【冰質財經】從CHATGPT到AI伺服器，臺灣科技島的新產業革命

作者：納蘭雪敏

臺灣科技產業經歷過幾次重要的革命，上一次重大的革命是2007年6月29日正式發售的iPhone，開啟了行動電話的革命，而2022年11月30日開始推出CHATGPT 以後，AI的應用正式進入實用階段，在2023年開始爆發，在美國現在被認為與AI時代最相關的就是Nvidia，因為新的AI伺服器需要大量採用GPU進行運算，因為GPU具有平行處理的能力，此外GPU可以進行浮點運算，像是Nvidia並針對GPU進行很多AI特殊的優化，比如Tensor Core增加深度學習的矩陣運算，而且提供軟體來進行接軌他們的產品，比如TensorFlow和PyTorch支援CUDA，靠著這幾點使得Nvidia成為AI的王者。

臺灣在這次的AI革命，為什麼角色特別重要？比如廣達在近期5月12日說到AI機櫃伺服器的價格是一般伺服器的3倍，在伺服器各品牌的市佔率之前8155博智法說會提供過市佔率的調查資料。

【冰質台股】20230512(五)-財報大轉身！廣達轉身AI概念股、大同漲停板，發生了什麼事？台股收15502點 (iceessencecapital.blogspot.com)

但從最源頭台積電替Nvidia生產GPU與SoC，AI是需要高性能運算的技術，所以台積電做為重要的晶片製造商，GPU對他消化產能有很大的幫助，當前先進製程技術全球只剩下INTEL與台積電和三星，台積電具有最高的領先地位，因為早期ARM的經驗，對於節能要求極高，使得台積電對當前製造GPU有極高的技術地位，台積電的入榜也不意外。

再者是ABF族群，被譽為ABF三雄的3037欣興、3189景碩、8046南電，但ABF並非只有臺灣可以做， Intel、AMD、Nvidia，有四家主要的重點供應商，Ibiden，Shinko，Unimicron和AT&S，所以我個人認為AI世代我會更重視的只有3037欣興。

再來是上游的CCL，銅箔基板相關我則比較看好2383台光電與6274台耀，所以這四家公司加上2382廣達，可能說是本次AI伺服器概念股中的第一梯隊。

晶圓製造：2330台積電
ABF:3037欣興
CCL:2383台光電( ELITE MATERIAL )與6274台耀(Taiwan Union )

系統廠：2382廣達

這個革命的來臨，其實最驚人的是他已經完全不需要人去記憶太多資訊，實際上有Google以後所有的教育都不應該再以記憶為重點，而是以理解概念、應用概念、實際應用後優化與修正為主，所以當前學校教的東西可以說完全變成了垃圾，甚至AI本身就能加快教育的速度，現在CHATGPT是因為很多知識版權的問題，如果他囊括了更多專業教科書與所有書籍的版權內容，甚至影片內容，他能直接變成老師，直接以問答的方式幫你解決實際你需要的問題。

課程的設計則是以各種實際的案例為主，學校的教育將逐漸失去意義，比如醫學院學生背骨頭名稱被認為是基礎課程，但AI未來可能直接告訴你，根據耳鼻喉科的臨床經驗，實際需要學習的骨頭僅有哪些，你就背那些就好，並不需要把全部的骨頭背起來，這才是有效率節省時間的學習方式，當前的教育模式犯了一個很嚴重的錯誤，就是把很多不需要的知識也浪費時間教育，這會使得學生浪費大量的時間學習的是老師會的東西，甚至老師自己也背不起，他只是因為年紀比較大，課本這樣編，他才這樣教。

在新的AI時代，像是臺灣這種教育將會使得學生繼續浪費大量的時間，而國際間的學校如果開始教授AI，雙方的差距不是幾十倍，幾百倍，而是0跟1，未來這些與AI使用幾十年的學生相比，這些還在學傳統教育的學生將變得毫無競爭力，只能做勞力工作。

但這不代表勞力工作賺不到錢沒有價值，而是如果一開始你就要從事勞力工作何必去大學接受高等教育，或者如果我們把AI與電腦知識的應用放到小學甚至幼稚園，甚至請AI設計具有高效率的教案，應用80/20法則來快速掌握那些重要的詞彙與聲音。

時間將會被大幅節省，這代表學生的競爭力將數千倍的增長，甚至十幾歲的學生就會使用AI進行軟體設計、翻譯、藝術創作、統計分析、金融投資、處理心理與感情生活、進行商業簡報、商業企劃。

未來的時代，你的下一代越早接觸將越早遠遠的甩開其他競爭者！

Author: Nalan Xuemin

The Taiwanese technology industry has undergone several major revolutions. The last significant one was the launch of the iPhone on June 29, 2007, which kicked off the mobile phone revolution. Starting from November 30, 2022, with the launch of CHATGPT, the application of AI officially entered the practical phase, and began to explode in 2023. Nvidia is now regarded in the United States as the most relevant company in the AI era, as new AI servers require a large amount of GPU processing. This is because GPUs have the ability for parallel processing and floating-point computation. Nvidia has carried out many AI-specific optimizations for its GPUs, such as the Tensor Core which increases matrix operations for deep learning, and software to support their products, like TensorFlow and PyTorch that support CUDA. These factors have made Nvidia the king of AI.

Why is Taiwan playing a particularly important role in this AI revolution? For instance, Quanta Computer recently announced on May 12 that the price of an AI rack server is three times that of a regular server. Before the market share of various server brands, 8155 ALLIED CIRCUIT said it would provide market share survey data.

【冰質台股】20230512(五)-財報大轉身！廣達轉身AI概念股、大同漲停板，發生了什麼事？台股收15502點 (iceessencecapital.blogspot.com)

From the source, TSMC is producing GPU and SoC for Nvidia. AI requires high-performance computing technology, so TSMC, as an important chip manufacturer, greatly benefits from GPUs to absorb capacity. Currently, only INTEL, TSMC, and Samsung are left in the world with advanced process technology, and TSMC has the leading position. Because of the early experience with ARM, which requires extremely high energy efficiency, TSMC has a very high technical position in current GPU manufacturing. It is not surprising that TSMC is on the list.

Moreover, there is the ABF group, known as the three heroes of ABF, 3037 Unimicron, 3189 Kinsus, and 8046 N.P.C. However, not only Taiwan can produce ABF, Intel, AMD, Nvidia have four key suppliers: Ibiden, Shinko, Unimicron, and AT&S. So, in the AI era, the company I personally value more is 3037 Unimicron.

Then there is the upstream CCL, I am more optimistic about 2383 Elite Material and 6274 Taiwan Union. Therefore, these four companies, along with 2382 Quanta, could be said to be the first echelon in the AI server concept stock.

Wafer Manufacturing: 2330 TSMC ABF: 3037 Unimicron CCL: 2383 Elite Material, 6274 Taiwan Union System Manufacturer: 2382 Quanta

The arrival of this revolution is actually amazing because it no longer needs people to memorize too much information. Since Google, all education should not focus on memorization, but understanding concepts, applying concepts, and optimizing and correcting after practical application. Therefore, what schools are currently teaching can be said to have become garbage. Even AI itself can accelerate the pace of education. Right now, CHATGPT is restricted due to many knowledge copyright issues. If it encompasses more professional textbooks and all copyrighted contents of books, even video content, it can directly become a teacher, directly solving your actual problems through Q&A.

Course design is mainly based on various practical cases. School education will gradually lose its meaning. For example, medical students memorizing bone names is considered a basic course, but AI in the future may directly tell you, according to the clinical experience of otolaryngology, what bones you actually need to learn, you just memorize those, and there is no need to memorize all the bones. This is an efficient and time-saving learning method. The current education model has made a serious mistake in teaching a lot of unnecessary knowledge, causing students to waste a lot of time learning what the teacher knows, and even the teacher cannot remember everything. It's just that because the teacher is older and the textbook is written in this way, he/she teaches this way.

In the new AI era, a place like Taiwan will continue to waste students' time on this type of education, while schools internationally starting to teach AI, the gap between the two is not tens of times, hundreds of times, but 0 and 1. In the future,these students who still learn traditional education will become uncompetitive compared to those who have been using AI for decades, and can only do manual labor.

However, this does not mean that manual labor jobs are worthless and can't make money. Rather, if you are going to engage in manual labor from the beginning, why go to college to receive higher education? Or if we put the application of AI and computer knowledge into elementary school or even kindergarten, and even have AI design highly efficient teaching plans, apply the 80/20 rule to quickly grasp those important words and sounds.

A lot of time will be saved, which means that the competitiveness of students will grow thousands of times. Even teenagers will use AI for software design, translation, artistic creation, statistical analysis, financial investment, handling psychological and emotional life, conducting business presentations, and business planning.

In the future, the sooner your next generation gets in touch, the sooner they will leave other competitors far behind!