽，兩種東西的分別。但是台積電趁你還在搞EMIB，它已經變成CoWoS，CoWoS就變成放在玻璃基板上，是四方形的玻璃基板上，節省了很多錢，和尺寸沒有限制，所以台積電除非大陸打台灣，不然5年內是絕對無敵的。
VRT是我正在買的股票，大家看到我只買了20多股，因為很高才開始追，很高開始追我不敢買這麼多做底貨，底貨只有10-20股，每逢跌就買，這個是很穩陣的事情。這件事我可以告訴大家，液冷是一個不可逆的因素，而所有新的東西都是要用液冷，而VRT就是液冷的領導者。
Palantir不用說，Palantir今天也是只有70至80億一年的turnover，這種公司是怎樣呢？是大公司和政府的captive software，就是它一幫你做了之後，你以後就不可以沒有它，它是底層架構，幫你將所有數據收集，然後作出戰略決策，這樣的公司你應該將它和IBM比，IBM也有10幾20倍今時今日，而IBM很明顯份額在輸，Oracle也是，這些公司的定價權是很大的。而Palantir就是現在雖然有人意圖和它競爭，但是還沒有build up到，它是有10幾20年的經驗，就是將所有的數據整理，而做戰略決策，而且加上執行，這個暫時沒有其他公司的。本來我一向認為它的價錢貴到離譜，現在跌下跌下就不是很貴，現在只是100倍PE，這樣就不是很貴，接下來一年它的業務一定double，這樣就變成50倍PE。
Tempus AI一直很多消息，很多人和它合作，你看到它過去5天升了4.81%，Roblox升了11%多，我的印象就是這些股份似乎在醞釀中，原底收集完，就是在等硬件告一個階段的時候，它們就會爆升。海力士不用說了，我不懂得說，海力士是世界上第五間賺1000幾億美金的公司，現在它的市值只有7000億，是幾倍PE，而接下來這兩年它還是double再double的。當然我們很幸運就是在300元的時候買下去，現在它連跌幾天，還溝不溝呢？它連跌兩天我就溝一點點。
AVAV這個最近似乎站穩了，開始我在160那裏一直慢慢賣，這個其實我們最初賺了一次很厲害的，220元賣，410元賣了，然後跌下來買。它主要是失了一張大單，它收購的那間BlueHalo，有一張大單它是算進去收購那條數，那張單失敗了，所以這麽大的回扣，現在似乎已經找到底了，好不好，這段講到這裏。
The news leaked yesterday that the peace talks between the United States and Iran have been finalized. In fact, Donald Trump's return to the White House is to await news of the peace treaty, not to prepare for a war against Iran. I find it hilarious when I see his supporters saying he is ready to go on a killing spree. According to the information released yesterday, both the US and Iran stated that the principles have been agreed upon, though there might be some issues with the wording. It was originally supposed to be signed today, but now it might be delayed by a day or two. Will there be any unexpected changes? I think the chance of a reversal is minimal, and I expect the stock market to start rebounding in a few hours.
Looking at other factors, Brent crude is at 104 dollars. I am telling everyone that if the US-Iran peace treaty is truly reached, my latest view is that oil prices will drop significantly. Why? It is simple. For the past three months, oil has been continually drawn down from inventories. Those drawn-down inventories will not be replenished immediately in order to avoid affecting the stock market; they will wait for the price to drop gradually before buying back. This means that all the oil that was not sold over the past 80 to 90 days can now be pushed into the market. Just the ships sitting in the Persian Gulf already hold hundreds of millions of barrels of oil, which will need to be pushed out and sold very quickly in a race to sell first. Meanwhile, the oil being produced in North America will continue to hit the market. In the short term, oil prices will fall, and as soon as oil prices retreat, the inflation factor will drop. Furthermore, the new Federal Reserve Chairman will definitely push for interest rate cuts, as he took the position precisely because he promised Donald Trump he would do this, which is good news for the stock market.
Gold has been holding at a similar price, over 4500 dollars, and cutting interest rates is beneficial for gold. Silver is standing at 75 dollars. The most interesting thing about silver lately is that photovoltaic panels no longer use silver, which really confuses me.
Bitcoin dropped all of a sudden back to over 7500 dollars. Bitcoin has now become a high-risk product, turning into a very high-danger investment, just like AI stocks, so cutting interest rates would also benefit Bitcoin. The most disadvantageous thing for Bitcoin is that it no longer has a story to tell. The story back then was that if traditional funds like ETFs started entering Bitcoin, it would make a fortune. This has already happened, and as it turns out, it did not make a fortune.
Tesla rebounded for several consecutive days, rising all the way from 400 dollars to 440 dollars before dropping back to 426 dollars. That rise to 440 dollars was due to expectations that China would release Full Self-Driving. Two days ago, someone published an article claiming that China had been added to Tesla's website as having Level 2 Full Self-Driving service available, but that was an error. China has always been included because an incomplete version of Full Self-Driving can be used in the mainland, but the utility of that system is limited. Tesla trained a new system, and that new one has not been approved yet, which is what everyone has been talking about.
As I have said, Tesla now depends entirely on one thing, which is its Robotaxi. Version 14.3.3 is out, and that smart summon is much better. Now, almost everyone driving a Tesla feels like they do not need to drive at all, and it is extremely comfortable. You can never go back to the past, and I think the subscriptions for Full Self-Driving will increase very quickly. However, the number of cars in the US is limited. In North America, 1.28 million cars already represent a very high penetration rate, reaching 30% to 40%. The rest will have to wait for Hardware 3 to release a Full Self-Driving 14.3.3 Light version by the end of the month, and if that works well, it will increase subscriptions.
Right now, the greatest hope rests on two things. One is approval from China, which represents the biggest subscription potential because China has the largest number of Tesla vehicles. Spain in Europe will approve it very soon, Lithuania has approved it, and Belgium will approve it shortly. By the end of June, will many other European countries approve it? People who have driven Tesla's Full Self-Driving in Europe find it shocking, as the situation in Europe is very mature for it. Meanwhile, let me tell you that Waymo has big problems. It turns out that going onto the highway is very dangerous for them, so whenever it comes to high-speed operation, Waymo vehicles have all been withdrawn. Currently, Waymo is not allowed on highways, and many of its flaws have been exposed. Why has Waymo shown so many weaknesses? Because I can tell you that the long tail of Full Self-Driving is so long that it is beyond your imagination, and this is a problem that cannot be solved by programming alone due to the sheer volume of long-tail scenarios.
How does Tesla handle it? Elon Musk, that crazy guy, spent just 11 weeks building a simulated road in Austin that features all kinds of curves, uphill slopes, and downhill slopes. Whenever a long-tail event occurs—meaning something they have never experienced before—they immediately place a Tesla vehicle on that road or a similar stretch, and then add simulation to generate ten thousand variations of it for continuous testing, combining the physical world with simulation. For example, if a car turns left and a squirrel suddenly falls from a tree, which is something they have never seen before, they will immediately simulate trees of different heights and various squirrels or other animals falling down to see how the Model Y, Model 3, and Cybertruck react. He immediately simulates numerous situations like this to solve the long-tail problem. This is how it must be done; without real-life data, relying solely on simulation does not work, it really does not work, though simulation does help.
In addition, the Semi is very popular and orders are starting to come in, meaning the Semi is receiving many orders. However, there is a piece of news that disappoints me quite a bit, which is that the energy density of the 4680 battery that finally came out is not good enough. It falls far short of what he originally promised, showing no significant improvements in energy density, longevity, or charging speed. Moreover, it is not cheaper either. Because while you spent several years developing the 4680, others were working on lithium iron phosphate batteries, which have dropped heavily in price, and the 4680 cannot beat lithium iron phosphate on price. This is why the 4680 battery cannot bring a massive cost advantage. But anyway, I am monitoring Tesla every day to see whether unsupervised Robotaxis can double within May. It doubled last month, reaching about 38 or 39 units now, and if it surpasses a hundred units, I think the commercial path will have matured.
On another note, regarding NVIDIA, the problem is that for the past few months I have been constantly asking this question, and those people believe that the criticisms against NVIDIA are completely absurd. NVIDIA has already priced in its high-speed growth, so its story is perfect, and if there is even a tiny flaw, people start turning bearish. How exactly have you priced in its high-speed growth? Right now you only have a PE ratio of twenty-something, pricing it in as if it were a low-growth entity. Even if its growth drops by half from 80% to 90% down to 40% to 50%, it seems like it should still command a PE ratio of 40 or 50. It seems that even if it cannot maintain 80% to 90%, it will maintain 40% no matter what. These reasons are completely absurd. They also say that there is a risk it cannot supply China, but its current figures do not even include supply to China. Supplying China would only add points, and it will not lose points just because it cannot supply China. The biggest issue is that the Cerebras situation is also absurd. How could Cerebras possibly threaten NVIDIA? When you manufacture a chip that large, the difficulty makes it incredibly expensive, and I tell everyone that you simply cannot secure the production capacity for it. What is this called? You make a massive 12-inch wafer that contains 800,000 cores running in parallel. It can run 800,000 cores simultaneously and uses SRAM built into the memory, which gives it very low latency. But the biggest problem is that with so much computing power, the SRAM memory embedded within it is not enough, so this disc spends a lot of time waiting because there is not enough data for computation. Therefore, they now want to split it into frequently used data and infrequently used data, but that is too technical so I will not elaborate. Anyway, you are essentially doing the same thing that TSMC does, and TSMC is the only one in the world that can do it because reticle sizes are limited. They print many reticles and then use stitches to link them together into a unified reticle. Do you understand what I am saying? The reticle size has limits, so how do you make a disc that large? This is called reticle stitching, which links them together. Only TSMC can do it, and you simply cannot secure large production capacity from TSMC.
The real question to ask about NVIDIA is whether ASIC chips, especially TPUs, will challenge it in the world of inference and agentic AI. I have answered this question many times and do not want to repeat myself. Let me give you the simplest answer: NVIDIA utilizes its economies of scale and its control over global production capacity. It can put down a massive amount of money to pre-order all the capacity, leaving you with none. It is the only company in the world that can afford to do this. It has booked TSMC's CoWoS and HBM; all of these have been locked up by NVIDIA, and the upcoming capacity is entirely reserved for it. You simply cannot get any capacity, and therefore its iteration speed is much faster than yours. For instance, when it comes to taping out—let me explain this to you, these are more technical matters, which is why Elon Musk gets so angry—take his AI5, now that it has been taped out, when do you think it will be manufactured? By the end of 2027, because it takes 12 to 18 months after taping out.
NVIDIA is different. It manufactures its chips just a few months after taping out because it paid the deposits and placed orders way in advance. Take its Rubin Ultra, which they only started working on early this year; they just bought Colossus, and it is already in mass production in the fourth quarter. Meanwhile, your TPU is still queued up until next year. By the time what you invented now actually comes out a dozen months later, TSMC will have already moved on to the second generation. Its pace of iteration is so fast that it makes it impossible for you to catch up, simply because it is a terrifying company. It earns about 210 billion US dollars this year. No company in the world has ever earned 210 billion US dollars before; this is a staggering figure, and its cash flow allows it to crush all of you.
TSMC is the same. I should make a video to explain CoWoS. Its CoWoS packaging alone is already a monopoly because you must stack many chips together now, which relies on advanced packaging to turn them into a massive chip to reduce latency. Right now, Cerebras is taking a different path by creating a single massive chip, and there is also the EMIB path where Intel builds a bridge. I should do a professional analysis for everyone to explain the difference between the two. However, while you are still working on EMIB, TSMC has already evolved CoWoS. CoWoS has now transitioned to being placed on glass substrates—square glass substrates—which saves a lot of money and removes size limitations. Therefore, unless mainland China attacks Taiwan, TSMC will be absolutely invincible within the next 5 years.
VRT is a stock I am currently buying. Everyone can see I only bought around 20-something shares because I only started chasing it when it was already very high. When chasing it at such a high price, I dare not buy too many for my core holding; the core holding is only 10 to 20 shares, and I buy whenever it drops. This is a very safe approach. I can tell you that liquid cooling is an irreversible trend, and all new technologies must use liquid cooling, and VRT is the leader in liquid cooling.
Palantir goes without saying. Palantir only has a turnover of 7 to 8 billion a year today. What kind of company is this? It is captive software for large corporations and governments, meaning once they help you implement it, you can never do without it again. It serves as the underlying architecture that helps you gather all your data and then make strategic decisions. You should compare a company like this to IBM. Even IBM still commands a multiple of a dozen or twenty times today, and IBM is clearly losing market share, as is Oracle. The pricing power of these companies is immense. While some people are attempting to compete with Palantir now, they have not built up to its level yet. It possesses 10 to 20 years of experience in organizing all data, making strategic decisions, and executing them, which no other company can offer at the moment. Originally, I always thought its price was ridiculously expensive, but as it kept falling, it became not so expensive. Right now it sits at just a 100x PE ratio, which is not that expensive. Over the next year, its business is bound to double, which will bring it down to a 50x PE ratio.
Tempus AI constantly has a lot of news, with many people partnering with it. You can see it rose 4.81% over the past 5 days, while Roblox rose over 11%. My impression is that these stocks seem to be brewing. After completing their accumulation at the bottom, they will explode upward once the hardware phase reaches a certain stage. Hynix goes without saying, I do not even know how to put it. Hynix is the fifth company in the world to earn over 100 billion US dollars, yet its current market cap is only 700 billion, giving it a PE ratio of just a few times. Over the next two years, it is still going to double and double again. Of course, we were very lucky to buy in at 300 dollars. Now that it has fallen for several consecutive days, should we average down? Since it has dropped for two straight days, I will average down just a little bit.
AVAV seems to have stabilized recently. I started selling it slowly around 160. Actually, we made a very great profit on it the first time, selling at 220 dollars and 410 dollars, and then bought back when it dropped. It mainly lost a large contract. The company it acquired, BlueHalo, had a major contract that was factored into the acquisition figures, and that contract failed, which caused such a massive pullback. Now it seems to have found a bottom. Alright, let's stop this section here.
免責聲明： 本內容提供之報告內容係根據本公司認可之資料來源，並基於特定日期所做之判斷，但不保證其完整性或正確性，報告中所有的意見及預估，如有變更恕不另行通知。
本研究報告所載之投資資訊，僅提供客戶做為一般投資參考，並非針對特定對象提供專屬之投資建議。 文中所載資訊或任何意見，不構成任何買賣有價證券或其他投資標的之要約、宣傳或引誘等事項。對於本投資報告所討論或建議之任何證券、投資標的，或文中所討論或建議之投資策略，投資人應就其是否適合本身財務狀況與投資條件，進一步諮詢財務顧問的意見。