【阿榕伯胡說科技Ep.87】獨家專訪前英特爾CEO季辛格，談矽谷硬體復興、矽光子與EMIB-T

音樂園本是預計在當地至少要花十年 人工智慧研究公司 偶給AI更多了 大喜舉演模型設計 從戶國神山到半導體供應鏈 從產業歷史到企業品牌密心 阿榮伯胡說科技 帶你聽熱鬧 更要聽門道 聽到朋友大家好 歡迎說你阿榮伯胡說科技 我是阿榮伯陳洋榮 我們這一集很特別 我們是專訪Intel前執行長Pat Jonesinger 那他現在是在一個戲鼓創頭 Playground擔任合獲人 然後他在10月下旬 帶了十幾家新創公司 帶台灣敗會供應鏈 然後在飯店的貴賓市裡面 讓我們錄影專訪 錄了大概一個小時 那這是我們 胡說科技第一次出外景 是很值得經驗的一集 那我們待會畫面會切過去 那我們先講一下背景 那我問的重點 其實是他怎麼看 AI這波帶的硬體復興 他們叫做Hawar Reynolds上司 那這個因為戲鼓的創頭 已經很長一段時間 他們不投硬體公司 其實也沒有 戲鼓也很久沒有硬體的新創公司了 可是現在因為AI的關係 所以投這些深科技新創 變成大熱門 聽說光是做AI晶片 就有100多家 那我覺得他講得蠻精采的 也願意評論現在Intel最好 想來台一店扣阿斯生意的Emit 提這個先進封裝技術 那我這部分已經先寫在本期的 胡說科技電子報 所以要訂閱的話 請看資訊欄或者是直接收取 胡說科技電子報 然後因為派著他講的是英文 那我也講的是英文 而且有些內容是比較技術的 裡面提到的新創公司 可能大家都沒有聽過 所以我待會會先補充一下背景 那我現在先簡單介紹一下 我們的講者既心革 他呢其實在龍村長大 因為家裡小孩是太多 沒有農地給他繼承 他才會去當工程師 那大家對於GyoSync的第一個印象 都是聽說在說2021年疫情期間 他就從本來是VNW的CEO 就回國老東家Intel要當舊世主 結果只做了3年半 他本來預計要做5年 就做了3年半 就在2024年底被董事會給請走了 那他在Intel的前半段人生 就是更早以前 其實真的很神奇 其實是一路跳擊 很誇張 他18歲就加入Intel 然後公司送他去Stanford做載子進修 然後他被Intel的傳奇人物 Andy Grove重點摘陪 等於是把當接班人量摘陪 32歲就成為Intel史上最年輕的副總 40歲就 進生首任的技術長 那個是2001年 Intel的全聖時期 那個時候Intel是市值是美國前幾名的 超級大公司 像現在會打一樣的地位 那現在AI硬體 有幾個很重要的技術 其實就是Intel在那個時候吹聲出來的 譬如說現在大學獲得ABF宅版 EUV還有細光紙 那可能很多人不知道 其實ABF宅版是1996年 Intel主動找日本未知事務 去共同開發出來的 ABF就是阿基摩托 Bewilder Film 阿基摩托就是未知事務 那個時候我問到一個比較難的 就是ITR-S 班奏級技術籃圖 可能大家比較陌生 我再解釋一下 那是Intel當年主導的一個技術組織 重點是要讓整個生態系統 都跟了Intel的步伐 其步向前走 從材料設備 所有報料理 相關的廠商都加入 會在那個 組織裡面討論每一代的摩爾定律 要做到什麼 然後 裡面用到新技術 哪一年就要做出來 要不然就會 敵類到大家 可是現在Intel已經不行了 也沒有ITR-S 就被人揮達主導 那就變成步調亂很多 然後台廠都在報竿都在報院 所以我就問他這個問題 他打的東西蠻有意思的 大家待會聽一下 然後裡面我提到他投資的新創 那在這裡解解一下 有一家是上月才成立的記憶體新創 叫做Nagari Labs 是 最近給他在Intel的老同事 什麽Data 創的 那個老同事好像很厲害 等一下聽他講 那公司的記憶體結構 就是很悶患 號稱說要像S-ray這麼快 然後能夠做到 像D-red那樣子的高密度 然後可以像Flash那樣的堆疊 根本就是超級悶患 就是要取代HBN 然後另一家公司 就是進行給他當任 執行董事長的X-Light 這家更猛 他是要用 例子加速去的技術 來產生新型的EUV光源 這個東西 大家已經講了幾十年了 從當年你本身在AVN的時候 就在討論的技術 可是都一直都沒有做出來 那他是第一個 要把這東西給乾脆學來的 他的Approach就是想要把 埃斯摩爾公司內部裡面 那個光源的公司 叫Simer給換掉 因為Simer現在生成EUV光源的方式 其實還沒有效率 也是EUV會那麼耗電的主要原因 好 那我們就轉換到訪問的現場 OK Nice to see you Very good Thank you I want to begin with a little story OK I have a very good friend he runs a VC like you and he is alumni of Stanford and he also worked in the semiconductor industry actually was the VP of UNC So I know him like 10 years ago and a lot of the partners of the VC were from semiconductor industry and most of them have PhD in Stanford but at that time all the company, their portfolios were lifestyle technologies travel and restaurant reservation 對 對 asked them why don't you invest in semiconductor company This background is that because in Taiwan we don't have new semiconductor startup in 10 years since 2000 something and they say in the US it was the STEM at that time 10 years ago people saying the semiconductor are boring even in Stanford but right now it's big we have all this interesting startups why is so big change? well fundamentally if we think over the last decade let's say who were the largest market cap companies a decade ago? they were largely software oriented Microsoft Apple we're sitting at the top of the heap and then what happened? AI happened and two things that result from that one is what has been the first industry that has been transformed by AI? and I will argue that the first industry is software we've seen this fundamental thing where you and I can vibe code a new application that replaces an existing SaaS application have a better line to what we need and we'll have that done this weekend it has transformed software so all of a sudden software was sitting at the top of the heap and all of a sudden it is the first one to be fundamentally transformed as a result so that has collapsed and who have become all of the top market cap companies? hardware and all of a sudden people realize that oh that software that AI runs on something and all of it is hardware so I'll say fundamentally Silicon Valley had become software Valley and all of a sudden all of a sudden all of a sudden all of a sudden all of the venture capitalists had gone over here I say how many more dating apps do we need to invest in? those days are done and now it's all back to the hardware and for playground we never left hard we were always doing hardware investing always doing deep technologies always doing deep tech always doing deep tech since the firm was started a little bit over 10 years ago I joined a year and a half ago largely for that reason and I have good friends that's interesting because I heard that playground was funded by the Walton family the family of Walmont no we have a range of venture capitalists that are part of it as I mentioned on stage Foxcon was one of our earliest venture partners we were always focused on hardware and that's part of the reason I joined a year and a half ago because I could have worked with a number of the different venture firms but all the rest had forgotten how to do hardware playground was one of the few that were still passionate about hardware and I think today's demonstration was clearly that way almost every company that you saw today was hardware and deep tech and the most interesting one of today's introduction I think is Nicari they just funded this year we just funded them literally a month ago a month ago and she was getting underway and its presentation said that he wants to find a foundry partner next year so fast very much, very important because we've been researching in university labs this technology has been worked on for seven or eight years in the university setting and we now believe that we have the recipe but we don't want to keep developing the recipe in a university now it's time to move to a 300 millimeter production fab build on a real-process technology not a university process and part of the visit here is to start to meet with potential fab partners so that's part of why I'm here and we're starting to meet 最重要的是,在最重要的部分,是最重要的部分。 that we could work with, so that we can start to build a 300 millimeter volume technology, so that we're not developing the technology and then transferring it. No, we're building it on a fan process that can scale to high production. So, although it is a new company, we're already in a commercializing phase. Oh, too. When you're looking at semiconductors, what we would call as the technology transfer phase, there's still lots of innovation. Okay, let's put it on the university too. Yeah, crazy. You know, you know, for that, and there's a lot of work to be done yet, because, you know, to do a memory, right, you have to show the memory works. Yeah. You know, but then you have to show that the memory, right, you know, is reliable. Yeah. Right. You know, so the whole, you know, set of, you know, things to be able to show that it's reliable, then you have to show it's manufacturable. When you fetch it, it's very potent. You know, as well. And until you do all three, you don't have a new memory. Yeah. Right. So, you know, when obviously it's proven that we can make it fast, you know, it's proven that we can make it dense. Yeah. Now, can we make it reliable, and manufacturable? We will do that with a fan partner. Yeah. When I thought introduction of the company, he says, the speed of S-RAM, with the density of D-RAM. Yes. Wow. That's huge. Oh, absolutely. Absolutely. It's a service killer. Yeah. And if you look at the graph that they showed, don't we believe that over the life of the technology, and he showed from Z1 to Z7, you know, that it will become denser than D-RAM? Yeah. Because. It's an D-RAM. Yeah. Because we can make it stackable. Right. Yeah. Like flash. And that's why I say, you know, it's fast like S-RAM. It's dense like D-RAM and stackable like flash. Wow. Because of those three characteristics. Right. We think that it can continue to be a technology that expands its uses over time. You know, for it. So we're quite excited about them. Yeah. You know, that's too good to be true. It is? You know, in past 20 years, we heard a lot of new memory structure. All of them were too good to be true. And eventually, they didn't come up in mass production. So why do you think this one, my maker? Because you're in Intel. Intel has a lot of new memory. O-Fail. Yeah. And I think two things now. Okay. You know, one is, you know, and as I commented, this is the team. The team. The team. You know, and Suman, you know, we've worked together for many years at Intel. Okay. And, you know, he was the guy that drove high-k metal gate. Right. He was 28 nanometer. Yeah. Yeah. He was the guy that did that. You know, and he knows that, you know, innovation is easy, right? A manufacturable, reliable innovation is hard. So he knows the difficulty of manufacturing. Yes. And back to design structures. Because he took high-k metal gate. Yeah. It's very, very difficult. Very, very, very production. Because other people had similar ideas, you know, for, you know, being able to essentially, you know, create, you know, the work functions and use different materials. The ideas were out there. But nobody could get them to work in a manufactured, reliable way. He did it. So he knows what that is. So, and they've been working on this for the last seven or eight years, right? This core concepts of oxide, transistors or oxide gates for storage. So they've been working on it. They've been maturing the ideas. And this team knows what it means to bring that kind of innovation into volume production. So that's one reason. But the second reason is is exactly what's going on today in memory. You know, there is an extraordinary demand for performance memory. Yeah. So you have, you know, an innovation with a team that I believe can take it to manufacturing. And you have a market demand. Yeah. AI is a memory workload that once banned with one density. And with that, you have a market need. And you have what I believe will be non-technology. So the timing is different. Even the technology came out maybe 20 years ago. Maybe you will stay in, well, 20 years ago, SRAM was still scaling. Yeah. Right. SRAM is no longer scaling. You know, literally if you look at N2 or Intel 18A, the SRAM on 18A or N2 are actually larger than the SRAM on the prior nodes. Right. You have no longer the transistors are getting smaller, but the SRAM is not. Okay. And so SRAM is no longer scaling. Right. And you know, what's HBM? Right. HBM is a D-RAM stack, but D-RAM doesn't like to be hot. Yeah. Right. So now you have essentially a thermal stack. Right. You're leveraging through silicon vías and others to get more bandwidth. You're shoreline bandwidth limited. You really want the memory on top of the CPU, XPU, or GPU to get exponential bandwidth. However, it's can't handle the heat to go do that. You create, you know, a thermal stack up. So, you know, you're really sort of stuck in the physics of a D-RAM. So SRAM is not scaling. You know, D-RAM can't enable the exponential bandwidth that you want. So we need something new. We need something new. And this is the answer. Yeah. And this is CMOS compatible. Right. So essentially, it can layer on a CMOS process, you know, since it's not using a capacitive storage. It doesn't have the reliability issues of a D-RAM at higher temperatures. And all those characteristics say, oh, this is potentially the breakthrough. Okay. And next question. I want to talk more about the hardware in Reynolds sense that you have previously said that right now is the best time to make hardware. Yeah. And I want to compare with the time when you were in Intel. You know, today, all the major technology of today's AI hardware, like a ABF substrate, and like a UV. The first half of the UV is invest by Intel and a second photonics. A lot of the major technology today was pioneered by Intel 20 or 30 years ago. And we're using it right now to make AI. And you're maybe probably CTO at that time. Yeah. And you know, that was the good old days. When I first time becoming a semiconductor journalist, I went to the ITRS meeting for the technology roadmap. And it was led by Intel, right? And so it's easy for us to predict the technology breakthrough, because in the ITRS, they tell me what years, which technology, which material should be ready for the most long? So it's if the 28 nanometers, or maybe 2010, high-key metal gate, maybe some equipment should be ready at what years. So everybody knows what to do. They have the roadmap, and everybody knows their role in this big roadmap. But surprise in Taiwan told me that right now, it's a chaos. It's a nightmare. Like a meeting 72, a media. Everybody works so hard and delay. And everybody worked overnight for the project. You know, they told me that right now is the situation to catch up the technology. It's more, much more difficult than 20 years ago when you were CTO. So why so big change? Can you compare the situation right now with the in-heel era? What's the difference? I do think there was Intel, in that period, had sort of, I'll say, an ecosystem perspective, right? You remember things like the Intel developer forum, ITRS that you spoke to, there was this, we had the EUV consortium. And I do think there was the central role that Intel played to sort of help guide in a quite open-handed way these many areas of the industry. You know, today you don't sort of quite have that leadership role. You know, when you see it in different pockets, you see, for instance, what TSMC does, right, with its, you know, their developer days and events, UCI events that they do. You see, for instance, iMac and their roadmaps that they present. You know, nobody sort of looks at the IPRS roadmaps anymore and say, that's the, you know, that's the future. But, you know, so I say, you know, that singular role has become more fragmented. Yeah, TSMC has its custom and media, the different little map. Yeah. And you have seen AI explode rapidly. Okay. Because, you know, if we were sitting here, right, you know, let's just take one parameter, right? If we were sitting here four years ago, five years ago, you know, and the A100 was just about to come out, right? Yeah. Right? Intel's, you know, at that point, it'd be a Zeon 5, right? What was the thermal envelope of the socket? It was about 300 watts, right? And, you know, the industry was grappling with going from 250 watts, right? And I think both the CPU and the GPU were going from 250 to 300 watts. That was a hard problem going from 250 to 300, right? Today, we're going from 300 to 5000, right? You know, in five years, right? We've seen an order of magnitude increase where before, we were sort of doing 10 percent, you know, per, you know, generation. We've seen that copper delivery. right, is fully inadequate because we've seen this explosive requirement of AI. So we've seen both the fragmentation, right, of there's no singular point that you can look at, you know, like an Intel or ITRS has become fragmented, and you've seen the explosion in the market. You know, what do I do, right, and then, you know, if anything, the Nvidia roadmap is the closest to the industry roadmap, you know, but even there, it's only pieces of it, right, and they're not trying to be holistic in necessarily facilitating the industry. So I think fragmentation and this dramatic acceleration that we've seen with AI. So what should we do as a technology driver? You were real right now, you know, VC, to drive kind of new technology companies to foster the ecosystem to go faster. Yeah, and so, you know, for, you know, we're trying to look across, right, the entire computing stack, right, as we would say, as we would think about, you know, what's happening as a semiconductor layer. Well, we're trying to create snowcap. Yeah, you know, you know, entirely new physics, right, we're trying to recreate power delivery with vertical semiconductor and power lattice. You know, we're trying to then, you know, recreate interconnect, right, move from copper to optical. I predicted the death of, yeah, yeah, yeah, 20 years ago. You know, so, you know, I think we are trying to build them. We're trying to say for semiconductors, you know, 13.5 nanoveter UV is a good enough, right? We have to have the next major breakthrough, the next major, you know, wavelength of light. So we're working on X light, you know, so every one of these, we're trying to say, okay, can we create the seminal company that defines that layer of the stack that enables the industry to catch up to what AI is requiring. Yeah, since you mentioned, I want to say this, you say that 20 years ago, today's update is a niche technology. Yeah. And tomorrow is the mainstream of every chip that we built. Yeah, right. That's 20 years ago. Yep. Yep. Eventually, I'll be right. Yeah. But why, why you say that at that time? Well, the same deal of course, copyright is still okay at that time, right? But we already started to see some limitations of power, right? The ability, and you know, there's been huge advancements. But relatively speaking, you know, the improvements in I/O, right? If you look at the improvements in transistors, you know, performance doubling every two years, you saw I/O doubling every five years, right? So even with the improvements in copper, the compute versus I/O ratio was increasing, right? You know, over that period of time, right? And so, you know, that fundamental separation is what was driving that comment, right? We need I/O to scale like compute does. And obviously, it didn't, right? And frankly, you know, that is what has led to many of these AI bottlenecks, right? You know, the NVL72 Taiwan was in chaos, right? Because as I call it, you know, the NVL72 was an engineering masterpiece and a manufacturing nightmare, right? It wasn't manufacturable, it wasn't scalable, but yet the demand for it was high because people wanted larger cluster sizes, you know, with an NV link connection across its memory portfolio. There was demand from the workload, you know, for it. But we have to fix the underlying technologies, we have to fix the physics of the platform. And that's what we're working on to playground. And I think that's part of the unique role that we play with Taiwan, right? You know, that we are looking for that next generation stack. Yeah, I want to go back to the simultaneous photonics because I have a friend who worked at that department in Intel 20 years ago. Yeah, and he told me what's a friend. Yeah, I had a lot of friends and I'm working in this field. Okay, because my friends in your department, he was right in the department after he graduated from Stanford PhD. And he told me that everybody's thinking at that time, the optics will commercialize in three years, but always three years. Yeah, because the copper is improved. Yep, you know, and you've seen better signaling for copper, you know, people have gone, you know, from Pam 2 to Pam 4, Pam 8. You know, they've been able to. Yeah, you know, and all of these things, they pushed it far further than I expected. Yeah. And, you know, frankly, I think that most people, you know, would have envisioned the incrementalism enabled them. But, you know, fundamentally, you still had I/O scaling much slower than compute. So that Taiwan industry care about this very much when will CPL finally happen? 2028. 2028? Yep. You know, the 2028, 2029 generation of AI systems, you know, and why will it happen then? Yeah, yeah. Right. You know, and there's two things, the maturing of the technologies, because, you know, people are working on it. People are working on it. Iron Labs is an 11 year old company. Yeah. You know, working on it, working on, you know, supply chains, TSMC Coop. Right. You know, and every piece of this is getting more mature. Right. You know, for it. And the experience of NVL72, everybody realizes we can't do NVL 288. Oh, really? Right. You know, copper doesn't work. Right. Because I want to go faster, which means I got to go smaller. Right. But, you know, to, you know, I have to do four racks now, not one. Right. You know, right. The physics is breaking. So the physics is breaking. So the need in the marketplace is much greater. And the model sizes have only gotten bigger. You know, we're talking about two terabyte models. Right. You know, so people, you know, so the models have gotten bigger. The physics is demanding that you make this and the industry is matured. So as a result, 2829 is the NPO CPO conversion point in the industry, my belief. Because what will happen is, you know, the early movers to CPO and NPO will be able to do a 288, maybe a 576 or maybe even a thousand plus cluster size. So the models, right, you know, that need large, you know, working memory, you know, that one large KV cache, that one large context and, you know, they will work on those and the copper based ones. Yeah. Maybe they can only build 36. You know, and, you know, so you're going to be looking at, you know, right, the equivalent 36 node, right, scale up system versus the optics based thousand node system. I want that one. Right. And I think, you know, the 2829 time frame is when the transition will occur. Now, that doesn't mean copper is going to be zero. Yeah. Right. But, you know, that's when we see this fundamental transition, you know, in next year, 2027 is important because that's when the supply chain has to start happening. You know, all of those things have to start maturing, you know, first, you know, system prototypes, all of the pieces of the supply chain, we need to route a route to supply chain next year so that we can start the, you know, ramp in a second half, 28. Yeah. But another friend of mine in TSMC, who is in charge, some of the pieces of lots of friends, you know, lots of friends. He told me that he's so worried about this supply chain issue because CPO is so new, a lot of interface and every piece of them will, this will not be easy. Okay. But in comparison to thinking that you're going to do the NVL 588. So if anybody says, hey, it's not going to happen. I just say, you know, remember NVL 72 and how hard it was. Yeah. You can show to happen. Yeah. It will happen. Right. And even with the supply chain challenges, you know, that we will have for coop and CPO and so on, they will be far better because they're moving in the direction of physics. When you are fighting physics, you're fighting a losing battle. Right. When I get to optics, I now have physics on my side again. Okay. I'm not fighting physics. That's right. That's right. So 20 years, finally come true. Finally true. And you have to say that every technology stack of AI needs to improve. Yes. Yes. So you have mentioned some of them. Yes. We have mentioned memory, sitting on photonics, optics, and which part do you think is the other bottleneck we need to fix? Well, the power power. Yeah. Power delivery and, you know, we, you know, presented power lattice. Yeah. Because the numbers are like IVR. Yeah. Yeah. And we did IVR for my group at Intel. You know, we did the first eye hardships. Yeah, it was, you know, we call it five or at the time. Okay. Right. You know, it was the code name and, you know, some of the power lattice team worked for me at Intel. Right. So everyone in the press conference work with you. Well, you know, you have many friends. Everybody worked for me. Right. So, yeah. So, and I joke that if you work for me once, you work for me forever. Right. So, right. You know, those relationships are powerful. But, you know, the IVR technology, it's time has also come because if you look at, you know, a GB 200, you know, border, all of the stuff that's not silicon is power delivery. Yeah. Right. You know, you have inductors, capacitors, switching elements and so on like that, all of that collapses into a chiplet, you know, that's immediately under it. Again, iSquare R losses, you know, the further you are away, the more resistance, the more iSquare R losses you have. So folding that into essentially a vertical power delivery is quite extraordinary. And, you know, if you think, you know, if I need to get, you know, if the GPU needs 3,000 watts and I now need to give it 8,000 watts to just get 3,000 watts to the GPU, where the 5,000 watts go. Essentially, it's lost in power delivery. So over half of the power of a data center is lost in power delivery. You know, that's just not sustainable at that level. So, you know, being able to solve power delivery, we also see that power conversion, you know, we believe data centers will become, and, you know, Jensen and I agree on this, that the new data center will be 800 volts, you know, a top of rack, right? And that'll be the power delivery. But then I want to have single-stage conversion, right? I want to be able to go from 800 to 48, 800 to 4.8, right, directly into my voltage regulation, and then you able to move data center efficiency from the 80s into the mid-90s. And so that will again free up more power for the data center. So power becomes another major bottleneck. And again, we're quite excited to have a couple of our companies in that space as well. Interesting. I have thought the Rome atom, TSMC, they have the next generation of their COAS. IVR is integrated, maybe in the poster, there's a component over there. And people told me that TSMC, maybe they want to make it in-house. But we think that power a lot, you know, again, power is hard, right? I joke, you know, that your power delivery, you know, it is more important to you than religion. You know, you don't change religions very often, right? And changing power delivery, it's hard, it's analog, you know, very few people understand it and can do it well. You know, we think the power lattice team is one of the unique ones in the industry. And nobody else has solved the inductor problem. You know, they have solved the inductor problem, you know, they've put it into a chiplet that's very small, so it can immediately fit into the substrate. You know, it's less than the BGA height, right? You know, so essentially, you know, in the BGA stack-up space, we can put the power delivery chip, you know, that's stunning, right? You know, so the physical form factor, the inclusion of inductors, the high performance switching, you know, and the demonstration they're doing, you know, it is a thousand times, you know, essentially the voltage drop, right? You know, that they're demonstrating today is a thousand times better than today's solutions today, not tenet. You know, so essentially, it takes you a millisecond to respond, we respond in a microsecond, right? With one-tenth the amount of voltage drop, you know, on major current swings, you know, that's what we're now demonstrating. So this is pretty transformational. So if your power delivery network needs an extra 5,000 watts, the drop is 10x higher, and it responds a thousand x slower, right? Okay, this is game-changing for power delivery. Yeah. What I see from your portfolio today, something surprisingly that they were all ready to come to fight with the giants, like the other companies, so I forgot them, they say the solution will compete with every link 72 in the scale-up territory. I say, you're competing with ambiding, you say, no, we're collaborating, but it's actually competing. Well, yeah, we're creating the breakthrough technology. Yeah, very breath. Yeah. And, you know, innovating, you know, fundamental ten hundred thousand x breakthroughs. Yeah. Hey, some of them might get acquired by Nvidia, right? Or TSMC, you know, if they're that good, and they're that breakthrough, you know, maybe they compete, maybe they get acquired, you know, we are not focused on that. We're focused on creating innovative companies, right? That fundamentally solve the bottlenecks of the entire AI computing staff. So bring me to another question. Have you found any that kind of company in Taiwan? So far, we haven't announced any yet. Okay. And we haven't found one that we're public on yet, but we've done our first Japanese investment, recently, we've done our first Korean investment recently. Scienta is an Australian company. So it's very impressive. Yeah. So I'll just say we want to find those unique breakthrough companies anywhere in the world. Okay. And, you know, we're we're actively looking to find some here in Taiwan. Please find one. Yeah, nothing would make me happier than when I'm here next year, you know, to have a Taiwan company. That's on stage. Okay. Now, you know, I sort of think though of Pikachu almost as a Taiwan company. Because yeah, Taiwanese. Yeah. I don't even tell you. And you heard in several of the companies, you know, we're opening Taiwan offices, yeah. For a number of the, yeah, your labs has power lattices now open to Taiwan office as well. Scienta will have a Taiwan office. So it's going to be hard to tell where Taiwan ends in U.S. or other countries, you know, begin. So we see that, you know, we're just going to be embedded to the ecosystem of Taiwan because this is where the world innovates. Yeah. I want to talk about X-Light because I've been to SML in Littleland and talked to CEO CTO before. So, you know, especially when there's a trade war with China, everybody's thinking, UV is the final frontier. If you want to, you know, blockade China. So X-Light is very a very brave company. Yeah. The most brave of all, I think. And you're the chairman executive chairman in the company. So I want to know what do you think of that company? Their goal is very ambitious. Yeah. To use the accelerator, the light source from the accelerator to replace the UV light source. And because everybody wants to use the accelerator maybe 20, 30 years ago. It's a very ancient idea. But right now they want to commercialize. I talk to the people today over there. And I think the way they speak to me is things to be practical. But, you know, a lot of things practical in the paper, but to work in a fact is another thing. And you have very good experience in commercialize everything. So, but do you think why do you think they can make it since you are chairman? So, you must have seen something. So, you know, the first is that the idea of a free electron laser. Yeah. Right. In fact, if you, you know, study ASML, had a free electron laser project. Okay. Right. So, the idea was considered, you know, by them like a decade ago. Okay. Right. You know, it's potentially a new light source. And when they looked at it from the optics, they couldn't figure out how to make it work. But what X-Lite has done is they've looked at it from the laser, not from the optics. So, they've actually created the multi beams at different power lengths, different wavelengths, different multiplications in the free electron laser. You know, so essentially, you know, they're accelerator guys. You know, they're not optics guys. And the problem before was that you had this whole 100k watt light source hitting the first optic. It just melted it. Yeah. Right. And what they have done is they've said, Oh, I'm going to create, you know, the multi 5k watt light sources in the laser. So, I'm going to be able to essentially, you know, create that separation. So, it's a total of 120k watts, right, that they're producing, but they're producing it as 20 separate light beams, right, out of the laser. And that was the fundamental breakthrough that allows them to say, Oh, I can connect up 20 scanners to one free electron laser because I've already created the partitioning of light. And when you've done that, it's like magic. All right. Because now nobody has to bear that full burden of the, you know, the full power of the light source. So, you're not melting the first optic. Right. You know, you also then have per beam flexibility. Oh, yeah. You know, that beam can be 5k watts. That beam can be 1k watt. That beam can be polarized. That beam can be 13.5 nanometers. That beam can be 12.8 nanometers. Or that beam could be 4.5 nanometers. So, now you can actually open up new wavelengths of light. Yeah. Because if you go back, you know, 193 nanometer light, you know, you then made this bold dramatic step to 13.5 nanometers. Right. You know, it was, that's what I'm using right now. Yeah. And now nothing has changed. We've been at, you know, 13.5. Yeah. Right. And to say what's next, we believe that we now open the opportunity for what's next. Because we believe that, you know, hey, we could take it to, right, to seven nanometer light. We could take it to 4.5 nanometer light. We expect that 4.5 nanometers is about right. Okay. One quarter of the wavelength of where it is. We believe their material structures, you know, for lenses and mirrors, et cetera. That will work at 4.5, resist at 4.5. So not only do we make today's EUV better, but we open the door to tomorrow's EUV. So yes, this is bold. And weftlands has changed, so the whole ecosystem, the photo resist, can they catch up? Well, you know, the photo resist, you know, everybody's a 13.5, because that's the only thing. But you know, part of, you know, I visited, we've been with the X-Lite team, particularly a lot of the resist work happens in Japan. Okay. So we're engaging with the Japanese companies around different resist, starting to explore what different mirrors, lenses, optics might look like as well. So just getting started, and I do believe somewhat like the EUV consortium was, you know, it took a decade of formation of that, you know, we do believe that we'll have to have that same sort of formation of what's next. Also, you experienced in Intel. Yes. To build a new ecosystem. Yes, yes. And you know, the Japanese, the Europeans, you know, TSMC, all of these will be important elements of building that, what's next. Yeah. And I want to ask the question, am I Intel? I want to start with the IMFP. Yes. It may be going to be the first major victory going to be against TSMC in maybe ten years, because Google is using it. And it might be success because they're debating about the maturity of the technology. And I think it's on you, it's part of the idea of 2.0 project, and the new Mexico factory was on you turn, right? Yes. Yeah. So can you tell me about this whole thing, IMFT? Because in the first, people don't believe, it's better than COAS. And right now, there's a rumor that TSMC is exploring their version of IMFT. Yeah. And you know, when you think about, you know, Intel had Foverose, TSMC had COAS, right? You know, they're similar, they're very similar technologies, and you know, but the problem with both of them is is that the substrate, you know, essentially needs to be the unit area of all the silicon that you're putting on top of it. So now you're taking a 300 millimeter wafer, you know, and creating maybe six COAS substrates per wafer, right? Yeah. Right. You know, when you're putting a rectangle on a circle, right, you can't get many, you know, in that sense, from a manufacturing perspective, it's not very scalable. Yeah. Waste a lot of wafer. Right. You know, when you go to IMFT, right, you're essentially creating an active die for the active area, right? Yeah. And then embedding those into the substrate. Yeah. So much more, you know, dramatically more silicon efficient, right? And since you're taking this now, right, much smaller area, you can put active devices into it as well. Yeah. Right. You know, because now, you know, if I take a very large area, it gets very expensive to start building, you know, connectivity into the substrate. Oh, no, but it was failed to in a substrate, not in the pose. Yeah. Yeah. So, you know, all of this makes it an attractive technology at that. And obviously, we were working on Foveros, right, which is very co-awesome, like, you know, but this was one where it's like, okay, you know, the silicon speaks, right? It's on our side. And if we can make it work, and obviously Intel, you know, was making it work in volume. So then to go to customers through the foundry. And given that, you know, essentially, you know, you're wafer limited, and that means that you have capacity limitations, as well as cost limitations. So when you're winning on capacity and on cost, right, all of a sudden, that becomes an exciting technology. So, you know, it's come to bear, and yeah, I was very happy to see that occur. You know, I was just in New Mexico a couple of weeks ago for a quantum event that was going on there. And I took great pride in seeing the emin to still be coming to life there. Yeah, I heard that news and I did creating jobs in New Mexico, looking very good. And I do think, you know, getting Intel to become a foundry for the world. You know, the world wants more foundry capacity. The world wants more foundry innovation, right? And, you know, 18A was hard, because essentially, we were catching up from Intel having fallen behind on the transistor, going to RibbonFat, they'll fall in behind on EUV, right? So we had to catch up there. And we also leaped ahead with backside power, right? So, you know, the three things, and 18A was hard, but 14A looks superb. Really? Yeah. Looks superb, because every time, if you look at Intel's process technologies, 45 nanometers was hard. You know, when it went to 32, it was a beautiful thing, right? You know, just, you know, the first one is always hard, right? You know, major new innovation in the transistor and lithography, you know, material structure, et cetera. The second one is always, okay, you're just refining all the learnings. So you mean that 18A, go first, and the 14A is, seeming to technology. Yeah. Similar technology, you know, going from four ribbons to three minor refinements on the backside power delivery, and it is just cleaning up 18A, you know, it's going to be a beautiful process technology. So I think, you know, EMIB-T14A, okay, Intel's back to having road class technologies. Five nodes in four years. That's the plan. Yeah. Yeah. So it's a pity that you don't have extra ear to see that. Yeah. Well, but the strategy is continued on, right? Yeah. Right? You know, and I'm proud that, you know, Intel was, you know, on the verge of leaving, you know, manufacturing, when I came, now it's back in manufacturing, and they're building more capacity, and obviously the commitment to build capacity was the right one. Yeah. Right. Because, you know, there's not enough to today's point of view. Well, you know, and they paused when I left, right on the factory buildout, and I think everybody today says, they should have the paused. You guys should have paused. They don't have money at that time. You don't have money at that time. You know, I have a lunch with few former executives at the TSMC. And one of them were 20 years in Intel, and then of course, and I told them that I'm going to interview Pat two weeks later, and what would you suggest to ask him? And actually they say that it's a pity that in your turn, agentic AI didn't come out at all. It's because in the CPU price, it's exploded, and save the company, save the cash flow. Yeah, it's just a one year. Well, you know, when we started the five years and four, you know, five notes and four years, and when I came back, you know, I said it was a five year assignment, right? And, you know, the board was comfortable with less than four. And, you know, I was disappointed, it wasn't, you know, how I laid out the plan. Yeah. At the same time, I'm in a good place, you know, I'm happy, I'm driving technologies that I think are super important, and the strategy that I laid out for Intel is continuing, right? And I think that's valuable for the company, valuable for the industry, and I'm doing other big things that I think are super important to the next generation of technology. And, you know, that's what we had today with PlayGrop, you know, many of those next generation technologies, and, you know, I think you're starting to see some of those come to life. Yeah. So I'm fulfilled, I'm challenged, and, you know, I do think the continuation of the Intel work combined with building the next, I'm in a good place. And do you think the factory business in Intel right now, because Libu, say that if there's no new customer, maybe it's a good technology, but no customer is using, maybe you will stop. So what would you look at the factory business in Intel and sustainability? Well, you know, obviously, I'm not in that chair, you can outside point of view is, you know, hey, to build major factories, you have to be a factory today. So if you build it, they will come, well, Intel needs to build it for its own products. Yeah. Right? And if you're building it for your own products, fill it up with the industry's products. It's going to be a great process technology, and I certainly hope that they, you know, don't pause, don't hesitate, you know, keep going. Is that good? You can act proper. You need to invest in advance to persuade customer to come out. Nobody knows that better than TSMC. Yeah. Right? Yeah. You know, for, and it really is spectacular, you know, what TSMC has done and become. And that, and that respect, and certainly, you know, my wish and hope is that Intel does that as well, but in the meantime, I'm driving innovations at places like Playground. I ask one question on the behalf of my TSMC friends, because that you have stayed at it, because of this AI with too huge and the progress of technology too fast. So maybe our calculation five years ago is no longer working right now. When people are complaining that TSMC is expanding too slow, you know, even CC way, say that the demand is three times their capacity. So in that kind of demand, TSMC is like 100% monopoly in advance note. And do you think it will be the need of the second and the third support? apply at the best node, with good mucky share in the future with this AI training. Well, you know, first, you know, we see the AI demand as relatively insatiable. Right? You know, there, you know, how much AI do you want? You know, like I said on stage, 10,000 X. Yeah. Right. You know, it's just, you know, the desire, you know, I want my glasses to have AI. I want my hearing aid to have AI. You know, I want every business process to be using AI. I want every manufacturer step and, you know, Taiwan manufacturers. You know, I want every step to capture images that are being AI analyzed at every step for every part, you know, for every manufacturer in line. Right? You know, forever. Right? You know, just dramatically more. Right? And being able to, you know, do metrology, you know, do, you know, forensics, et cetera, you know, across it. So, I don't want a little bit more. I want a lot more AI. So, you know, in that sense, absolutely, you know, we should be building more directly. You know, I firmly believe the world needs more than one supplier, you know, for everything. Right? You know, Amcore needs an ASE, right? You know, right? And, you know, hoping to see more, you know, competition there. But we need more, more than that. We need fundamental innovation. Okay. Right? You know, and why is he mid-T interesting? Right? You know, because it's a fundamentally more efficient way. You know, why is science interesting? It's a fundamentally better way, you know, to think about packaging. You know, and maybe our most radical company we haven't even touched on yet is snowcap. Yeah. Right? You know, saying, you know, superconducting Josephson Junctions. Yeah. And it runs perfectly well on a fab process that sort of is like, you know, think about it like a 15 nanometer process. Right? Is what, you know, it's sort of like well suited to at that. I don't even need EUV for that. Right? You know, for it. And, you know, it's maybe the most radical one. I need new physics. Yeah. You know, copper to optical is going from electrons to photons. It's new physics. And, you know, so we continue to see that innovation is what we need to see in addition to capacity. So, you know, we believe, you know, those innovative cycles clearly, how we do better job with today's AI and CMOS products, but how we usher in, you know, the future devices that may or may not be based on CMOS. Yeah, but that's interesting because when I, once I saw the podcast of OpenAI, they say that in the beginning, they start superintenergize. It's just a new computer program. You write a new model. And later, they find out, okay, there's a scaling law in the more GPU to make the model better. Yeah, more data, more training data, more devices getting larger. Yeah, but maybe the scaling law, maybe we'll slow up. Since you use a new architecture, maybe new model, we'll use less computing power to do a better job. If that entrepreneurship came out, we don't need so many fans. I fundamentally disagree, because you remember, right, if new models, let's say they're 10x better, yeah, I still want 10,000x more AI. Right, it's not like a little bit, not enough. Right, you know, I mean, there's so much more, you know, capacity, you know, that we want, you know, that these breakthroughs, I feel again, is reasonably insatiable. Right, in that respect. So, you know, clearly, you know, we don't see any upper bound to silicon, because we're going to just keep seeing it in more and more places. You know, what we put in silicon may change. Right, you know, there's going to be new algorithms that emerge. There's going to be new ways to put systems together. You know, there's going to be new physics, like electrons to photons for optics. But the upper bound here, you know, is everybody will want more intelligence in everything for every person on the planet. Yeah, that's the path that we're on. And right now, you're fundamentally constrained by manufacturing capacity here in Taiwan. You need more. Right, so yes, you should be investing more. Thank you. Thank you for your time today. Okay.