YouTube 頻道

發財二極體

@發財二極體 · —

馬瑞辰 Jui-Chen Ma 教授 x 林嘉洤Jia-Chuan Lin 教授當金融界的巨頭遇上科技產業的博士，讓兩位專家，帶領你一邊跟上金融投資趨勢，一邊搞懂台灣科技產業。聽教授們的深入解析，揭開科技產業的核心秘密，裡面還有許多意想不到的專業知識，讓投資者運用在投資策略上，預測未來、超前佈局！【馬瑞辰 Jui-Chen Ma 教授】專長領域：房地產金融學、金融交易實務、金融行銷學重要學經歷：北京大學政府管理學院博士東京大學碩士麻省理工學院(M.I.T.)學士清華大學兼任教授清華安富金融工程研究中心執行長【林嘉洤Jia-Chuan Lin 教授】專長領域：半導體科技、奈米科技、 IC設計與製造、光電技術重要學經歷：國立臺北大學電機系特聘教授國立臺北大學行政副校長聖約翰科技大學副校長國立成功大學電機博士

查看 YouTube 頻道

PERFORMANCE SNAPSHOT

核心表現

公開快照中的近期可用數據

資料維度載入中

訂閱數234,000

觀看數1,393,719

影片數175

CONTENT INTELLIGENCE

近期內容表現

以同一觀看單位比較最多三支公開影片

3 支可比較內容

AI模型削價競爭，誰才是真正贏家？傑文斯悖論一次看懂 Who Really Wins the AI Model Price War? Understanding the Jevons ParadoxYouTube 影片

274

Kimi K3為何這麼強？看懂MoE「專家混合」架構，AI算力如何越用越省？Why Kimi So Powerful? How AI Can Use Compute More EfficientlyYouTube 影片

362

一次搞懂Token與API：AI到底怎麼運作、怎麼計費？Understanding Tokens and APIs: How Does AI Work and How Is It Billed?YouTube 影片

1,095

近期影片

AI模型削價競爭，誰才是真正贏家？傑文斯悖論一次看懂 Who Really Wins the AI Model Price War? Understanding the Jevons Paradox

#AI #JevonsParadox #傑文斯悖論 #Token #LLM #ChatGPT #KimiK3 #AIAgent #生成式AI #人工智慧 #GPU #HBM #資料中心 #SaaS #SI #AI模型 #AI產業 #AI投資 #雲端運算 #科技科普 Token越便宜，AI公司真的會越來越難賺嗎？還是AI反而將迎來更大的成長循環？近年來，大型語言模型（LLM）的能力持續提升，但Token價格卻快速下降。許多人擔心，模型業者之間的價格競爭，是否會壓縮獲利、降低資本支出，甚至拖慢AI產業發展。然而，從經濟學角度來看，答案可能正好相反。本集節目中，林教授透過**「傑文斯悖論（Jevons Paradox）」**，解析AI產業正在發生的重要變化：當技術進步降低成本後，使用量往往會大幅增加，最終帶動整體市場規模持續擴張。除了探討AI模型的Token價格戰之外，也深入分析AI Agent、SaaS、SI軟體服務，以及GPU、HBM、資料中心等AI基礎建設，如何共同形成新一波AI產業成長循環。本集重點： ✔ 什麼是傑文斯悖論（Jevons Paradox）？ ✔ 為什麼Token價格下降，不一定代表AI公司獲利惡化？ ✔ AI模型降價，為何可能帶動更多企業導入AI？ ✔ AI Agent普及後，Token與API需求將如何成長？ ✔ 為什麼Token越便宜，反而可能需要更多GPU與算力？ ✔ AI算力需求增加，如何帶動HBM、記憶體、儲存與網路設備？ ✔ 電力、資料中心、液冷散熱等基礎建設為何仍具成長動能？ ✔ SaaS、SI與AI軟體服務商將迎來哪些新商機？ ✔ 模型成本下降後，AI產業鏈的定價權將如何重新分配？節目中更以Buffet吃到飽與雞排降價等生活化案例，說明「價格下降、需求反而增加」的經濟現象。當AI模型成本降低，企業便能以更低成本將AI整合到產品與工作流程中；更多AI應用、更多AI Agent、更頻繁的API呼叫，將進一步推升整體Token消耗量與算力需求。這意味著，AI產業真正的價值，不只是模型本身，而是從模型延伸到應用、軟體、資料中心與硬體設備所形成的完整生態系。未來AI市場真正值得關注的，或許不是**「每個Token可以賣多少錢」，而是「全球每天會新增多少Token的使用量」**。如果你想從產業鏈、商業模式與投資角度，看懂Token價格戰、AI Agent興起，以及AI軟硬體生態的未來發展，這集內容值得收藏！ Will cheaper Tokens really make AI companies less profitable—or could lower AI costs actually trigger an even bigger growth cycle? In recent years, the capabilities of Large Language Models (LLMs) have continued to improve, while Token prices have fallen rapidly. This has raised concerns that intensifying price competition among AI model providers could squeeze profit margins, reduce capital expenditures, and eventually slow down the development of the AI industry. However, from an economic perspective, the answer may actually be the opposite. In this episode, Professor Lin uses the Jevons Paradox to explain a major trend emerging in the AI industry: when technological advances reduce the cost of using a resource, demand for that resource may actually increase significantly, ultimately expanding the overall market. Beyond the Token price war among AI model providers, this episode explores how AI Agents, SaaS, system integrators (SI), GPUs, HBM, and data centers could collectively form the next major growth cycle for the AI ecosystem. Key Topics Covered: ✔ What is the Jevons Paradox? ✔ Why doesn't a decline in Token prices necessarily mean lower profits for AI companies? ✔ Why could cheaper AI models encourage more businesses to adopt AI? ✔ How could the rise of AI Agents drive Token consumption and API demand? ✔ Why could cheaper Tokens actually lead to greater demand for GPUs and computing power? ✔ How could rising AI workloads drive demand for HBM, memory, storage, and networking equipment? ✔ Why could power infrastructure, data centers, and liquid cooling continue to benefit from AI growth? ✔ What new opportunities could emerge for SaaS, system integrators, and AI software providers? ✔ How could lower model costs reshape pricing power across the AI value chain? The episode also uses everyday examples such as all-you-can-eat buffets and discounted fried chicken to explain the economic phenomenon in which lower prices can actually stimulate greater consumption. As AI model costs decline, businesses can integrate AI into their products and workflows at a lower cost. More AI applications, more AI Agents, and more frequent API calls could ultimately drive a significant increase in Token consumption and computing demand. This means that the true value of the AI industry may not lie solely in the AI models themselves, but in the broader ecosystem extending from models to applications, software, data centers, and hardware infrastructure. The key question for the future of AI may not simply be "How much can each Token be sold for?" but rather: "How many more Tokens will the world consume every day as AI becomes cheaper and more accessible?" If you want to understand the Token price war, the rise of AI Agents, and the future of the AI hardware and software ecosystem from an industry-chain, business-model, and investment perspective, this episode is worth watching and saving. 【發財二極體相關平台】 YouTube頻道：https://www.youtube.com/channel/UCVrP FB粉絲頁：https://www.facebook.com/ 【聯絡我們】 email: [email protected] 🇹🇼 台股概念股類別公司代號主要受惠邏輯 AI晶片/ASIC 世芯-KY 3661 AI ASIC、客製化晶片 AI晶片/ASIC 創意 3443 ASIC設計、AI加速器 AI晶片台積電 2330 AI GPU/ASIC先進製程、CoWoS GPU/ASIC周邊智邦 2345 AI高速網通、交換器 AI伺服器廣達 2382 AI伺服器、整機櫃 AI伺服器緯穎 6669 CSP/雲端AI伺服器 AI伺服器緯創 3231 AI伺服器、資料中心 AI伺服器鴻海 2317 AI伺服器、機櫃、資料中心 PCB 欣興 3037 ABF載板、AI/HPC PCB 金像電 2368 AI伺服器高速PCB PCB 台燿 6274 高速高頻CCL PCB 台光電 2383 AI高速傳輸材料散熱奇鋐 3017 AI伺服器液冷、散熱模組散熱雙鴻 3324 液冷、冷板、散熱散熱健策 3653 AI伺服器散熱、機構件散熱富世達 6805 AI伺服器液冷/機構件電源台達電 2308 AI伺服器電源、電力管理、液冷電源光寶科 2301 AI伺服器電源電源康舒 6282 AI資料中心電源連接器嘉澤 3533 CPU/GPU Socket、AI伺服器連接器貿聯-KY 3665 高速線材、AI伺服器電源/連接網通智邦 2345 800G/高速交換器網通台光電 2383 高速網通材料記憶體南亞科 2408 DRAM需求記憶體華邦電 2344 記憶體、AI周邊被動元件國巨 2327 AI伺服器MLCC 被動元件華新科 2492 MLCC、AI電源需求 🇺🇸 美股概念股類別公司代號主要受惠邏輯 GPU NVIDIA NVDA AI訓練/推論GPU GPU AMD AMD Instinct GPU、AI加速器 ASIC Broadcom AVGO AI ASIC、網通晶片 CPU/AI Intel INTC AI CPU、封裝/資料中心

觀看 274按讚 0留言 0

觀看影片 ↗

Kimi K3為何這麼強？看懂MoE「專家混合」架構，AI算力如何越用越省？Why Kimi So Powerful? How AI Can Use Compute More Efficiently

#AI #MoE #MixtureOfExperts #KimiK3 #AI模型 #Token #AI推論 #GPU #AI算力 #人工智慧 #LLM #生成式AI #AI產業 #資料中心 #AI晶片 Kimi K3為何能在擁有超大規模參數的同時，兼顧強大性能與低Token成本？關鍵之一，就是近年AI模型越來越受到重視的「MoE（Mixture of Experts，專家混合）」架構。當我們談到大型語言模型時，直覺會認為模型參數越多，就代表需要越大的算力。但MoE的核心思維恰恰不同：不是每一道題目，都需要所有專家一起出動。本集節目中，林教授用「專家團隊」與「Router路由器」的概念，帶大家認識MoE架構究竟是如何運作，以及它為什麼可能成為未來AI模型的重要發展方向。您將了解： ✔ 什麼是MoE（Mixture of Experts，專家混合）？ ✔ 傳統大型模型與MoE架構有什麼不同？ ✔ AI模型裡的「Expert專家」到底是什麼？ ✔ Router如何判斷一個問題應該交給哪些專家處理？ ✔ 為什麼簡單問題不需要動用全部模型？ ✔ AI如何根據問題的複雜程度，動態分配運算資源？ ✔ MoE如何降低Token運算量與AI推論成本？ ✔ 為什麼「參數規模很大」不代表每次推論都需要使用全部算力？ ✔ MoE對未來AI晶片、GPU算力與資料中心有什麼影響？節目中以「三個臭皮匠勝過一個諸葛亮」、「殺雞焉用牛刀」等生活化的例子，讓大家理解MoE最重要的概念：把不同能力的專家分門別類，再由Router根據問題內容，挑選最適合的專家出場。例如簡單的數學問題，可能只需要一位數學專家；遇到更複雜的研究型問題，則可以調度更多、更強的專家共同處理。如此一來，模型可以在維持龐大「總參數規模」的同時，避免每次推論都啟動全部參數，進而提高算力使用效率。這也讓MoE不只是AI模型架構上的技術變化，更與未來的AI推論成本、Token經濟、GPU需求、AI資料中心以及AI Agent密切相關。如果你想了解Kimi K3等新一代AI模型背後的技術邏輯，以及為什麼未來AI發展不只是「堆更多GPU」，而是如何更有效率地使用算力，這集節目值得一看！ Why can Kimi K3 deliver strong performance and low Token costs while featuring an extremely large number of parameters? One of the key factors is the increasingly important MoE (Mixture of Experts) architecture. When we talk about large language models, it is natural to assume that more parameters require more computing power. But MoE takes a different approach: not every question requires every expert to work at the same time. In this episode, Professor Lin uses the analogy of a team of specialists and a Router to explain how the MoE architecture works and why it could become an important direction for the future development of AI models. In this episode, you'll learn: ✔ What is MoE (Mixture of Experts)? ✔ How is MoE different from traditional large language models? ✔ What exactly are “Experts” inside an AI model? ✔ How does the Router decide which experts should handle a particular task? ✔ Why don't simple questions require the entire model to be activated? ✔ How can AI dynamically allocate computing resources based on the complexity of a task? ✔ How can MoE reduce Token usage and AI inference costs? ✔ Why doesn't a model with a huge number of total parameters need to activate all of them for every inference task? ✔ What impact could MoE have on future AI chips, GPUs, and data centers? The episode uses simple analogies such as “three ordinary people can outperform one brilliant strategist” and “don't use a sledgehammer to crack a nut” to explain the core concept behind MoE: divide AI capabilities into different specialized experts, then let the Router select the most relevant experts for each task. For example, a simple math problem may only require one mathematical expert, while a much more complex research-level question may require several more capable experts to work together. In this way, a model can maintain a massive total parameter count without activating all of its parameters for every inference task, improving overall computing efficiency. This makes MoE more than just a change in AI model architecture. It is also closely connected to the future of AI inference costs, the Token economy, GPU demand, AI data centers, and AI Agents. If you want to understand the technology behind next-generation AI models such as Kimi K3—and why the future of AI is not simply about adding more GPUs, but about using computing resources more efficiently—this episode is worth watching! 【發財二極體相關平台】 YouTube頻道：https://www.youtube.com/channel/UCVrP FB粉絲頁：https://www.facebook.com/ 【聯絡我們】 email: [email protected] 台股概念股公司股號概念台積電 2330 AI GPU／ASIC、先進製程聯發科 2454 AI ASIC、Edge AI 世芯-KY 3661 AI ASIC／客製化晶片創意 3443 AI ASIC／高速運算智原 3035 ASIC／AI Chip 廣達 2382 AI Server／AI Data Center 緯穎 6669 CSP／AI Server 緯創 3231 AI Server／雲端基礎建設鴻海 2317 AI Server／AI Data Center 英業達 2356 AI Server 技嘉 2376 GPU Server／AI Server 華碩 2357 AI Server／GPU 智邦 2345 AI Networking／高速交換器台達電 2308 AI Data Center Power／散熱光寶科 2301 AI Server／電源奇鋐 3017 AI Server 液冷／散熱雙鴻 3324 AI GPU／液冷散熱 🇺🇸 美股概念股公司股號概念 NVIDIA NVDA GPU／AI Inference／MoE AMD AMD AI GPU／Inference Broadcom AVGO AI ASIC／Networking Marvell Technology MRVL Custom XPU／AI Networking Micron Technology MU HBM／高頻寬記憶體 Arista Networks ANET AI Data Center Networking Microsoft MSFT Azure／AI Model／Inference Alphabet GOOGL Gemini／TPU／Google Cloud Amazon AMZN AWS／Bedrock／AI Inference Oracle ORCL OCI／AI Cloud Meta Platforms META Llama／AI Inference CoreWeave CRWV AI Cloud／GPU Compute Nebius NBIS AI Cloud／GPU Compute Vertiv VRT AI Data Center Power／Cooling

觀看 362按讚 1留言 0

觀看影片 ↗

一次搞懂Token與API：AI到底怎麼運作、怎麼計費？Understanding Tokens and APIs: How Does AI Work and How Is It Billed?

#token #ai #openai #google #anthropic #api AI時代，Token 幾乎已經成為每天都會聽到的關鍵字。ChatGPT、Gemini、Claude 等 AI 模型的使用費用，經常都以 Token 作為計價單位。但問題來了：一個 Token 到底有多大？為什麼不能直接用「字數」計算？而我們常聽到的 API，又究竟是什麼？本集節目中，林教授將用最生活化的方式，帶大家一次搞懂 AI 世界裡兩個非常重要的概念——Token 與 API。您將了解： ✔ Token 到底是什麼？一個 Token 有多大？ ✔ 為什麼 AI 不直接用「字數」作為計算單位？ ✔ 英文、中文在 Token 計算上有什麼差異？ ✔ Token 為什麼可以反映 AI 模型背後的運算與硬體資源？ ✔ API 的全名是什麼？它到底在做什麼？ ✔ 為什麼不同軟體、資料庫與 AI 模型之間需要 API？ ✔ AI Agent 為什麼需要透過 API 串接不同模型與服務？ ✔ 為什麼使用 AI API 就會產生費用？ ✔ AI 模型、API、Token 三者之間到底是什麼關係？節目中更用「國家之間的外交官」以及「餐廳服務生」來比喻 API，讓原本看似艱深的軟體架構變得非常直觀：AI 模型就像雲端廚房，而 API 就像負責傳遞需求與結果的服務生；沒有 API，再強大的 AI 模型也很難直接被不同應用程式使用。從 Token 的運算成本，到 API 的資料串接，再到 AI Agent 與多模型協作，這些看似簡單的名詞，其實正是理解 AI 商業模式、雲端運算與生成式 AI 產業鏈的重要基礎。如果你想真正搞懂 ChatGPT 背後是怎麼運作、AI 為什麼可以計價，以及未來 AI Agent 為什麼會帶來大量 API 與 Token 需求，這集內容值得收藏！ n the age of AI, Token has become one of the most frequently heard terms. AI models such as ChatGPT, Gemini, and Claude often use Tokens as the basis for calculating usage costs. But what exactly is a Token? How large is one Token? Why don't AI companies simply charge by the number of words? And what exactly is an API that we hear about so often? In this episode, Professor Lin uses simple, real-world analogies to explain two fundamental concepts in the AI world: Tokens and APIs. In this episode, you'll learn: ✔ What exactly is a Token, and how large is one Token? ✔ Why doesn't AI simply calculate usage based on the number of words? ✔ How does Tokenization differ between English and Chinese? ✔ Why can Token usage reflect the computing and hardware resources behind AI models? ✔ What does API stand for, and what does it actually do? ✔ Why do different software applications, databases, and AI models need APIs? ✔ Why do AI Agents rely on APIs to connect different models and services? ✔ Why does using an AI API generate costs? ✔ What is the relationship between AI models, APIs, and Tokens? The episode uses two simple analogies to make these complex concepts easy to understand: an API is like a diplomat connecting different countries, or a waiter connecting a restaurant kitchen with its customers. An AI model is like a cloud-based kitchen, while the API acts as the waiter that delivers requests to the model and brings the results back to the user. Without APIs, even the most powerful AI models would have difficulty being integrated into different applications and services. From the computing cost represented by Tokens, to data connections through APIs, and ultimately to AI Agents and multi-model collaboration, these seemingly simple concepts are fundamental to understanding the AI business model, cloud computing, and the broader generative AI ecosystem. If you want to understand how ChatGPT works behind the scenes, why AI can be billed based on usage, and why the rise of AI Agents could drive massive demand for APIs and Tokens, this episode is a must-watch. 【發財二極體相關平台】 YouTube頻道：https://www.youtube.com/channel/UCVrP FB粉絲頁：https://www.facebook.com/ 【聯絡我們】 email: [email protected] 公司代號主要概念台積電 2330 AI GPU／ASIC、先進製程聯發科 2454 AI ASIC、Edge AI 廣達 2382 AI Server、雲端資料中心緯創 3231 AI Server、Cloud Infrastructure 緯穎 6669 CSP／AI Server 鴻海 2317 AI Server、資料中心、AI Factory 英業達 2356 AI Server 技嘉 2376 AI Server、GPU Server 華碩 2357 AI Server、AI Cloud 智邦 2345 AI Data Center Networking 台達電 2308 AI Data Center Power 光寶科 2301 AI Server／電源創意 3443 AI ASIC／ASIC Design 世芯-KY 3661 AI ASIC 智原 3035 ASIC／AI Chip 美股第一層：最直接的 AI／Token／API 概念公司代號核心邏輯 NVIDIA NVDA GPU + Inference + AI Software Microsoft MSFT Azure + AI API + Copilot Alphabet GOOGL Gemini + Google Cloud + API Amazon AMZN AWS + Bedrock + AI Infrastructure Oracle ORCL OCI + AI Cloud + GPU Cloud Meta META Llama + AI Inference CoreWeave CRWV AI Cloud／GPU Compute Nebius NBIS AI Cloud／GPU Compute 第二層：AI 算力與 ASIC 公司代號核心邏輯 AMD AMD Instinct GPU／AI Accelerator Broadcom AVGO AI ASIC／Networking Marvell MRVL Custom AI ASIC／Networking Micron MU HBM／Memory Arista Networks ANET AI Networking Vertiv VRT AI Data Center Power／Cooling

觀看 1,095按讚 2留言 0

觀看影片 ↗

AI 費用怎麼控管？從 Kimi K3 降價看企業如何用小成本發揮大算力 How to Control AI Costs? Lessons from Kimi K3’s Price Cut

#ai #token #kimik3 ＃推理＃訓練許多人想用 AI 提升工作效率，卻常被「按 Token 計價」的暴增費用嚇退，甚至擔心預算失控收到天價帳單！最新登場的 Kimi K3 不僅擁有 2.8 兆參數與開源優勢，更投下「價格震撼彈」——輸入與輸出費用比市面上主流封閉模型便宜了一半甚至三分之二！究竟什麼是「輸入（Input）」？什麼是「輸出（Output）」？為什麼「創作燒腦」會比「閱讀資料」貴這麼多？林教授這次用「請 AI 寫武俠小說」的超生動比喻，帶你一次搞懂 Token 算力計價機制與 3 個幫你節省 AI 費用的關鍵技術！ 📌 本集精華亮點📖 什麼是輸入 vs. 輸出？（請 AI 寫武俠小說比喻）輸入（Input / 閱讀）：你給 AI 參考的資料與背景設定。就像給機器人讀幾十本參考小說，過程相對單純、費用較便宜。輸出（Output / 創作）： AI 吸收資料後自己寫出全新章節。這是進行推理（Reasoning）的過程，也就是俗稱的「燒腦」，需要消耗大量算力，因此費用顯著高昂！ 💥 Kimi K3 價格震撼：輸入費用低至每百萬 Token $3 美金（其他頂級模型約 $10）；輸出費用僅 $15 美金（其他頂級模型高達 $30~$50），讓企業與個人使用 AI 的門檻大幅降低。 🛠️ 3 招省錢大作戰：AI 降本增效的核心技術蒸餾技術（Distillation）：不用把整部大書丟給 AI，僅挑選最精彩的重點精華輸入，從源頭省下 Token。大小模型梯隊搭配：短篇草稿或簡易任務先讓「小模型」跑，最後需要彙整巨著時再交給「大模型」。位元數與精度縮減（Quantization）：如同圓周率不一定要帶到 pi = 3.1415926，有時精簡到 3.14 甚至 3 即可，算力需求直接砍半！ Many people want to boost productivity with AI, but are often deterred by spiraling costs under token-based pricing, fearing budget overruns and astronomical bills! The newly released Kimi K3 not only features a 2.8-trillion parameter architecture and open-source advantages, but also drops a "price bombshell"—with input and output costs reduced by half to two-thirds compared to mainstream closed-source models! What exactly are "Input" and "Output"? Why is "creative reasoning" so much more expensive than "reading data"? In this episode, Professor Lin uses a vivid analogy—"asking AI to write a wuxia novel"—to help you understand token pricing mechanisms and master 3 key technologies to slash your AI expenses! 📌 Episode Highlights📖 What is Input vs. Output? (The Wuxia Novel Analogy) Input (Reading): The reference materials and context you provide to the AI. It's like feeding dozens of reference novels to a robot—a relatively straightforward process with lower costs. Output (Creation): The brand-new chapters written independently by the AI after absorbing the data. This involves Reasoning (commonly referred to as "heavy brainwork"), which consumes significant compute resources and results in substantially higher costs! 💥 Kimi K3 Pricing ShockInput costs as low as $3 USD per million tokens (compared to ~$10 USD for other top-tier models). Output costs at just $15 USD per million tokens (compared to $30–$50 USD for other top-tier models). Significantly lowers the barrier to entry for both enterprises and individuals. 🛠️ 3 Cost-Cutting Strategies: Key Technologies for AI Cost Reduction & EfficiencyModel Distillation: Instead of feeding an entire encyclopedia to the AI, select only the most essential highlights to cut token consumption at the source.Tiered Model Deployment (Large & Small Models): Route short drafts or simple tasks to smaller models first, and engage large models only when compiling the final major work.Quantization (Precision Reduction): Just as $\pi$ doesn't always need to be calculated to $3.1415926$ (sometimes $3.14$ or even $3$ suffices), reducing bit precision cuts compute requirements in half! 【發財二極體相關平台】 YouTube頻道：https://www.youtube.com/channel/UCVrP FB粉絲頁：https://www.facebook.com/ 【聯絡我們】 email: [email protected] 🇹🇼 台股企業私有化部署與系統整合：精誠 (6214)、零壹 (3029)、邁達特 (6112) 邊緣運算與端側 AI：聯發科 (2454)、華碩 (2357)、研華 (2395) 客製化推理晶片 (ASIC)：世芯-KY (3661)、創意 (3443) 🇺🇸 美股邊緣 AI 與端側晶片： Qualcomm (QCOM)、Apple (AAPL) AI 推理與客製化晶片： Broadcom (AVGO)、Marvell (MRVL) AI 軟體與企業級落地： Palantir (PLTR)、Salesforce (CRM) 輕量化模型與蒸餾技術推手： Meta (META)

觀看 98按讚 0留言 0

觀看影片 ↗

AI前沿模型震撼彈！Kimi K3 開源 AI 能比肩國際巨頭？A Bombshell in Frontier AI! Can Open-Source AI Match Tech Giants?

#ai #chatgpt #anthropic #moonshot #opensource 經歷了世界盃的熱血沸騰與近期全球股市的高空彈跳，AI 科技圈又傳來重磅震撼彈！繼 DeepSeek 之後，中國 AI 新星「月之暗面（Moonshot AI）」發表了最新大語言模型 Kimi K3。這款新模型不僅具備 2.8 兆（2.8T）的驚人參數規模，更堅持「開源」路線，直接強勢對標 OpenAI、Google 與 Anthropic 等頂級封閉式模型！到底什麼是開源與封閉？模型的「參數」又代表什麼意思？本集馬教授與林教授將帶大家用最白話的方式拆解這波全新的 AI 浪潮！ 📌 本集精華亮點🔓 開源 vs 🔒 封閉式模型哪裡不同？開源模型（Open-Source）：使用者能直接取得模型核心參數，可下載至自有環境進行深入客製化與微調。封閉式模型（Closed-Source）：模型本體在雲端，使用者僅能透過 API 連結輸入並取得答案，無法獲取底層參數。 🧠 模型「參數（Parameters）」到底是什麼？林教授用「愛因斯坦的大腦容量」做了絕妙比喻——參數就像是 AI 思考機器人的「智力與腦容量」。雖然不是絕對的一對一，但參數規模越大，通常代表模型的思考與智慧容量越高！ ⚔️ 2.8 兆參數的震撼！全球頂尖模型大比拼過去開源模型常被詬病參數規模較小，但 Kimi K3 一舉將參數推升至 2.8 兆，一舉跨入與國際一線封閉式模型平起平坐的戰場！歡迎在下方留言告訴我們：你平時比較喜歡使用「開源 AI」還是「封閉式 AI」呢？👇 喜歡我們的內容，請別忘了【訂閱、按讚、分享】並開啟小鈴鐺！ Following the excitement of the World Cup and the recent rollercoaster ride in global stock markets, another major bombshell has dropped in the AI tech sphere! Following DeepSeek, China's rising AI star "Moonshot AI" has released its latest large language model, Kimi K3. This new model not only boasts a staggering parameter scale of 2.8 trillion (2.8T), but also stays committed to an open-source route—directly competing with top proprietary (closed-source) models like OpenAI, Google, and Anthropic! What exactly is the difference between open-source and closed-source? And what do model "parameters" actually mean? In this episode, Professor Ma and Professor Lin will break down this brand-new AI wave in the plainest terms! 📌 Episode Highlights 🔓 Open-Source vs. 🔒 Closed-Source Models: What's the Difference? Open-Source Models: Users can directly access the model's core parameters and download them into their local environment for deep customization and fine-tuning. Closed-Source Models: The model resides in the cloud. Users can only connect via APIs to input queries and receive answers without accessing the underlying parameters. 🧠 What Exactly Are Model "Parameters"? Professor Lin makes a brilliant analogy using "Albert Einstein's brain capacity"—parameters are like the "intelligence and brain capacity" of an AI thinking robot. While not a strict one-to-one correlation, a larger parameter scale generally signifies a higher capacity for reasoning and intelligence! ⚔️ The 2.8 Trillion Parameter Shocker! Global Top Models Face Off In the past, open-source models were often criticized for their smaller parameter scales. However, Kimi K3 pushes the boundary to 2.8 trillion parameters, stepping right onto the battlefield to go toe-to-toe with top-tier international closed-source models! Let us know in the comments below: Do you usually prefer using "Open-Source AI" or "Closed-Source AI"? 👇 If you enjoyed our content, don't forget to [Subscribe, Like, Share], and turn on the notification bell! 【發財二極體相關平台】 YouTube頻道：https://www.youtube.com/channel/UCVrP FB粉絲頁：https://www.facebook.com/ 【聯絡我們】 email: [email protected] 🇺🇸 美股 LLM 與 AI 算力核心概念股超大參數模型（如 2.8 兆參數）的訓練與推理，高度依賴巨型 GPU 集群、高速網絡傳輸以及雲端基礎設施。 1. 模型開發、開源生態與雲端巨頭股票名稱股票代號產業地位與核心優勢Meta PlatformsMETA開源 LLM 領頭羊（Llama 系列），推動全球開源 AI 生態系與邊緣 AI 發展。MicrosoftMSFT獨家合作 OpenAI，Azure 雲端提供全球最大的 LLM 訓練與 API 推理算力。Alphabet (Google)GOOGL旗艦 Gemini 模型開發商，擁有自研 TPU 算力集群與 Google Cloud AI 服務。AmazonAMZNAWS Bedrock 提供多模型平台、投資 Anthropic (Claude)，並自研 Trainium 晶片。 2. AI 運算晶片與高速網路股票名稱股票代號產業地位與核心優勢NVIDIANVDA全球 AI 算力霸主。其 GPU（Hopper/Blackwell）為訓練兆級參數模型的不二之選。BroadcomAVGO超大規模 AI 資料中心網絡交換晶片（Tomahawk）與客製化 AI 晶片（ASIC）龍頭。AMDAMDInstinct MI300/MI350 系列晶片，提供開源社群與雲端廠商高效能 GPU 替代方案。MarvellMRVL專攻高速光互連（Optical Interconnect）晶片與雲端 ASIC 客製化設計。 🇹🇼 台股 AI 模型硬體與軟體整合概念股台灣在全球 AI 供應鏈中佔據「晶片代工」與「AI 伺服器硬體製造」的關鍵壟斷地位，同時本土系統整合商也正積極搶攻企業級 LLM 落地商機。 1. 晶片製造與 ASIC 設計服務股票名稱股票代號產業地位與核心優勢台積電2330全球 AI 晶片獨家代工。先進製程與 CoWoS 封裝包辦 NVIDIA、AMD、TPU 產能。世芯-KY3661高階 AI ASIC 設計服務龍頭，協助美系 CSP 雲端巨頭開發自研 AI 算力晶片。創意3443台積電轉投資 ASIC 設計服務大廠，專攻先進封裝與高階 HBM 介面 IP。 2. AI 伺服器與算力機櫃組裝股票名稱股票代號產業地位與核心優勢鴻海2317NVL72 / NVL36 AI 頂級機櫃組裝龍頭，全球 AI 伺服器市佔率極高。廣達2382北美四大雲端巨頭（Microsoft、Google、Amazon、Meta）AI 伺服器核心供應商。緯穎6669專注大型資料中心與 AI 伺服器整機櫃研發，算力基礎設施出口主力。川湖2059全球高階 AI 伺服器導軌獨霸廠商，承受超重 AI 機櫃支撐需求。 3. 企業級 LLM 微調與系統整合 (SI)股票名稱股票代號產業地位與核心優勢精誠6214台灣資服龍頭，協助企業進行開源模型（如 Llama、DeepSeek、Kimi）私有化部署與 Fine-tuning。零壹3029代理全球 AI 算力設備與雲端架構，提供企業端生成式 AI 解決方案落地。

觀看 194按讚 3留言 0

觀看影片 ↗

AI伺服器電力從48V到1V，電容為何成為關鍵？From 48V to 1V: Why Capacitors Are Critical to AI Server Power Delivery

#AI伺服器 #GPU #PowerDeliveryNetwork #PDN #電容 #MLCC #鉭電容 #鋁電容 #48V架構 #資料中心 #電源管理 #電力電子 #AI晶片 #高速運算 #HPC #科技科普 #半導體產業當我們談論 AI 晶片、GPU 與高速運算時，很少有人注意到另一個同樣重要的課題——供電系統。一顆先進 AI 晶片在運作時，實際使用的電壓往往不到 1V，但資料中心輸入的卻是數百伏特的交流電。這中間究竟經歷了哪些轉換？又有哪些關鍵元件在背後默默運作？本集節目中，林教授將帶大家深入了解 AI 伺服器的供電架構，以及電容在 Power Delivery Network（PDN）中的核心角色。您將了解： ✔ 為什麼資料中心輸入的是高壓交流電？ ✔ 為何晶片最終需要的是低電壓直流電？ ✔ 交流電轉直流電的關鍵元件是什麼？ ✔ AI伺服器為何從12V逐步升級到48V架構？ ✔ Power Delivery Network（PDN）到底是什麼？ ✔ 電容在整流、儲能與降壓過程中扮演什麼角色？ ✔ 為什麼不同環節需要不同種類的電容？ ✔ MLCC、鉭電容、鋁電容、固態電容與液態電容如何分工合作？ ✔ 接面電容與串聯電阻（ESR）對高速運算有何影響？節目中更以「大禹治水」作為生動比喻，帶領觀眾理解電能如何在伺服器系統中被管理、儲存與分配。從高壓交流電進入機櫃，到最終穩定地供應給GPU與AI加速器，每一個環節都考驗著電源設計與被動元件技術。隨著 AI 算力持續提升，供電系統的重要性已不亞於晶片本身。未來無論是 AI 資料中心、電動車快充系統，甚至下一代高效能運算平台，Power Delivery Network 與高階電容都將成為產業競爭的關鍵焦點。 When we discuss AI chips, GPUs, and high-performance computing (HPC), few people notice an equally critical topic: the power delivery system. While an advanced AI chip typically operates on less than 1V, the data center inputs hundreds of volts of Alternating Current (AC). What transformations occur in between? Which key components are working silently behind the scenes? In this episode, Professor Lin will take us on a deep dive into AI server power architectures and the pivotal role capacitors play in the Power Delivery Network (PDN). What you will learn: High-Voltage AC: Why do data centers use high-voltage AC power as their main input? Low-Voltage DC: Why do chips ultimately require low-voltage Direct Current (DC)? AC-to-DC Conversion: What are the critical components driving AC-to-DC conversion? The 48V Shift: Why are AI servers upgrading from legacy 12V to 48V architectures? Demystifying PDN: What exactly is a Power Delivery Network (PDN)? Capacitor Roles: What roles do capacitors play in rectification, energy storage, and voltage step-down (buck) processes? Component Diversity: Why do different stages of the power network require different types of capacitors? The Capacitor Ecosystem: How do MLCCs, tantalum capacitors, aluminum capacitors, solid capacitors, and liquid electrolytic capacitors collaborate? Parasitic Elements: How do junction capacitance and Equivalent Series Resistance (ESR) impact high-performance computing? The episode also uses the vivid metaphor of "Great Yu Controlling the Waters" to help the audience grasp how electrical energy is managed, stored, and distributed within a server system. From the moment high-voltage AC enters the server rack to its final, stable delivery to GPUs and AI accelerators, every single step challenges the boundaries of power design and passive component technology. As AI computing power continues to soar, the importance of the power delivery system now rivals that of the chips themselves. Looking ahead, whether in AI data centers, EV fast-charging systems, or next-generation high-performance computing platforms, the Power Delivery Network (PDN) and high-end capacitors will undoubtedly become the ultimate battlegrounds for industry competition. 一、伺服器電源供應系統（AC-to-DC 與 48V 電源櫃）負責在電力進入伺服器機櫃的第一關，將資料中心的高壓交流電（AC）轉換為直流電（DC），並升級至 48V 架構以降低傳輸損耗。台股：台達電（2308）：全球 AI 伺服器電源龍頭，主導 48V 電源櫃與大功率 DC-DC 轉換模組。光寶科（2301）：高功率 AI 伺服器電源供應器（CRPS）核心供應商。群電（6412）：高階伺服器與資料中心電源供應器。康舒（2317）：雲端資料中心高階電源供應器。美股： Vicor Corporation（VICR）：48V 高密度電源分配架構（PDN）與模組的技術先驅。二、電源管理晶片與電壓調節（PMIC / VRM / IVR）負責在板端進行最後一哩路的「降壓與精準穩壓」，將 48V 降至 GPU 核心運作所需的 1V 以下超低電壓。美股： Monolithic Power Systems（MPWR）：AI 伺服器電壓調節模組（VRM）晶片核心供應商，與大廠如 NVIDIA 合作緊密。 Texas Instruments（TXN）：全球類比與電源管理晶片巨頭。 Analog Devices（ADI）：高階電源管理與精準訊號鏈晶片大廠。台股：矽力-KY（6415）*：高階電源管理晶片（PMIC）。力智（6719）：核心電壓調節與多相控制器（Multi-phase VRM）。三、高階穩壓電容群（MLCC / 鉭電 / 固態電容 / 鋁電）在 PDN 網路中扮演「大禹治水」般的緩衝與濾波角色。AI 伺服器的高速與高功率，對電容的「低等效串聯電阻（ESR）」與「耐高壓高容」提出了嚴苛要求。高壓/高容 MLCC（積層陶瓷電容）與鉭質電容：國巨（2327）：台股被動元件龍頭，併購美商 Kemet 後成為全球鉭質電容與高階 MLCC 的領先者，綜合戰力最強。華新科（2492）：主力鎖定高壓、高容、特殊規格 MLCC，AI 伺服器應用營收彈性大。信昌電（6173）：專攻 AI 伺服器電源模組所需的 100V 以上高壓 MLCC 與陶瓷粉末材料。禾伸堂（3026）：深耕利基型高壓 MLCC 與 AI 電源模組核心電容。板端高分子固態電容（GPU 與主機板週邊去耦）：鈺邦（6449）：全球導電高分子固態電容龍頭，在處理器板端高頻濾波、降低 ESR 扮演最直接的受惠角色。鋁質電解電容（前級大容量儲能與整流）：立隆電（2472）：資料中心用鋁質電解電容、高分子混合型電容供應商。凱美（5317）：鋁質電解電容與固態電容，受惠於電源模組升級。【發財二極體相關平台】 YouTube頻道：https://www.youtube.com/channel/UCVrP FB粉絲頁：https://www.facebook.com/ 【聯絡我們】 email: [email protected]

觀看 413按讚 5留言 0

觀看影片 ↗

資料擷取時間：2026-09-08T07:08:24.000000Z

公開欄位僅來自目前發布的 YouTube 頻道快照；缺值以「—」呈現。