# 【直播｜AI即時中字翻譯】輝達 NVIDIA 黃仁勳攜手Marvell執行長在COMPUTEX 2026主題演講談AI新戰略

https://www.youtube.com/watch?v=dPHKRmMSsNE
Translation: zh-TW

[00:00] Heat.
  熱。

[00:00] Heat.
  熱。

[01:39] N.
  嗯。

[01:49] Heat. Heat.
  熱。熱。

[02:13] N.
  嗯。

[02:23] Yeah,
  對，

[02:44] I know.
  我知道。

[03:38] Heat up.
  加熱。

[04:04] Heat.
  熱。

[04:08] Hey, heat.
  嘿，熱。

[04:08] Hey, heat.
  嘿，熱。

[04:32] Ladies and gentlemen,
  女士們先生們，

[04:56] Heat.
  熱。

[04:56] Heat.
  熱。

[04:56] N.
  N。

[05:12] Heat.
  熱。

[05:18] Hey, Heat.
  嘿，熱。

[05:43] Heat.
  熱。

[05:49] Heat.
  熱。

[05:49] Heat.
  熱。

[10:41] Ladies and gentlemen,
  女士們先生們，

[10:44] Please continue.
  請繼續。

[10:49] Thank you.
  謝謝您。

[11:04] Heat.
  熱。

[11:48] Heat.
  熱。

[12:10] Heat.
  熱。

[13:32] Heat.
  熱。

[15:43] Heat up here.
  這裡很熱。

[16:10] Please welcome Marll chairman and CEO Matt Murphy.
  歡迎 Marll 主席兼執行長 Matt Murphy。

[16:33] It's great to be here to kick off day one at Computex and it's great to be back here in Taiwan.
  很高興來到這裡為Computex的第一天揭開序幕，也很高興回到台灣。

[16:38] You know, the first time I came here was nearly 30 years ago.
  你知道，我第一次來這裡將近是30年前。

[16:40] It was my first business trip to
  那是我第一次的商務旅行到

[16:45] Asia and I remember back then visiting some of the key technology companies here at the time.
  亞洲，我還記得當時參觀了當時這裡一些主要的科技公司。

[16:52] Many of them were still young small companies, emerging companies and today those same companies have become the most important technology leaders in the world.
  當時許多公司都還年輕，是新興的小公司，而如今這些公司已成為世界上最重要的科技領導者。

[17:02] Now I've had the opportunity to come back many times and see Taiwan continue to grow in importance as one of the world's leading technology centers.
  現在我有機會多次回來，看到台灣作為世界領先的科技中心之一，持續成長並日益重要。

[17:12] And today so much of the future of AI infrastructure is being built right here.
  而如今，AI基礎設施的未來有很大一部分正在這裡建造。

[17:19] I have a question for all of you.
  我有一個問題想問大家。

[17:22] What defines the performance of AI infrastructure?
  是什麼定義了AI基礎設施的性能？

[17:25] Now maybe you're thinking about the processor, the GPU, the XPU, or maybe it's the process node used to build it.
  現在，你或許在想處理器、GPU、XPU，或者也許是建造它所使用的製程節點。

[17:35] 3 nanometer, 2 nanometer, or soon A14, A16.
  3奈米、2奈米，或者很快的A14、A16。

[17:39] Those are great metrics.
  這些都是很棒的指標。

[17:41] They tell you a lot about the speed, the efficiency, and the density of the compute.
  它們能告訴你很多關於運算的が速度、效率和密度。

[17:43] And AI
  而AI

[17:46] workloads are certainly compute intensive, but that's not the whole story.
  工作負載當然是計算密集型的，但這並不是全部的故事。

[17:50] Now, you might say, well, what about memory?
  現在，你可能會說，那記憶體呢？

[17:53] AI workloads are incredibly memory intensive as well.
  人工智能工作負載同樣是極其記憶體密集型的。

[17:56] More memory, higher bandwidth, all of that matters.
  更多的記憶體，更高的頻寬，這一切都很重要。

[17:59] It's all critical, no doubt.
  毫無疑問，這一切都至關重要。

[18:03] But that's still not the defining characteristic of the system.
  但這仍然不是該系統的決定性特徵。

[18:08] Because one processor, no matter how fast it is, no matter how much memory it has attached to it, is simply not enough for today's AI workloads.
  因為一個處理器，無論它多快，無論它有多少記憶體連接到它，都遠遠不足以應對當今的人工智能工作負載。

[18:15] You need tens of thousands and eventually millions of processors working together as a single massive compute engine.
  你需要數萬甚至最終數百萬個處理器協同工作，作為一個龐大的計算引擎。

[18:23] That's why computing at this scale is fundamentally a connectivity challenge.
  這就是為什麼在此規模下進行計算本質上是一個連接性挑戰。

[18:25] And increasingly it is the architecture and characteristics of connectivity that defines the performance of the system.
  而且越來越多地，是連接性的架構和特徵決定了系統的性能。

[18:31] Now look, we've seen incredible breakthroughs in accelerated computing and we've seen the emergence of high bandwidth memory to meet the AI challenge.
  現在看看，我們在加速計算方面看到了令人難以置信的突破，並且我們看到了高頻寬記憶體的出現，以應對人工智能的挑戰。

[18:46] But I'm here to tell you the
  但我要告訴你的是

[18:49] Next major wave of innovation and scale will come from the underlying connectivity of these systems.
  下一波主要的創新和規模將來自這些系統的底層連接性。

[18:57] And as those and as those connections move from copper to optical, they will unlock new architectural possibilities.
  隨著這些連接從銅線轉移到光纖，它們將解鎖新的架構可能性。

[19:05] So today, I'm going to explain why connectivity is becoming one of the defining characteristics and challenges of the AI era and why this technology transition matters to optics.
  因此，今天我將解釋為什麼連接性正成為人工智能時代的決定性特徵和挑戰之一，以及為什麼這次技術轉型對光學至關重要。

[19:15] Now, this isn't something far out in the future.
  現在，這並不是遙遠的未來。

[19:18] It's happening right now, this year, next year.
  它正在發生，就在今年，明年。

[19:20] We're in the ramp.
  我們正處於爬坡階段。

[19:23] And at Marll, we've been preparing for this moment for nearly a decade.
  在 Marll，我們為這一刻已經準備了近十年。

[19:28] We built the company very deliberately around the infrastructure required to move data at massive scale.
  我們非常有意識地圍繞著大規模移動數據所需的基礎設施來建立公司。

[19:36] And to understand why we made that bet, let's go back in time 10 years ago when I joined Marll as a CEO.
  要理解我們為什麼做出這個決定，讓我們回到十年前，當時我作為首席執行官加入 Marll。

[19:46] So prior to Marll, I spent 22 years at one company, Maxim Integrated Products,
  所以在加入 Marll 之前，我在 Maxim Integrated Products 這家公司工作了 22 年，

[19:50] which was a leading analog semiconductor company.
  這是一家領先的模擬半導體公司。

[19:53] And one of the unique things about working at an analog company is that your products go into virtually every piece of end equipment, every electronic system, every on market in the planet.
  在類比公司工作，獨特之處在於您的產品幾乎進入了每一個終端設備、每一個電子系統、地球上每一個市場。

[20:03] So over those two decades, I had a front row seat to just about every major technology trend.
  因此，在這二十年間，我親眼見證了幾乎所有重大的技術趨勢。

[20:07] First personal computing, then notebooks, digital still cameras, smartphones, eventually data center.
  首先是個人電腦，然後是筆記型電腦、數位相機、智慧型手機，最終是資料中心。

[20:16] And I watched wave after wave of technology reshape the whole industry.
  我見證了科技一波又一波地重塑整個產業。

[20:22] So I joined Marll and I didn't start off actually thinking about well what products do we have.
  所以我加入了 Marll，一開始並沒有真正思考我們有哪些產品。

[20:26] I reflected on where the industry was headed and it seemed clear to me even at that time back in 2016 that the next major growth cycle for semiconductors in the world really was going to be driven by the data platform companies.
  我反思了產業的發展方向，即使在 2016 年，對我來說也很清楚，全球半導體的下一個主要成長週期將真正由資料平台公司推動。

[20:43] Back then it was still the same ones as today.
  當時和現在一樣。

[20:44] companies like Google, Amazon, Microsoft, Meta and more specifically the semiconductor
  像 Google、Amazon、Microsoft、Meta 等公司，更具體地說，是半導體

[20:51] technologies that were required for those markets to move data, store data, process data and secure data, do it at massive scale.
  這些技術是那些市場移動數據、儲存數據、處理數據和保護數據所必需的，而且能夠大規模地進行。

[21:00] That was the vision we had.
  這就是我們的願景。

[21:05] But when I looked at the products we had at that time, very few of these were actually exposed to that trend.
  但當我審視當時我們擁有的產品時，很少有產品真正受到該趨勢的影響。

[21:10] It was kind of a problem.
  這有點像個問題。

[21:11] Less than 10% of our revenue 10 years ago was coming from data center.
  十年前，我們不到 10% 的收入來自數據中心。

[21:16] That's it.
  就這樣。

[21:16] A couple hundred million bucks.
  幾億美元。

[21:18] But more than 60% of our revenue back then was coming from consumer.
  但當時我們超過 60% 的收入來自消費市場。

[21:23] And so it was exciting time.
  所以那是一個令人興奮的時期。

[21:23] We were in virtual reality headsets.
  我們涉足虛擬實境頭戴裝置。

[21:26] We were in gaming consoles, streaming devices, wearables.
  我們涉足遊戲機、串流裝置、穿戴裝置。

[21:29] In fact, our claim to fame back then was Marll was designed in to the first Wi-Fi connected Barbie Dreamhouse.
  事實上，我們當時的成名之作是 Marll 被設計進了第一款 Wi-Fi 連接的芭比夢幻屋。

[21:38] That was our big design win.
  那是我們重大的設計勝利。

[21:41] It was real.
  這是真實的。

[21:41] In fact, the first week I was at Marll, they team briefed me on on what a great design win this was.
  事實上，我到 Marll 的第一週，團隊就向我簡報了這是一次多麼偉大的設計勝利。

[21:48] So that's where we were.
  所以，這就是我們當時的狀況。

[21:51] So we had a
  所以我們有一個

[21:53] Vision. There was a pretty big gap though between the reality that we were facing and where we saw the industry heading.
  願景。然而，我們所面臨的現實與我們所看到的產業發展方向之間，存在著相當大的差距。

[21:58] But we had conviction.
  但我們有信念。

[22:03] So we decided to bet the whole future of Marll on it.
  所以我們決定將 Marll 的全部未來押注在上面。

[22:06] So to do that, we needed a clear vision.
  因此，為了做到這一點，我們需要一個清晰的願景。

[22:09] And our vision at that time was pretty simple.
  而我們當時的願景相當簡單。

[22:11] And by the way, this is still the same vision that we have today, 10 years later, which is build a best-in-class pure play company focused on semiconductor solutions for data infrastructure.
  順帶一提，這仍然是我們今天所擁有的願景，十年後的今天，那就是建立一家專注於數據基礎設施半導體解決方案的一流純粹公司。

[22:22] Now, at that time, data infrastructure was not a recognized market category.
  現在，當時數據基礎設施並不是一個公認的市場類別。

[22:27] It was the term that we used to describe the infrastructure that was going to be required to move the world's data, store the world's data, process the world's data, and secure it.
  這是我們用來描述將要處理、儲存、處理和保護世界數據所需的基礎設施的術語。

[22:38] But like I said, we were not in that business yet.
  但正如我所說，我們當時還沒有從事那項業務。

[22:40] And frankly, we didn't even have a lot to work with.
  坦白說，我們手頭能用的東西也不多。

[22:45] Uh as we went after it, we had some.
  呃，當我們著手去做時，我們有一些。

[22:48] So my team and I came to a conclusion which is that we would need to build these capabilities internally and others we would need to build through strategic M&A and we had to get
  所以我和我的團隊得出了結論，我們需要內部建立這些能力，而其他的則需要通過策略性併購來建立，而且我們必須獲得

[22:57] focused because when you're transforming, it's not just deciding about what you're going to do.
  專注，因為當你轉型時，這不僅僅是決定你要做什麼。

[23:02] It's equally important to decide what you are not going to do.
  同樣重要的是決定你將不做什麼。

[23:08] So with that strategy in place, we got to work.
  因此，有了這項策略，我們開始工作。

[23:10] We began systematically building Marll around that vision.
  我們開始系統性地圍繞該願景建立 Marll。

[23:13] And it wasn't just one move.
  而且這不只是一個動作。

[23:15] There was a series of deliberate choices.
  這是一系列深思熟慮的選擇。

[23:17] We looked for the premium assets in the markets that mattered the most.
  我們尋找對我們最重要的市場中的優質資產。

[23:22] The best companies, best technologies, the best teams with the strongest market positions.
  最好的公司、最好的技術、擁有最強市場地位的最好的團隊。

[23:33] Now, we first started by divesting businesses that weren't aligned with our strategy.
  現在，我們首先剝離了與我們策略不符的業務。

[23:36] You can see some of those there.
  你可以在那裡看到其中一些。

[23:39] Then very quickly we acquired Cavium to strengthen our compute and networking capabilities.
  然後我們很快收購了 Cavium，以加強我們的計算和網絡能力。

[23:42] That was back in 2018.
  那是在 2018 年。

[23:46] 2019 we divested our Wi-Fi business.
  2019 年我們剝離了我們的 Wi-Fi 業務。

[23:48] Again we we we were focusing but we acquired a Vera to establish our custom silicon business and then Aquanchia to bolster our connectivity portfolio.
  我們再次專注，但我們收購了 Vera 以建立我們的客製化矽業務，然後收購了 Aquanchia 以加強我們的連接性產品組合。

[23:58] In 2021 we followed all that up by acquiring Infi for $10 billion.
  在2021年，我們透過以100億美元收購Infi來跟進這一切。

[24:04] was our largest acquisition to date and we got world-class data center connectivity technology into the company through that.
  這是我們迄今為止最大規模的收購，並透過這次收購，我們獲得了世界級的數據中心連接技術。

[24:10] and we acquired Anovium the same year adding high-end data center switching capability to the portfolio.
  同年，我們收購了Anovium，為產品組合增加了高端數據中心交換功能。

[24:17] So then we took a break we took a few years to digest and focused on unifying and building out our whole technology platform to address the data infrastructure opportunity.
  然後我們休息了一下，花了幾年時間消化，並專注於統一和構建我們的整個技術平台，以應對數據基礎設施的機會。

[24:27] But over the last 12 months we fired up the M&A engine again.
  但在過去12個月裡，我們再次啟動了併購引擎。

[24:32] We divested our automotive Ethernet business, again, Power of Focus, and acquired Celestial AI for its photonic fabric technology and XCON for scaleup switching.
  我們剝離了汽車以太網業務，再次強調專注的力量，並收購了Celestial AI及其光子交換技術，以及XCON用於擴大規模的交換業務。

[24:42] So, if you add it all up, over the last decade, we've invested roughly 22.5 billion through acquisitions.
  所以，如果你把所有這些加起來，在過去的十年裡，我們透過收購投資了約225億美元。

[24:50] We spent $18 billion organically inside of Marvel to develop the platform.
  我們在Marvel內部有機地花了180億美元來開發該平台。

[24:56] And then we divested approximately $4.5
  然後我們剝離了約45億美元的業務。

[24:59] billion worth of assets.
  價值十億美元的資產。

[25:01] So all in we've invested roughly $ 36 billion investing in this platform.
  所以總計我們在這個平台上投資了約 360 億美元。

[25:08] Let me show you the result of some of these investments.
  讓我向您展示這些投資的一些成果。

[25:12] First of all, we have built an incredible technology platform and it all starts with the advanced process node.
  首先，我們建立了一個令人難以置信的技術平台，而這一切都始於先進的製程節點。

[25:17] It's one of the most important decisions we made actually was to become a process node leader.
  實際上，我們做出的最重要的決定之一是成為製程節點的領導者。

[25:21] Now, Marll, Cavium, and some of the companies we acquired had all been fast followers, meaning you're like a node or two behind on everything you do.
  現在，Marll、Cavium 以及我們收購的一些公司都一直是快速跟隨者，這意味著你在所做的一切事情上都落後一到兩個節點。

[25:29] And that's largely a result of just not having enough scale.
  這在很大程度上是因為規模不足。

[25:35] That's usually why people do that.
  這通常是人們這樣做的原因。

[25:37] But as we integrated these businesses, we made the decision that if we're going to compete in data infrastructure, we had to be at the absolute leading edge.
  但當我們整合這些業務時，我們做出了決定，如果我們要進入數據基礎設施領域競爭，我們必須處於絕對領先地位。

[25:45] No choice.
  別無選擇。

[25:48] Now, here's a little known fact.
  現在，這裡有一個鮮為人知的 વાસ્તવિકતા。

[25:50] Marll skipped 7 nanometer completely.
  Marll 完全跳過了 7 奈米製程。

[25:54] We made a full no jump at that time from 14 and 16 nanometer all the way to five.
  當時我們從 14 和 16 奈米製程直接跳到了 5 奈米製程。

[26:00] Nobody does this.
  沒有人這樣做。

[26:02] Nobody takes that kind of a risk or a bet.
  沒有人會冒那種風險或打賭。

[26:04] But we did and it worked.
  但我們做了，而且奏效了。

[26:07] It worked really well flawlessly.
  它運行得非常好，完美無瑕。

[26:09] Actually our engineering team did an outstanding job executing this transformation.
  實際上，我們的工程團隊在執行這次轉型方面做得非常出色。

[26:14] So in early 2020 we released our first worldclass IP platform complete with die-to-do interfaces custom SRAMM high-speed sudis and more.
  因此，在 2020 年初，我們發布了第一個世界級的 IP 平台，配備了 die-to-do 接口、定制的 SRAMM 高速 sudis 等。

[26:25] Now certis is a good example of how we built this platform.
  現在，Certis 是我們如何構建這個平台的良好範例。

[26:27] It combined Marll's own core engineering strength with exceptional talent from a Vera Aquancha infi and others.
  它結合了 Marll 自身的工程核心實力以及來自 Vera Aquancha、infi 等公司的卓越人才。

[26:35] Now today that is a 1,500 person organization at Marll, second to none in terms of engineering scale and capability.
  如今，這已成為 Marll 擁有 1500 名員工的組織，在工程規模和能力方面無與倫比。

[26:44] So to support the process data portion of our mission, we built a best-in-class custom compute platform working in deep partnerships with the world's leading hyperscalers.
  因此，為了支持我們任務的流程數據部分，我們建立了一個一流的定制計算平台，與世界領先的超大規模雲服務提供商建立了深度合作夥伴關係。

[26:55] That business has been been doing very well for us.
  這項業務對我們來說一直做得非常好。

[26:58] In store data, we built a whole
  在內部數據方面，我們建立了一個完整的

[27:00] portfolio of storage controllers, CXLbased memory poolers, and nearmemory compute.
  儲存控制器、CXL記憶體集區器和近記憶體運算的組合。

[27:06] But here's where we really went allin, and that was in data movement.
  但這才是我們真正全力投入的地方，那就是資料移動。

[27:12] And this is where our high-speed connectivity portfolio.
  這就是我們的高速連接組合。

[27:15] And when you look at Marvel's data center business today, the vast majority of our revenue actually comes from connectivity.
  當您看看今天美滿電子（Marvell）的資料中心業務時，我們的大部分收入實際上來自連接性。

[27:22] from high-speed optical interconnect inside the data center to long reach optics between data centers to high-speed switching infrastructure.
  從資料中心內的高速光學互連，到資料中心之間的长距離光學，再到高速交換基礎設施。

[27:32] So today we are the undisputed connectivity leader and when you step back and look at what we built and where the market ultimately went I think the results speak for themselves.
  因此，今天我們是無可爭議的連接領導者，當您回顧我們所建立的以及市場最終的走向時，我認為結果不言而喻。

[27:43] So back in 2016 Marll was a $2.3 billion company.
  所以早在 2016 年，Marvell 是一家價值 23 億美元的公司。

[27:47] As we embarked on the transformation, actually in the first five years, we doubled the company, $4.5 billion in revenue.
  當我們開始轉型時，實際上在最初的五年裡，我們將公司規模翻了一番，營收達到 45 億美元。

[27:54] Over the next five years, our growth accelerated and according to consensus estimates on Wall Street for the current
  在接下來的五年裡，我們的增長加速，根據華爾街對當前的共識估計

[28:01] year we're in, we're set to grow about 2

[28:04] and a half times over the last 5 years

[28:06] to 11.4 billion.

[28:09] But in the recent couple of years, if

[28:10] you actually drill down, Marll has been

[28:12] growing like 40% a year. So the growth

[28:15] rate is actually accelerating in the

[28:17] last few years.

[28:19] So at this point Marll is off to the

[28:21] races. Okay. Um and based on the outlook

[28:23] that we shared in our earnings call last

[28:25] week, consensus estimates have come up

[28:27] and they expect us now to deliver 16.4

[28:31] billion in revenue next year.

[28:34] So as I said earlier when we started

[28:36] this journey, data center represented

[28:38] less than 10% of our revenue and we bet

[28:40] the farm on it. Last quarter it was over

[28:43] 75% of our revenue and growing very

[28:45] rapidly. This is a very different

[28:47] company than we used to be and the

[28:48] thesis has largely played out but but

[28:53] we're still in the early innings of this

[28:55] infrastructure buildout. The next phase

[28:57] is all in front of us. It'll have a

[28:59] different set of requirements and that

[29:02] brings us back to connectivity.

[29:05] So for the past several years as AI has

[29:07] created new demands on the

[29:08] infrastructure, we've seen the industry

[29:10] solve one major bottleneck after

[29:13] another. And first it was compute. I

[29:16] mean the industry needed dramatically

[29:17] more compute to enable modern AI and

[29:20] Nvidia did an incredible job leading

[29:23] that revolution and along the way became

[29:25] the world's first $5 trillion market cap

[29:29] company. Congratulations to Jensen and

[29:31] his whole team that's here. It was just

[29:33] a phenomenal phenomenal result.

[29:40] Next

[29:42] came the memory bottleneck.

[29:44] Larger models required enormous amounts

[29:46] of memory and bandwidth. And the memory

[29:49] companies are scaling aggressively now

[29:50] to meet that demand. And just recently,

[29:52] we've seen three new 1 trillion dollar

[29:55] market cap companies emerge in that

[29:57] market.

[29:58] But the bottleneck is shifting again.

[30:01] Now, it's connectivity that will define

[30:03] the limits of the infrastructure. Just

[30:05] like with compute and memory, the

[30:08] industry will rally to meet this

[30:10] challenge.

[30:12] Now, this isn't just me saying this.

[30:15] This is what we're hearing from our

[30:16] largest customers. The world's largest

[30:18] hyperscalers are now reimagining their

[30:20] entire network architectures.

[30:23] They recognize that scaling AI

[30:25] infrastructure is now first and foremost

[30:28] a connectivity challenge. As reasoning

[30:31] models, mixture of experts

[30:32] architectures, agentic AI, it all

[30:35] continues to evolve. More data has to

[30:37] move across the infrastructure demanding

[30:39] higher bandwidth and lower latency. And

[30:43] as workloads no longer fit within one

[30:44] data center, guess what? They need to

[30:46] build larger data centers or full

[30:48] campuses full of data centers and all

[30:50] the high-speed connectivity between

[30:53] them.

[30:54] Thus, the connectivity becomes a

[30:56] critical enabler of scaling compute. And

[30:59] increasingly, our customers recognize

[31:02] that optics is the way forward and

[31:04] they're looking to leaders like Marll to

[31:07] help them build larger, faster networks

[31:09] and at scale.

[31:13] So, when you look across the

[31:14] semiconductor industry at the leading

[31:16] companies supporting this infrastructure

[31:18] buildout, it becomes clear each of us is

[31:20] focused on a different part of the

[31:21] infrastructure.

[31:24] that shows up in the revenue mix.

[31:26] Some of the companies are compute first

[31:29] means the vast majority of their revenue

[31:30] is tied to compute with some of it tied

[31:33] to connectivity but most of its compute

[31:35] and it's obviously a critical part of

[31:37] the stack and that's why we have several

[31:39] mult you know trillion dollar plus

[31:41] companies in this group. Then you have

[31:43] the companies focused on memory and

[31:44] again all trillion dollar market cap

[31:46] companies at this point. It's

[31:47] unbelievable.

[31:49] And then you have Marll. We're

[31:51] different. We're unique

[31:53] today. The vast majority of our revenue

[31:55] actually comes from connectivity.

[31:58] So we built this company around data

[32:01] movement and today the vast majority of

[32:04] our revenue comes actually from

[32:05] connectivity. Now this spans a broad

[32:07] range of technologies and even the

[32:10] portion of our revenue that's from

[32:11] compute which you can see is

[32:13] fundamentally because customers embed

[32:15] our connectivity in their compute

[32:16] engines. So this gives us a unique

[32:19] position and perspective on these

[32:21] technology transitions that are

[32:22] happening and it creates a very

[32:24] different relationship that we can have

[32:26] with the rest of the ecosystem. We

[32:28] partner deeply with the compute

[32:29] companies. We partner deeply with the

[32:32] memory companies. These are very

[32:34] strategic relationships and in many ways

[32:36] we are the Switzerland of the industry

[32:38] and we work with everybody.

[32:41] Now, one of the best examples of the

[32:43] role that Marll plays in this ecosystem

[32:45] is the recently announced strategic

[32:48] partnership and expansion with Nvidia.

[32:51] And as part of this announcement that we

[32:52] made a few months back, Nvidia invested

[32:55] $2 billion into Marll. And we're

[32:57] expanding our partnership now across

[32:59] multiple multiple dimensions including

[33:02] optics, photonix, NVLink fusion.

[33:07] And I'm thrilled to announce that Jensen

[33:08] himself is here today. He's going to

[33:10] join me on stage. We're going to spend a

[33:13] few minutes chatting about the

[33:14] partnership and we're going to see where

[33:17] AI infrastructure goes from here. Uh so

[33:20] with that, let me please welcome to the

[33:22] stage Jensen Wong.

[33:32] >> Hey man,

[33:33] >> what's up Jensen? How you doing?

[33:34] >> Boy, that's a huge stage. That run a

[33:37] long ways.

[33:38] >> Are you out of breath? You okay? I know.

[33:40] >> Let's fire up. Good to see you.

[33:45] >> There you go. Yeah. Congrats on a great

[33:48] kickoff yesterday, GTC. You guys are off

[33:50] to the races this week.

[33:51] >> Thank you. Thank you.

[33:53] >> Um, look, maybe you heard some of what I

[33:55] just said. So, we're talking about

[33:56] connectivity today.

[33:57] >> The next trillion dollar company, ladies

[33:59] and gentlemen.

[34:00] >> Whoa.

[34:03] >> That would be exciting. Let's do it

[34:04] together. Let's do it together. Um, but

[34:07] it it really all starts with what's

[34:08] happening today in AI infrastructure

[34:10] kind of more broadly. So, how do you see

[34:12] that like just from the big picture

[34:14] standpoint? We're at this extraordinary

[34:15] moment. Customer demands through the

[34:18] roof. How do you see connectivity

[34:20] playing into this and the in the

[34:21] interconnect that's required?

[34:22] >> Yeah, that's really great. You know,

[34:24] yesterday um I I said that useful AI has

[34:28] arrived. It's the reason why your demand

[34:31] is going through the roof. It's the

[34:32] reason why my demand is going through

[34:33] the roof.

[34:34] >> Yeah. And and this new computing pattern

[34:36] that makes it possible is called agents.

[34:39] And these agents has a particular

[34:42] computing platform computing pattern

[34:44] that is disagregated and distributed.

[34:48] >> When you take a computing problem and

[34:50] you disagregate it into a lot of parts

[34:53] and you distribute it across the entire

[34:55] data center, what's necessary is

[34:58] connectivity. That's the reason why

[35:01] Matt's doing so well. That's the reason

[35:03] why Marll is so essential. We've

[35:05] distributed and disagregated computing

[35:08] so that it runs across these enormous

[35:10] clusters so that we could get agg we're

[35:13] aggregating the total compute, the total

[35:16] memory, the total bandwidth that we have

[35:19] and the what makes it possible is

[35:21] connectivity.

[35:22] >> Yeah, we're we're I mean we're we're

[35:23] we're seeing it. And then

[35:25] >> as you going to be the next trillion

[35:27] dollar company,

[35:28] >> we got a little work to do, but we're

[35:29] we're on our way. We're on our way.

[35:31] Thank you, Jensen.

[35:33] Well, let's talk about let's talk about

[35:35] scale. I mean, we used to talk about

[35:36] tens of GPUs and CPUs and and XPUs

[35:39] connected now thousands now maybe

[35:41] millions at some point. So, as you scale

[35:43] the compute and you scale the

[35:45] connectivity, I think we talked about

[35:47] things like agents, but how do you think

[35:48] about that, you know, across data

[35:50] centers within data centers, how do you

[35:53] think about connectivity at large

[35:54] playing that role and what kinds of

[35:56] technologies do you think are important

[35:57] there? Well, at the foundation of it,

[35:59] the agent computing pattern requires um

[36:03] an orchestration system that allows the

[36:06] large language models, the computing to

[36:09] be able to think and reason and come up

[36:11] with plans, but it also has to use tools

[36:14] and, you know, browse the internet,

[36:15] access memory, access long-term memory,

[36:18] deal with short-term working memory. All

[36:21] of that requires a lot of connectivity.

[36:23] But it's also the it's also the case and

[36:26] if you look at the way we introduced

[36:27] Vera Rubin Hopper was designed for

[36:31] training.

[36:32] >> Yeah.

[36:32] >> Grace Blackwell introduced MVLink 72 our

[36:36] first scale up fabric. Uh it introduced

[36:40] the idea of extremely fast inference for

[36:44] MOE models that are very large mixture

[36:46] of expert models that are extremely

[36:48] large. And so Grace Blackwell was for

[36:50] inference. Vera Rubin is to run agents.

[36:53] Yeah.

[36:53] >> Which is the reason why the Vera Rubin

[36:55] system includes of course the Vera Rubin

[36:59] thinking AI but it also includes Vera

[37:02] CPUs for orchestration. It includes Vera

[37:06] CX for storage acceleration for managing

[37:10] long-term memory. And the way that I

[37:12] think about these systems, you know,

[37:13] sometimes uh maybe the CSP wants to

[37:16] design their own custom chip and between

[37:18] us, we also partner together on MVLink

[37:21] Fusion

[37:22] >> which makes it possible for you to use

[37:24] the same system architecture

[37:28] and with Vera Rubin inside some of your

[37:31] semi-custom chips, a lot of your

[37:34] interconnect silicon photonics and

[37:36] optics and technology such and we can

[37:39] create essentially

[37:40] a disagregated, distributed and

[37:43] heterogeneous data center. And so that's

[37:46] that's the big idea. And yet their their

[37:48] system architecture is identical. Their

[37:51] networking technology can leverage a lot

[37:53] of NVIDIA stack. Um the CPU could be

[37:56] Vera and yet it can leverage a lot of

[37:58] your stack. So MVLink Fusion is about

[38:00] taking Nvidia's technology and our

[38:03] platforms, Marll's technologies and PL

[38:05] and we fuse it. That's why it's called

[38:07] Fusion.

[38:07] >> Yeah. No, I think, you know, I think

[38:09] about the partnership and we've been

[38:10] working together a long time. I think

[38:12] memorializing it with the investment,

[38:14] which we really appreciate. I think it's

[38:16] been it's been huge for us. We're

[38:17] honored to have it.

[38:18] >> I I you know, who doesn't love making

[38:20] money? It's nice to give

[38:24] >> It's done well since you invested. So,

[38:26] yeah.

[38:27] >> I I love getting rich.

[38:29] >> Just follow him.

[38:31] >> Give Matt all my money and just watch

[38:33] him make money.

[38:34] >> That's what I'm doing every day. That's

[38:36] what I'm doing every day. But I think

[38:37] these things you talked about which we

[38:38] brought to fruition, NVLink Fusion

[38:42] working together on optics, I mean I

[38:44] think the era of agents and kind of your

[38:46] new platform now I think it's ideally

[38:48] suited. I mean NVLink Fusion we we had

[38:51] this idea years ago, right? But I think

[38:53] it was a little ahead of its time. And

[38:55] now when and I wanted to see if you

[38:56] agree when when you think about kind of

[38:59] your platform and then some of the

[39:01] custom networking and compute needs that

[39:03] our customers have and the the ability

[39:05] and the the need to interoperate and

[39:07] work together. It seems like the time is

[39:09] now between Marll and Nvidia to really

[39:11] go enable our customers to have that

[39:13] flexibility that they're looking for and

[39:15] really use the era of agents to scale

[39:18] our platforms together.

[39:19] >> Yeah. you know, ultimately I do think

[39:22] that if you buy nothing but Nvidia, it's

[39:25] okay.

[39:27] Okay. I mean, if but if you absolutely

[39:30] must uh design your own AS6, um we're

[39:34] still happy having Nvidia be inside that

[39:36] data center.

[39:37] >> And so, you know, you don't have to buy

[39:39] everything from us. Just buy something

[39:42] from us. You know, we're we're we're

[39:45] happy to support you and support the

[39:47] customer. And so so I think that between

[39:50] the two of us you have the benefit

[39:52] >> of a general purpose very high

[39:55] efficiency you know a a system that is

[39:57] very well built starting with you know

[39:59] of course Vera Rubin but anything that

[40:01] you want to extend to specialize uh you

[40:04] can do so as well which is the reason

[40:05] why your customers and mine Nvidia is in

[40:08] AWS Marll's in AWS Nvidia is in all of

[40:12] the clouds and it's wonderful to see

[40:14] Marll expand into all of these different

[40:16] clouds. Yeah, great. Thanks. Hey, one

[40:18] last one for you.

[40:18] >> Just leave some business for me. You

[40:20] know,

[40:21] >> look, we're your best salespeople right

[40:22] now. Are you a great saleserson?

[40:23] >> Your best salesperson

[40:25] >> working together. Final question for

[40:27] you. A lot of my talk is about some of

[40:30] the transition, especially as you go to

[40:32] inside the rack from copper to optical.

[40:34] It's obviously not going to be a one

[40:35] zero. It's going to take, you know,

[40:37] there's time and there's different use

[40:38] cases. But how do you see that playing

[40:40] out right now, the transition from

[40:41] copper uh to optics and maybe how we can

[40:44] work together there, too? Well, we

[40:46] should use copper as much as we can for

[40:48] as long as we can, but copper has its

[40:50] limits. Copper has its limits with

[40:52] bandwidth and also with distance. And

[40:55] so, so ultimately um this the right

[40:58] strategy is to scale up with copper as m

[41:02] as long as you can. After that, you

[41:05] scale up further with optics and you

[41:09] scale out with optics and you scale

[41:12] across with optics. And so you use

[41:15] optics wherever you must. You use copper

[41:17] wherever you can. And so I think that

[41:19] that that in that intersection is going

[41:22] to continue for a long time. Here's

[41:24] here's the the the bottom line is in the

[41:26] next 5 10 years we're going to use a ton

[41:29] of copper and we're going to use tons

[41:32] and tons of optics. And so these these

[41:34] data centers are part of infrastructure

[41:37] now. And the reason why I say that AI is

[41:40] now useful, useful AI has arrived is

[41:43] because now AI is profitable and tokens

[41:46] are profitable. When token production is

[41:49] profitable, everybody wants to make more

[41:51] tokens, which is the reason why, you

[41:53] know, Marll's demand is so high is our

[41:56] demand is so high because everybody

[41:58] wants to produce more tokens because

[41:59] it's used all over the place by agents.

[42:02] >> Absolutely. Well, I think you touched on

[42:03] a bunch of things I'm going to cover

[42:04] later. If you want to do the rest of my

[42:06] presentation, you can. So ladies and

[42:08] gentlemen,

[42:08] >> these beautiful slides, you know,

[42:10] >> Matt, just just sit right there. I'll be

[42:13] >> You take it from here. All right, Jensen

[42:14] Juan. Good to see you, brother. All

[42:16] right. Take care.

[42:17] >> Okay, guys. Thank you.

[42:18] >> Thank you, Jensen.

[42:20] >> Bye, Marll.

[42:23] >> Bye, Marvel.

[42:25] All right. Outstanding. Outstanding.

[42:28] Super fun to have Jensen here as always.

[42:31] All right. So, we've been talking a lot

[42:32] about connectivity. Jensen and I just

[42:34] covered this. So, let's like dive in

[42:35] now, right? Let's go one level deeper.

[42:38] So AI infrastructure

[42:40] spans every distance. It spans from

[42:43] hundreds or even a thousand kilometers

[42:45] between data centers to just millimeters

[42:48] inside the package. Every one of those

[42:50] distances, it requires a different

[42:52] solution. It's a different technology,

[42:55] different engineering team. It's a

[42:56] different completely different set of

[42:57] experts and in many cases it's a

[43:00] different supply chain. So these are not

[43:03] variations of the same problem. What you

[43:05] have here is fundamentally different

[43:06] engineering challenges and that's what

[43:08] we're going to walk through next.

[43:10] All right. So let's start with the

[43:12] longest distance. Jensen referred to

[43:14] this. This is scale across connecting

[43:17] data centers together. Now every major

[43:19] cloud provider has hundreds of data

[43:21] centers around the world and all of

[43:23] those data centers need to communicate

[43:25] with each other. This is fundamentally a

[43:27] longd distance connectivity problem.

[43:30] We're talking about links that can span

[43:32] hundreds or even a thousand kilometers.

[43:34] This requires very specific, very

[43:36] complex technology called coherent

[43:38] modulation. At the heart of it is a

[43:41] specialized digital signal processor or

[43:43] DSP. It's designed to push enormous

[43:46] amounts of data cross fiber optic cables

[43:49] over very long distances with extremely

[43:52] high reliability. There's only a few

[43:55] companies in the world that build these

[43:56] coherent DSPs and we're one of them.

[44:00] Marll has been a leader in this

[44:01] technology for many generations. We

[44:03] build optical modules that contain all

[44:06] the electronics needed to drive and

[44:08] modulate the laser and transmit

[44:11] data over long distances.

[44:15] So, I've got a little showand tell here

[44:18] in my pocket. Not holding up a chip this

[44:20] time. I'm holding up an optical module.

[44:23] This is one of our coherent optical

[44:24] modules. This is an incredibly complex

[44:27] piece of engineering.

[44:30] At Marll, we build the entire module.

[44:32] This is ours. It includes the advanced

[44:34] node CMOS DSP. It's among the most

[44:37] complex chips, just the DSP alone that

[44:39] we design at Marll, but it also

[44:42] incorporates inside our fourth

[44:44] generation silicon photonix technology.

[44:48] That's inside here. We've been

[44:50] developing that technology and in

[44:52] production for a decade on silicon

[44:54] photonics.

[44:56] It also includes our own broadband

[44:58] analog components that we designed which

[45:00] is designed in silicon germanmanium. So

[45:03] Marll pioneered this technology starting

[45:05] with 100 gbits per second a decade ago

[45:08] then moving to 400 gig and now shipping

[45:11] 800 gig in volume. And later this year

[45:15] we'll be sampling the world's first 1.6

[45:18] 6 terab 2nanmter coherent optical

[45:21] solution and that couldn't come at a

[45:23] better time. Demand for bandwidth has

[45:26] never been greater.

[45:28] All right, now let's go inside the data

[45:30] center. So these data centers can be

[45:32] very large spanning hundreds of meters

[45:33] and they contain racks and racks of

[45:36] compute servers. Now each rack typically

[45:38] has a switch at the top with servers

[45:41] connected into that switch. Those rack

[45:43] level switches connect to the spine and

[45:45] then the core switches. This creates the

[45:47] network fabric that ties the entire data

[45:49] center together and all of that is

[45:51] connected through fiber optic cables.

[45:54] Now once again optical modules drive

[45:56] data transmission over those fiber optic

[45:59] cables. But this time the modulation

[46:01] scheme is different. Instead of coherent

[46:04] technology we use a more power optimized

[46:07] modulation technology which is called

[46:09] PAM 4. So the two key semiconductor

[46:12] solutions for this part of the market

[46:14] are the PAM 4 chipset inside the module

[46:17] and then the cloud switching

[46:18] infrastructure that ties the data center

[46:21] together.

[46:23] Marll builds both. Starting with the PAM

[46:26] 4 chipset, we build the industry's

[46:28] leading PAM4 DSP solution

[46:32] and also the high-speed analog

[46:34] components that go around them including

[46:36] trans impedance amplifiers or TAS and

[46:38] laser drivers. These are also in silicon

[46:40] geranium by the way. And we've led the

[46:42] industry through every major transition

[46:44] of PAM technology starting at 50 gig,

[46:47] 100 gig, 200, 400, and 800. Then last

[46:51] year, we began ramping Marll's 1.6T

[46:54] 3nanmter PAM 4 solutions, leading the

[46:57] industry's transition to 1.6T

[46:59] connectivity.

[47:01] Now, for Ethernet switching, Marll has a

[47:03] similarly complete portfolio of products

[47:05] from 12.8 terabs to 51.2 2 terabs. And

[47:09] today

[47:11] we announced our new 100T Ethernet

[47:14] switch specifically designed for AI data

[47:17] centers with the industry's lowest

[47:18] power.

[47:20] Woo!

[47:22] Special announcement for Computex. We

[47:25] waited. So you put it all together, we

[47:28] provide a complete solution for

[47:29] connectivity inside the data center.

[47:33] Now let's move inside the rack.

[47:35] The goal here is to connect the largest

[47:37] possible number of processors together

[47:39] in a full anytoy configuration. In other

[47:42] words, every processor can communicate

[47:44] directly with every other processor. And

[47:47] Jensen talked about this. The first

[47:49] company to bring this architecture to

[47:50] market was Nvidia with NVL 72 named for

[47:54] the 72 GPUs connected together inside a

[47:57] single rack. And this required a

[47:59] completely different approach to

[48:00] connectivity was a different class of

[48:02] switch and the ability to drive very

[48:04] high-speed signals over copper back

[48:06] planes inside the rack. So today this is

[48:10] not the domain of optics. This is the

[48:13] domain of copper and the core

[48:15] differentiator here is the electrical

[48:17] certis technology not the optical. Now,

[48:20] Marll also has leading electrical certis

[48:24] at 200 Gbits per second today. And we've

[48:27] demonstrated already over the last

[48:29] couple of years, 400 Gbits per second

[48:31] for the future. So, we're building this

[48:34] series technology into our customers

[48:36] custom silicon and their XPUs and also

[48:38] into our own scaleup switches.

[48:42] All right. Now, let's go all the way

[48:43] inside the package

[48:46] here. We're not talking about meters

[48:47] anymore. We're talking about

[48:48] millimeters. And you might not actually

[48:50] think about this as a connectivity

[48:52] challenge, but today most advanced chips

[48:55] have multiple chiplets inside the

[48:57] package. So when you have 2 and 1/2D or

[48:59] 3D packaging, it's fundamentally a

[49:01] connectivity technology actually. And it

[49:04] allows these chiplets to sit very close

[49:05] together inside a package and

[49:07] communicate through ultra high-speed

[49:09] short-reach die-to-dfaces.

[49:12] And Marll has leading die to dieerties

[49:14] and leading capability in advanced

[49:16] packaging allowing our customers to

[49:18] build some of the most complex unique

[49:20] multi-d chips in the industry.

[49:24] So as you can see connectivity for AI

[49:27] data centers requires a very broad

[49:29] portfolio of technologies. Each distance

[49:32] requires a very different solution. And

[49:34] Marll has the industry's most complete

[49:36] portfolio from millimeters to

[49:38] kilometers. every hop, every distance.

[49:41] And it turns out having all of those

[49:43] capabilities under one roof is unusual.

[49:46] It's unique. When we go and compete,

[49:49] normally there's a different set of

[49:51] companies that we compete against in

[49:52] each one of these categories across

[49:54] these different distances. But this is

[49:56] what makes us unique. We're the one-stop

[49:59] shop. We're the leader across the entire

[50:02] connectivity stack. And that brings us

[50:04] to the next major challenge facing the

[50:06] industry. So,

[50:09] what you probably notice as I described

[50:11] these different solutions in the last

[50:12] couple of slides is there's different

[50:15] solutions for different distances and

[50:17] that some of those connections today are

[50:19] optical and some of those connections

[50:20] today are electrical. And it's it's

[50:23] actually defined by distance. And so the

[50:26] the connections on the left side of this

[50:27] chart are optical today. That means they

[50:30] use fiber optic cables to transmit light

[50:32] with complex electronics on either side

[50:34] of the cable to drive and modulate the

[50:36] laser that's transmitting that light.

[50:39] The connections on the right side of

[50:40] this are electrical. So they use copper

[50:42] cables or just copper traces that are

[50:45] printed on the circuit board or even

[50:46] microscopic copper routing inside the

[50:48] package. So the common theme here is

[50:51] copper. And in the middle you see the

[50:54] wall, the copper wall. And the wall is

[50:57] defined by the longest distance you can

[50:59] transmit a signal over copper. So before

[51:01] you have to move to an opt before you

[51:03] have to move to an optical connection.

[51:05] So this is an important distinction

[51:07] because copper is simple and it's low

[51:09] cost and as Jensen said you want to use

[51:11] it for as long as you can. It's very

[51:12] practical.

[51:14] Um but optics and optics is more

[51:16] complicated. It requires lasers,

[51:18] photonics, complex electronics. So it's

[51:21] it's a bigger lift but it's going to be

[51:22] needed. And the copper wall, what I'm

[51:25] here to tell you today is it's about to

[51:26] move. It's going to move again and it's

[51:28] going to take over the rack itself. So,

[51:31] this is creating an explosion in demand

[51:33] for the optical industry. An incredibly

[51:36] complex engineering challenges are

[51:37] coming and along the way. So, why is

[51:40] this happening?

[51:42] So, it's not just somebody's preference

[51:43] to go do this. This is physics. The

[51:45] distance a signal can travel over a

[51:47] copper cable is inversely

[51:50] proportional to the bandwidth. So every

[51:52] time you double the bandwidth you have

[51:54] to cut the distance in half. Today the

[51:56] highest speed production systems in the

[51:58] world run at 200 Gbits per second per

[52:00] lane just to give you an example. So at

[52:02] that bandwidth the cable length is

[52:04] limited to roughly 2 and 1/2 mters. Now

[52:06] by comparison systems running at 100 gig

[52:09] could use about 5 m cables and the

[52:11] height of the rack is about 2 m. So once

[52:13] you account for all the routing inside

[52:15] the rack 2 and 12 m is right at the

[52:17] limit. So when we move to 400 gig, we

[52:20] can no longer fully connect the rack

[52:22] with copper. So the wall is moving and

[52:26] it's moving now.

[52:29] And going forward, even the connections

[52:30] within the rack will become optical and

[52:32] the whole industry knows this is coming.

[52:34] So we've been preparing for this moment,

[52:35] not just Marll, but the industry. And

[52:37] you see this in Taiwan, by the way, and

[52:39] the supply chain and the ramp up that's

[52:41] happening. The ramifications for this

[52:43] are actually enormous because each time

[52:45] the m the the wall moves one step to the

[52:47] right, the number of connections that

[52:50] you have goes up by at least an order of

[52:53] magnitude. So it's creating this

[52:55] explosion in demand as I mentioned and

[52:57] the optical supply chain needs to scale

[52:59] up massively and be ready. But we've

[53:02] seen this movie before. Okay, I mean 20

[53:04] years ago and I remember this when

[53:05] state-of-the-art was 10 gigabits per

[53:08] second inside the data center. It was 10

[53:10] gig and we used copper cables all across

[53:12] the data center. Optics back then was

[53:14] reserved for just very very long

[53:16] distances. It was essentially like a

[53:18] telecom technology. But when the wall

[53:20] moved, the optics industry actually rose

[53:22] to the challenge. And today all the

[53:25] hyperscaled data centers in the world,

[53:26] they're all optically connected. And as

[53:28] we saw in that transition, it did

[53:30] require new solutions. You couldn't use

[53:32] the same power hungry kind of telecom

[53:34] approach, which is where PAM 4 came in.

[53:37] was optimized for power, density, and

[53:39] reach and requirements specifically

[53:41] tuned to inside the data center. And

[53:43] Marll was one of the key innovators

[53:45] there. So, we're about to see the same

[53:47] wave of innovation needed as optics

[53:50] moves inside the rack. And that's with a

[53:53] technology called co-ackage optics or

[53:55] CPO. You hear a lot about this now. I'm

[53:57] gonna I'm going to tell you more.

[54:00] CPO is a technology where we bring the

[54:02] optical connections all the way to the

[54:04] package itself right next to the compute

[54:07] either the custom compute or the

[54:09] switching silicon and the fundamental

[54:12] challenge we're solving with CPO is

[54:13] density and power now remember the

[54:17] number of connections inside the rack is

[54:19] like 10x the number of connections

[54:22] between the racks so we try so if we

[54:25] just try to use the same optical

[54:26] technology used across the racks in the

[54:28] data center, you you wouldn't have

[54:30] enough power. You wouldn't have enough

[54:32] physical space. You cannot fit all these

[54:34] standard optical modules and cables as

[54:37] they are today. It just doesn't work.

[54:38] It's not possible. So, the industry has

[54:40] been inventing this co-ackage optics

[54:42] concept which brings the optical fiber

[54:44] right to the package and it tightly

[54:46] couples the electronics that drive the

[54:47] signal over the fiber directly with the

[54:49] custom compute or switching silicon. So

[54:52] this is a massive change and it's hard

[54:56] because you're combining some of the

[54:57] most advanced technologies in the chip

[54:59] industry. Leading edge cos silicon

[55:02] photonics advanced packaging optical

[55:05] interconnect all manufactured in a in a

[55:08] small tightly integrated system. So the

[55:10] complexity is very high but it's the

[55:12] only way to continue scaling bandwidth

[55:14] and overcome this limitation that I

[55:16] talked about with copper while reducing

[55:18] power at the same time. So this is where

[55:21] the industry is headed and this is one

[55:22] of the reasons that Marll has invested

[55:25] for more than a decade in silicon

[55:27] photonics optical DSPs all the analog

[55:30] broadband components around it and all

[55:32] the advanced packaging you need to pull

[55:34] this off. It needs to all come together

[55:37] actually in CPO. So this isn't some

[55:39] futuristic thing guys. Okay, it's

[55:42] happening now. And in fact I brought a

[55:44] couple of Marll examples with me today.

[55:46] So let's do a quick uh a quick show and

[55:49] tell. Okay.

[55:52] Okay. So over here you have a

[55:53] traditional Ethernet switch. This is our

[55:56] 100T terlink switch that we announced

[55:58] today. And you guys are the first to see

[56:00] it actually everybody here in the room.

[56:02] You can see the switch in the middle of

[56:03] the board. Copper traces inside the PCB

[56:07] carry the signal to the front panel

[56:08] which is here. And this is where all the

[56:10] optical modules plug in. Now let's move

[56:12] over here.

[56:14] This is a CPO based switch right here.

[56:17] Now, notice that there's still the

[56:18] switch silicon in the middle. That's

[56:20] right in the center of the die of the

[56:21] package. In this case, this is our 51.2T

[56:24] switch. And all around the edges are 16

[56:27] 3.2T

[56:29] optical engines. So, the 16 * 3.2 you

[56:33] get 51.2. So, this is this is d the

[56:36] fiber is directly attached now to the to

[56:39] these engines. It's not to the front

[56:40] panel. So we've completely eliminated

[56:42] the copper traces on the PCB. Light

[56:45] comes directly out of the package. Okay,

[56:47] this is a very very complex piece of

[56:49] engineering and it was very uh it's very

[56:51] cool to be able to show this off today.

[56:54] Okay, so co-ackage optics is here and

[56:56] the industry is scaling up to meet the

[56:58] challenge and as we've seen time and

[57:00] time again, every time we reach a

[57:02] physical barrier, we break through it

[57:03] with technology and innovation. In this

[57:06] case, by replacing copper with fiber.

[57:09] Because unlike electrons traveling over

[57:11] copper wires, the distance that photons

[57:13] can carry a signal through glass is

[57:15] largely unrelated to the bandwidth. So

[57:17] as AI infrastructure demands even higher

[57:19] transmission speeds and needs to scale

[57:21] to larger and more complex systems,

[57:23] spanning millions of processors woven

[57:25] together now, not thousands or hundreds,

[57:27] optical connectivity will increasingly

[57:29] become the deacto solution. So the real

[57:32] question becomes, what does it take to

[57:34] deliver optics across the full AI

[57:36] infrastructure stack? What's it going to

[57:38] take? Well, it starts with recognizing

[57:41] there is no single technology for the

[57:44] entire data center. It's not how this

[57:46] works. There's no one-sizefits-all all

[57:48] there's no oneizefits-all solution.

[57:50] There's no shortcuts. There's no easy

[57:52] way to the end here.

[57:54] There's not a single architecture,

[57:56] modulation scheme, frequency band,

[57:59] you know, unique technology that's going

[58:00] to do it all. There's no free lunch.

[58:03] That's why we are pursuing

[58:06] a bunch of different unique optical

[58:08] paths across every distance to get here.

[58:11] Each one of these technologies they have

[58:12] up here is optimized for a different

[58:14] design point. Each one enables a

[58:16] critical part of the infrastructure and

[58:18] addressing different requirements for

[58:20] density, bandwidth and power and

[58:22] integration all across the stack.

[58:26] So if optical interconnect is the

[58:27] underlying technology for which next

[58:29] generation AI infrastructure is built,

[58:31] then Marll is building the broadest

[58:33] portfolio with the deepest bench in the

[58:35] industry. But no company can deliver

[58:37] this transformation alone. And as Jensen

[58:40] talked about earlier, right, it takes an

[58:42] ecosystem to get here.

[58:45] So like I said, technology innovation is

[58:48] great. Um it's it's part of the

[58:50] challenge, but not all of it. Um, but

[58:53] demonstrating this at scale is really

[58:55] what matters. And at this point, if

[58:57] you're just operating on a PowerPoint or

[58:59] a demo, Press release, it's not going to

[59:04] get you there. Customers need solutions

[59:06] now that are ready. They're reliable.

[59:08] They need to be manufacturable and be

[59:10] ready to deploy at scale. So, Marvel and

[59:12] our ecosystem partners have been doing

[59:14] this for a long time. We've already

[59:16] shipped hundreds of millions of DSPs.

[59:17] We've accumulated through our through

[59:19] our um through our volumes tens of

[59:22] billions of device hours of data in the

[59:25] field. This experience matters because

[59:28] these products have to work not just in

[59:30] the lab but in the world's largest data

[59:32] centers at very high volume and very

[59:33] reliably for years. So that requires

[59:37] investing ahead in the manufacturing

[59:38] ecosystem. You've got to build the

[59:40] capacity and the supply chain

[59:41] infrastructure before the market

[59:44] arrives. This is why the ecosystem

[59:46] matters so much.

[59:47] And it matters a lot here in Taiwan by

[59:49] the way. Now, one of our most important

[59:52] partners at Marll in this journey has

[59:54] been Advanced Semiconductor Engineering

[59:56] or ASSE. Now, ASSE is one of the world's

[59:59] leading semiconductor manufacturing

[01:00:01] companies. They have more than a 100,000

[01:00:03] employees with operation in Asia and

[01:00:05] actually all around the globe with a

[01:00:07] decadesl long track record of helping

[01:00:08] enable pretty much every major

[01:00:10] technology transition we've gone through

[01:00:12] in the semiconductor industry. Now

[01:00:15] leading ASSE through this period of

[01:00:17] transformation is someone that I know

[01:00:19] quite well. He spent more than 25 years

[01:00:21] helping shape both the company and the

[01:00:23] industry.

[01:00:25] Today I'm thrilled to have my next guest

[01:00:27] speaker come up which is ASSE CEO Dr.

[01:00:31] Tien Woo. Tien, please join me on the

[01:00:34] stage. Thank you.

[01:00:42] Hey

[01:00:45] Tienne, how are you?

[01:00:46] >> Thank you for you invite in inviting me

[01:00:48] to Computch.

[01:00:49] >> Great to see you. It's an honor to have

[01:00:51] you on stage with us. Oh,

[01:00:52] >> it's my honor.

[01:00:53] >> Um, look, you've

[01:00:55] >> we've been working together a long time.

[01:00:57] And,

[01:00:58] >> you know, when I became the CEO,

[01:01:01] you know, we had a set of ambitions. We

[01:01:03] talked to a lot of our suppliers. I've

[01:01:05] known you even before I was the Marvel

[01:01:07] CEO, when I was

[01:01:08] >> an executive back at Maxim and we worked

[01:01:10] together there. But part of what and

[01:01:13] maybe explain to the audience too that

[01:01:15] sometimes people don't realize is that

[01:01:16] as a as a key supplier into this

[01:01:18] ecosystem, you have to make bets, right?

[01:01:21] You got to make bets on the companies

[01:01:22] you work with. You got to make bets on

[01:01:23] who you think is going to be successful.

[01:01:25] And we really appreciate that ASC bet on

[01:01:30] Marll very early, very early. And we've

[01:01:34] seen great success actually based on

[01:01:36] that. But I'm just curious if you could

[01:01:38] share your perspective maybe on where

[01:01:40] Marll was, what your thought process is

[01:01:42] and then where are we today in our

[01:01:44] journey together. Um so so be great to

[01:01:47] hear from you Ten. Thank you.

[01:01:48] >> Okay. I think the best way to describe

[01:01:49] this is a gradual process. The first

[01:01:53] decision was not difficult. Marllist

[01:01:56] company has a very good reputation has

[01:02:00] gone through a lot of transition. So the

[01:02:02] track record of Marll has already been

[01:02:05] there. the product set uh was a little

[01:02:08] bit obsolete at the time when you

[01:02:10] joined. So the first one is the business

[01:02:14] model needs to be aligned. Taiwan ASC is

[01:02:19] in the manufacturing sector. So we're

[01:02:21] looking for bet

[01:02:24] not only on betting on your success.

[01:02:26] We're also betting on somebody who can

[01:02:29] provide the insight for the next

[01:02:32] generation architecture and also the

[01:02:34] technology requirement. As you know the

[01:02:37] Taiwan company invest infrastructure and

[01:02:40] capac 10 years ahead of time. Big bet.

[01:02:44] We're only counting on whatever capacity

[01:02:47] we put it in will be needed and will be

[01:02:50] utilized. That's how we make money. So

[01:02:53] betting on company that we believe will

[01:02:56] give us very good insight well into the

[01:02:59] future becomes very important. So that's

[01:03:02] how the decision was made at the very

[01:03:03] beginning. And for the last 10 years,

[01:03:05] I'm just really happy. Everything that

[01:03:08] we talk about it, it was a dream 10

[01:03:10] years ago. was a dream

[01:03:11] >> and today we are going to ship it and

[01:03:14] you just mentioned that you're going to

[01:03:16] have 40% growth for the next few years.

[01:03:19] I believe you're going to up you're

[01:03:20] going to beat that.

[01:03:22] >> So we're busy now preparing the capacity

[01:03:25] for you.

[01:03:26] >> Yes. Um we also appreciate that over the

[01:03:28] last 10 years we have gone through a lot

[01:03:31] of strategic discussion right you make

[01:03:34] commitment to us we make investment for

[01:03:37] you and over time we're going to produce

[01:03:39] more of your parts I think that's the uh

[01:03:42] really a short story for how that

[01:03:44] decision will come

[01:03:45] >> yeah no it's been a great story maybe

[01:03:46] one more for you um you know the

[01:03:49] ecosystem here in Taiwan is so unique

[01:03:51] and like you said it takes like decade

[01:03:53] of investment before you really can see

[01:03:55] the return And there's just such such a

[01:03:57] power that's happening here. How do you

[01:03:59] describe it to to people here and also

[01:04:01] there's a lot of people around the world

[01:04:03] watching and then what makes it possible

[01:04:06] here? Why is it unique? And then what

[01:04:07] also makes it difficult to replicate

[01:04:09] this in the rest of the world and but

[01:04:11] but at the same time there's

[01:04:12] globalization. So how do we think about

[01:04:14] those dynamics? I think that'd be an

[01:04:16] interesting one.

[01:04:17] >> I think the reason why you're asking the

[01:04:18] question there's a lot of competing

[01:04:20] forces and also uncertainty across the

[01:04:23] world. So I think my belief is any

[01:04:27] business needs to have vision as well as

[01:04:31] long-term alignment on value. So in the

[01:04:34] business model the whole Taiwan sector

[01:04:37] is built on capacity utilization and

[01:04:40] also innovation and technology

[01:04:42] investment way ahead of the curve.

[01:04:44] That's what Taiwan's value. So with the

[01:04:47] fabulous company or with specific IDM

[01:04:50] company that business model aligns

[01:04:53] beneath that will be the economy of

[01:04:56] scale Taiwan accumulated 40 years based

[01:05:00] on the PC transition to the wireless to

[01:05:02] the mobile computing to the data center

[01:05:04] now we're into HPC so that 40 years of

[01:05:08] experience accumulated 350,000

[01:05:12] semiconductor employee also accumulated

[01:05:15] 1.1 1 million high-tech employee and

[01:05:19] many of them are here that experience

[01:05:22] becomes extremely valuable combined with

[01:05:25] the economy of scale as well as the

[01:05:28] cluster efficiencies. So when you think

[01:05:31] about the workforce with years of

[01:05:33] experience behind it when you think

[01:05:35] about the cluster efficiency when you

[01:05:38] think about the capacity economy scale

[01:05:42] we already put it in. But one more thing

[01:05:45] I think Taiwan good or bad we had fewer

[01:05:49] choices than the other region like

[01:05:51] United States. So most of the engineer

[01:05:54] when they come out they have few choices

[01:05:56] to make semiconductor

[01:05:58] IT industry becomes a attractive choices

[01:06:02] in Taiwan not necessarily in the other

[01:06:05] region. So with all of this combined I

[01:06:08] think this ecosystem is very very

[01:06:10] difficult to replicate. It is not

[01:06:13] impossible but will take years.

[01:06:16] >> Right. Great. Well, thank you so much. I

[01:06:18] appreciate the partnership so much.

[01:06:20] Thank you.

[01:06:20] >> We're off to the races. Ten. Thank you,

[01:06:22] Ten. Woo.

[01:06:23] >> Thank you.

[01:06:30] >> Okay, so

[01:06:33] like we said, the future of AI data data

[01:06:35] centers is all optically connected

[01:06:37] infrastructure. And you heard him say

[01:06:39] it, right? This is going to drive a

[01:06:40] tidal wave of growth innovation that's

[01:06:42] needed in scale and in manufacturing.

[01:06:45] But what does that inevitable future

[01:06:48] actually look like? I mean, if you just

[01:06:49] take a step back for a minute and you

[01:06:51] actually don't think about right now,

[01:06:52] think about 10 years in the future and

[01:06:54] it's a world where a lot of the copper

[01:06:56] connections are gone and just think

[01:06:58] about a world where data transmission

[01:07:00] now at some point is all optical.

[01:07:04] This is a world where then distance

[01:07:05] doesn't matter matter actually and

[01:07:07] that's that's a profound change.

[01:07:10] Servers, racks, and overall data center

[01:07:12] architectures today have all been

[01:07:14] designed around the constraints of

[01:07:16] distance. And software workloads

[01:07:18] actually have also been optimized around

[01:07:20] those same constraints. But what if

[01:07:22] distance no longer matters? How might

[01:07:24] the architecture itself change? And what

[01:07:26] new capabilities become possible when

[01:07:28] the infrastructure is no longer

[01:07:30] constrained by distance?

[01:07:33] So let's start with the scaleup network

[01:07:34] in the rack. As we discussed earlier,

[01:07:37] this is where we can connect the largest

[01:07:38] possible number of processors together

[01:07:40] in a full any to any configuration. Now,

[01:07:43] in the past, the size of this domain was

[01:07:45] limited by the length of the copper

[01:07:46] connection. But with optics, distance

[01:07:48] doesn't matter. So now we can change the

[01:07:50] size of the scaleup domain from 72 or

[01:07:54] 144 XPUs or GPUs to a thousand or more

[01:07:58] all optically interconnected. The

[01:08:00] implications for workloads are enormous.

[01:08:03] Today AI workloads must be broken down

[01:08:06] into smaller subpros that fit within the

[01:08:08] scaleup cluster because communicating

[01:08:11] outside the cluster today is slower,

[01:08:13] much lower bandwidth.

[01:08:15] But optically interconnected systems can

[01:08:17] manage workloads on an order of

[01:08:19] magnitude larger. And it does not stop

[01:08:21] there. By the way, what happens when the

[01:08:24] optical connectivity comes inside the

[01:08:25] server itself? Modern AI servers are

[01:08:28] composed of a certain number of CPUs,

[01:08:30] XPUs, memory, and network interfaces.

[01:08:33] And the reason they're all in the same

[01:08:34] system is because of distance.

[01:08:37] CPUs and XPUs need to access memory at

[01:08:40] very very high bandwidth, which means

[01:08:41] they need to sit right next to each

[01:08:43] other on the board with copper traces

[01:08:45] serving as the connections between them.

[01:08:47] But in a future where these connections

[01:08:48] are all optical, distance actually

[01:08:51] doesn't matter. You can imagine a

[01:08:53] completely disagregated architecture.

[01:08:56] XPUs in one system, memory in another,

[01:09:00] gentic CPUs in another, which unlocks

[01:09:03] another possibility. In today's systems,

[01:09:05] the ratio of CPU and XPU or GPU, it's

[01:09:09] fixed. So these ratios have to be

[01:09:12] defined at the time the system is built

[01:09:14] and deployed. But no two workloads

[01:09:16] require exactly the same ratio. Jensen

[01:09:18] talked about this actually, which means

[01:09:20] at any given time some portion of the

[01:09:22] computer memory could be underutilized

[01:09:24] for a given workload that costs money.

[01:09:27] But once we decompose the system into

[01:09:29] separate pools of compute and memory and

[01:09:31] they're all op optically interconnected,

[01:09:34] we can then decompose dedicated systems

[01:09:36] on the fly which are then optimized for

[01:09:38] whatever the workload is.

[01:09:40] So imagine future data centers, a

[01:09:43] globally opticated, sorry, a globally

[01:09:46] optically interconnected data

[01:09:48] infrastructure.

[01:09:50] These rigid boundaries we have today in

[01:09:52] the systems we have, they begin to

[01:09:53] disappear. Compute can now be pulled.

[01:09:56] Memory can be pulled and infrastructure

[01:09:59] can be composed dynamically at scale.

[01:10:03] For the first time, architects can begin

[01:10:05] designing AI systems around the needs of

[01:10:07] the model, not around the limits of the

[01:10:11] interconnect.

[01:10:14] So, this is where AI infrastructure is

[01:10:15] headed. It's a data center without

[01:10:18] distance where compute, memory,

[01:10:20] networking, and photonics operate as one

[01:10:23] unified system

[01:10:26] where millions of resources across the

[01:10:27] data center can work together as if they

[01:10:29] were one machine. an architecture

[01:10:32] defined by the needs of the workload,

[01:10:35] not the limits of the connectivity.

[01:10:38] We believe this is the next era of

[01:10:40] computing infrastructure and Marll is

[01:10:43] helping build the connectivity

[01:10:44] foundation that will make all this

[01:10:46] possible.

[01:10:48] Thank you very much for your time today.

[01:11:21] Heat up here.

[01:11:24] Heat. Heat.