# 【直播｜AI即時中字翻譯】Intel 英特爾 CEO 陳立武 COMPUTEX 2026主題演講

https://www.youtube.com/watch?v=wvKpcag9LA4
Translation: zh-TW

[00:01] Silicon, the foundation of modern technology.
  矽，現代科技的基石。

[00:07] Every transistor placed with purpose.
  每一個電晶體都帶著目的被放置。

[00:09] Every watt wrestled from physics.
  每一瓦特都從物理學中爭取而來。

[00:11] Where every instruction set earns its right to execute.
  指令集得以執行的權利在此獲得。

[00:17] This is how performance gets driven, how efficiency gets built, how intelligence acts, not just answers.
  這就是效能如何被驅動，效率如何被建立，智慧如何行動，而不僅僅是回答。

[00:26] Our future is with ecosystems shaped by architecture capable of connecting technologies, tools, and partners.
  我們的未來在於由架構塑造的生態系統，這些架構能夠連接技術、工具和合作夥伴。

[00:36] Powerful enough to execute, efficient enough to scale, familiar enough to build with speed.
  強大到足以執行，高效到足以擴展，熟悉到足以快速建構。

[00:45] Tomorrow, it's about the progress we scale through open platforms, shared standards, and partnerships amplifying each other's strengths.
  明天，在於我們透過開放平台、共享標準和互相增強優勢的合作夥伴關係來擴展的進步。

[00:54] A unified architecture engineered for systems.
  一個為系統而設計的統一架構。

[00:57] The next chapter is being written at Silicon with our engineering at its heart.
  下一章正在矽谷書寫，以我們的工程為核心。

[00:58] Built different, built together,
  獨特打造，共同建構，

[01:03] Built on Intel.
  建立在英特爾之上。

[01:06] And now, ladies and gentlemen, please give a warm welcome to the CEO of Intel, Li Pan.
  現在，各位女士們先生們，請熱烈歡迎英特爾的首席執行官李潘。

[01:29] Intel computer steps.
  英特爾電腦的步驟。

[01:44] Uh this is the elephant mountain and so 1,00 step to 184 meter and so I survived and then came down with one piece.
  呃，這是象山，所以走了 1000 步到 184 公尺，所以我倖存下來，然後完整地下來了。

[01:58] So I'm here.
  所以我來了。

[02:00] So but if I walk a little bit slower then you know uh I'm exhausted.
  所以，但如果我走得慢一點，那麼你知道，呃，我會筋疲力盡。

[02:07] But anyway it's a beautiful view.
  但總之這是一個美麗的景色。

[02:10] I mean highly recommend all of you to do that.
  我的意思是強烈推薦你們所有人都這麼做。

[02:13] Yep.
  對。

[02:16] I think first of all I think uh delighted to be here and uh this is an important event and then I'd like to uh get started and about nearly six decades ago a group of brilliant and highly motivated engineers and venture capitalists uh including Artur Rock, Don Valentine and many others.
  我想首先，我想很高興來到這裡，這是一個重要的活動，然後我想開始，大約在將近六十年前，一群才華橫溢、積極進取的工程師和風險投資家，包括阿特爾·洛克、唐·瓦倫丁以及許多其他人。

[02:41] It found companies like Intel, Apple and others broadly set the motion the largest economic known to mankind uh create what became known as Silicon Valley and uh this is a very exciting time and that is a very same ambition and mindset with the semiconductors make
  它創立了像英特爾、蘋果等公司，廣泛地推動了人類已知最大的經濟發展，創造了後來被稱為矽谷的東西，這是一個非常令人興奮的時代，這與半導體製造的雄心和思維方式非常相似。

[03:08] across the ocean.
  跨越了海洋。

[03:11] that sparked the creation of Silicon Island right here in Taiwan.
  這激發了矽島在台灣的誕生。

[03:18] And I have been very fortunate to associate with the creation of semiconductor industry in Taiwan 40 years ago.
  我很幸運能參與台灣半導體產業40年前的創立。

[03:26] uh because uh 40 years ago uh Minister KT Lee Leorting uh invite me to lay the foundation of venture capital concept in Taiwan.
  因為40年前，李國鼎部長邀請我為台灣的創投概念奠定基礎。

[03:35] It's a very new concept and uh you know you people put money with you and you play the money and then you share the profit of 20%.
  這是一個非常新的概念，你們把錢交給我，我來運用這筆錢，然後分享20%的利潤。

[03:48] But you don't share 20% of the losses.
  但你們不分擔20%的虧損。

[03:51] So this is a very unique uh venture capital concept but I managed to do that and then with the help of Minister KT Lee and then development fund and I set up my venture fund.
  所以這是一個非常獨特的創投概念，但我成功做到了，並在李國鼎部長和發展基金的協助下，成立了我的創投基金。

[04:03] and about the same time you know uh uh
  大約在同一時間，你知道嗎？

[04:10] Maurice Chang uh from TI came back to E3 and then uh then set up the TSMC.
  Maurice Chang，來自德州儀器的，回到了E3，然後，然後創立了台積電。

[04:17] So it's a very exciting time that I'm being involved in this whole science park and uh Shinzu science park and the foundation of become the silicon island.
  所以，我參與了整個科學園區和新竹科學園區，以及成為矽島的基礎，這是一個非常令人興奮的時刻。

[04:27] And so I really see the benefit of you know uh from OEM ODM from design to manufacturing it's all here in Taiwan.
  所以我真的看到了，你知道的，從OEM、ODM到設計到製造的好處，這一切都在台灣。

[04:40] Taiwan PC ecosystem has played a critical role uh in Intel growth and success.
  台灣的PC生態系統在英特爾的成長和成功中扮演了關鍵角色。

[04:46] I in fact last year I was here uh to celebrate Intel 40th anniversary in Taiwan.
  事實上，去年我曾在這裡慶祝英特爾在台灣的40週年紀念。

[04:53] I want to thank all of the suppliers, partners and customers for 40 years of partnership with Intel.
  我要感謝所有供應商、合作夥伴和客戶與英特爾40年的合作夥伴關係。

[05:04] As partnership continue to grow.
  隨著合作夥伴關係的持續成長。

[05:12] Thank you.
  謝謝你。

[05:14] As partnership continue to grow and uh stronger every year.
  隨著合作夥伴關係持續成長，並且 uh 每年都更加穩固。

[05:25] It had been the year since I stepped in uh the role of Intel CEO to be more precise 14 months uh to being CEO of Intel and it may be the first CEO can speak Mandarin.
  自從我擔任英特爾執行長的職位以來已經一年了，更精確地說，是 14 個月 uh 擔任英特爾執行長，而且我可能是第一位會說中文的執行長。

[05:43] [applause]
  [掌聲]

[05:46] and in fact couple of uh customer of Taiwan very important partner for us is that Lipo also.
  事實上，台灣的幾位 uh 客戶，對我們來說是非常重要的合作夥伴，也是立寶。

[05:53] So very unusual if a CEO can drink liquor with us.
  所以，如果一位執行長能和我們一起喝酒，那就很不尋常了。

[06:01] but anyway it's I'm part of this community uh execution has always always at the top of my list to do so we had to bring focus back to the core uh at our
  但總之，我屬於這個社群 uh 執行始終始終是我最重要的任務之一，所以我們必須將重心拉回到核心 uh 在我們的

[06:15] Intel is an engineering company and that's what I decided from day one.
  英特爾是一家工程公司，這是我從第一天就決定的。

[06:21] I came to become CEO of Intel, I have all the engineering report to me.
  我來擔任英特爾的首席執行官，所有的工程報告都交給我。

[06:26] So they're understanding to really drive the engineering, drive success in engineering, uh performance.
  所以他們理解要真正推動工程，推動工程的成功，呃，性能。

[06:31] Our customer and partners are always uh you know already seen a shift uh in the Intel showed up.
  我們的客戶和合作夥伴總是，呃，你知道，已經看到了英特爾出現的轉變。

[06:43] We are just getting started and so stay tuned.
  我們才剛剛開始，所以請繼續關注。

[06:46] We have a journey in front of us.
  我們前方有一段旅程。

[06:48] Opportunity ahead is enormous and our job is to stay focused, execute and deliver.
  前方的機會是巨大的，我們的任務是保持專注，執行並交付。

[06:59] Every year, Intel ship hundred millions of uh hundreds and millions of uh SOC, orchestrating silicon across every industry, working tightly with our partners ecosystem across each.
  每年，英特爾都會出貨數億個，呃，數億個，呃，系統單晶片，在各個行業協調矽晶片，與我們的合作夥伴生態系統緊密合作。

[07:18] layer of the stack from silicon to SOC to system and to software.
  堆疊的一層，從矽到 SOC 到系統再到軟體。

[07:29] This generates trillions of dollars in value across four core compute ecosystem.
  這在四個核心計算生態系統中產生了數兆美元的價值。

[07:35] First personal computers, second edge agentic AI and later physical AI, third foundational data centers and finally emerging intelligence centers that will power digital agents of the future.
  首先是個人電腦，其次是邊緣自主 AI 和後來的實體 AI，第三是基礎數據中心，最後是將為未來數位代理提供動力的新興智慧中心。

[08:01] Each of these ecosystem represent generational opportunity and increasingly each of these will need purpose-built CPU, GPU and A6 solutions cers for specific workloads and
  這些生態系統中的每一個都代表著一代人的機會，而且這些中的每一個都將越來越需要專為特定工作負載和...

[08:21] application. The silicon we are building now will be for human use and the digital agent use.
  應用程式。我們現在正在建構的矽晶片將用於人類用途和數位代理人用途。

[08:33] Let us begin with the ecosystem that started all to talk more about the PC ecosystem.
  讓我們從開始談論個人電腦生態系統的生態系統開始。

[08:40] Please join me to welcome to stage our new leaders for client compute and physical AI, Alex.
  請加入我，歡迎我們客戶端運算和實體人工智慧的新領導者 Alex 上台。

[08:59] Thank you. Thank you. Thank you.
  謝謝。謝謝。謝謝。

[09:01] Thank you, Lupu.
  謝謝你，Lupu。

[09:03] Little story about me.
  關於我的一個小故事。

[09:05] The first time I came to Taiwan was the year 1990 and I was fresh out of school.
  我第一次來台灣是 1990 年，那時我剛畢業。

[09:11] first job, first international trip, first East Asia country I have ever been to and I came right here in Taipei.
  第一份工作、第一次國際旅行、我曾經去過的第一個東亞國家，而我就是來到台北。

[09:21] And I quickly realized even back then that, you know, this the the desire to grow, uh, the mindset of win-win and cooperation is all over all of our customers and our partners here in Taiwan.
  我很快就意識到，即使在當時，你知道，這種成長的渴望、雙贏和合作的心態，在我們在台灣的所有客戶和合作夥伴身上都隨處可見。

[09:37] And so it quickly became uh one of my favorite countries to visit and work.
  因此，它很快就成為我最喜歡的旅遊和工作國家之一。

[09:43] And now forward the clock 36 years, I'm here in my first international trip with all the great people here at Intel.
  現在把時間往前推36年，我來到這裡，這是我第一次與英特爾的各位優秀人士進行國際旅行。

[09:51] And where do I come? Right here to Taipei.
  我來自哪裡？就在這裡，台北。

[09:53] Thank you.
  謝謝。

[10:01] I can't wait to plan and build a future with you guys.
  我迫不及待地想和大家一起規劃和建立未來。

[10:03] So let's get started.
  那麼，我們開始吧。

[10:05] We have a lot to cover.
  我們有很多內容要講。

[10:08] Intel has continuously increased the pace of progress across all PC segments.
  英特爾在所有個人電腦領域不斷加快進步的步伐。

[10:14] Workstations, desktops, creators, gamers, premium and mainstream laptops.
  工作站、桌上型電腦、創作者、遊戲玩家、高階和主流筆記型電腦。

[10:22] Every major segment, every major segment is driven by an Intel system solution.
  每一個主要領域，每一個主要領域都由英特爾系統解決方案驅動。

[10:31] With that vast coverage in mind, we're adding another dimension to scale these products even more effectively.
  考慮到如此廣泛的覆蓋範圍，我們正在增加另一個維度，以更有效地擴展這些產品。

[10:41] The Intel 18A process is now at full scale.
  英特爾 18A 製程現已全面投入生產。

[10:46] We have a full lineup of products with hundreds of design wins to prove that.
  我們擁有一系列完整產品，並有數百項設計獲勝來證明這一點。

[10:52] At CES, we launched the Core Ultra Series 3, Intel's first product built on 18A process technology.
  在 CES 上，我們推出了 Core Ultra 系列 3，這是英特爾首款基於 18A 製程技術的產品。

[11:02] It's setting a new standard for premium mobile performance and battery life.
  它為頂級移動性能和電池壽命設定了新標準。

[11:10] The Ultra Series 3 enables great user experiences across any tasks with a very fast response CPU, a highly improved GPU,
  Ultra 系列 3 憑藉極快的響應 CPU、大幅改進的 GPU，可在任何任務中實現出色的用戶體驗，

[11:22] low power processing NPU, and the latest multimedia capabilities.
  低功耗處理NPU，以及最新的多媒體功能。

[11:27] It's a perfect blend of IP performance and power for any AI and agentic experience.
  它是IP性能和功耗的完美結合，適用於任何AI和代理體驗。

[11:36] It's allowing us to lead the way to transform every PC, every PC to an agentic capable platform.
  它使我們能夠引領潮流，將每一台PC，每一台PC轉變為具備代理能力的平台。

[11:48] Today, more than 300 designs are shipping across consumer and commercial segments, over 300.
  今天，有超過300種設計正在消費和商業領域出貨，超過300種。

[11:58] And to scale these capabilities even further, we've taken the latest core ultra IPs and specifically tailored them for the mainstream market.
  為了進一步擴展這些功能，我們採用了最新的核心超IP，並專門為主流市場量身定制。

[12:12] The result, the Intel Core Series 3 introduced in April.
  結果是，於四月推出的Intel Core Series 3。

[12:16] Let me repeat, we just introduced this in April and it's already scaled up to 70 plus designs.
  讓我重複一遍，我們剛在四月推出這個，它已經擴展到70多種設計。

[12:25] That brings. Thank you.
  這帶來了。謝謝您。

[12:29] Thank you.
  謝謝您。

[12:30] That brings the total series lineup to nearly 400 designs in just a few short months.
  這使得系列產品陣容總計在短短幾個月內就達到了近400種設計。

[12:36] Now, that is massive scale.
  現在，這是一個龐大的規模。

[12:42] Let's look at some of the capabilities of the Core Series 3.
  讓我們來看看Core Series 3的一些功能。

[12:47] And we can start with battery life.
  我們可以從電池壽命開始。

[12:49] You know, I can go over these numbers here that are printed and I can talk about how we measured them and things like that, but I I have a question for you guys.
  您知道，我可以查看這裡列印的這些數字，我可以談談我們是如何測量它們的等等，但我有一個問題要問各位。

[12:55] How long is your day?
  您的一天有多長？

[12:57] 10 hours, 12 hours, 14 hours, or more like us at Intel?
  10小時，12小時，14小時，還是像我們英特爾一樣更長？

[13:08] But the great user experience is if your PC lasts longer than your workday.
  但絕佳的使用者體驗是，如果您的PC比您的工作日還持久。

[13:13] And that's exactly what we're delivering in all segments.
  這正是我們正在所有細分市場中提供的。

[13:17] And we support ample number of ports for all of your connectivity needs, unlike some of our competitors who only have
  並且我們支援足夠多的連接埠以滿足您所有的連接需求，不像我們的一些競爭對手，他們只有

[13:26] one USBC interface.
  一個USB-C接口。

[13:30] But I'll let you be the judge of that one.
  但我會讓你來評判這一點。

[13:36] The goal of this uh uh Core Series 3 is to bring premium feel and experiences to incredibly thin form factors for mainstream PCs.
  這個呃呃Core Series 3的目標是為主流PC帶來極致纖薄外形中的高級感和體驗。

[13:46] And you know, you don't have to take my word for it.
  而且你知道，你不必相信我說的話。

[13:48] You can look at this wall of incredibly designed, sleek PCs that are here, and all of it is thanks to you, our partners, and our customers.
  你可以看看這面牆上這些設計極佳、造型時尚的PC，這一切都要歸功於你們，我們的合作夥伴和客戶。

[13:56] We couldn't have done it without you.
  沒有你們，我們不可能做到。

[13:59] So, please, a round of applause.
  所以，請大家給予掌聲。

[14:06] Isn't it awesome?
  是不是很棒？

[14:06] Super light, super thin, really great.
  超級輕，超級薄，真的很棒。

[14:12] Okay.
  好的。

[14:12] Now, the next proof is scaling 18A IP into growing markets.
  現在，下一個證明是將18A IP擴展到增長中的市場。

[14:18] And the fastest growing portion of the PC market is the handheld gaming.
  而PC市場中增長最快的部分是掌上遊戲機。

[14:21] Let's take a look.
  讓我們來看看。

[14:35] Heat.
  熱。

[15:02] Heat.
  熱。

[15:16] Heat. Heat.
  熱。熱。

[15:36] This is the Arc G3.
  這是 Arc G3。

[15:41] I think a beautiful chip.
  我認為它是一款漂亮的晶片。

[15:44] More beautiful than what was presented yesterday at a keynote.
  比昨天在主題演講中所展示的更漂亮。

[15:51] Okay.
  好的。

[15:54] The G3 is derived from the Core Ultra Series 3 and the ARC G3 is a tuned higher-performance GPU specifically for handheld gaming and it's providing great battery life.
  G3 源自 Core Ultra Series 3，ARC G3 是一款專為手持遊戲調校的高效能 GPU，並提供出色的電池續航力。

[16:07] The performance tested across multiple games is consistent and stable versus competition.
  跨多款遊戲測試的效能與競爭對手相比，一致且穩定。

[16:14] We are more than 40% faster, 40% faster, and at the same performance, we're half the power.
  我們的速度快了 40% 以上，快了 40%，在相同的效能下，我們耗電量減半。

[16:23] And on top of that, we're running all AAA games at 1080p resolution, many of them above 120 frames per second.
  此外，我們還能在 1080p 解析度下運行所有 AAA 遊戲，其中許多遊戲的幀率都超過每秒 120 幀。

[16:32] Now, that is giving gamers a great user experience.
  現在，這為遊戲玩家提供了絕佳的使用者體驗。

[16:37] All of these devices will be available later this month and is just the beginning.
  所有這些設備將於本月晚些時候上市，而這僅僅是個開始。

[16:42] We're going to have plenty more designs coming throughout the year.
  我們將在今年推出更多設計。

[16:50] Thank you.
  謝謝。

[16:50] Thank you.
  謝謝。

[16:54] Okay, it is indeed true that Intel has a leading lineup of processors and with the ver versatility of the ATNA process technology, our newest offerings, we're bringing powerful performance and efficiency to scale across the breadth of premium mainstream and handheld gaming segments.
  好的，英特爾確實擁有領先的處理器陣容，並且憑藉 ATNA 製程技術的多功能性，我們最新的產品將強大的性能和效率帶入高端主流和手持遊戲領域。

[17:17] These same fundamentals, the same IP, the same capabilities can deliver far beyond the PC ecosystem.
  這些相同的基本原理、相同的知識產權、相同的功能可以超越個人電腦生態系統。

[17:27] The demand for our processors at the edge has been booming.
  我們邊緣處理器的需求一直在蓬勃發展。

[17:34] As you've seen, we've already taken existing product
  正如你所見，我們已經採用了現有的產品

[17:38] lines and pivoted them into adjacent markets, enabling our customers to grow their businesses.
  線並將其轉移到鄰近的市場，使我們的客戶能夠發展他們的業務。

[17:47] Now, the edge is demanding the latest products from Intel.
  現在，邊緣正在要求英特爾提供最新的產品。

[17:49] And that's why this year we're taking our latest series 3 products into the edge business with over 130 designs in multiple verticals.
  這就是為什麼今年我們將最新的第三系列產品帶入邊緣業務，在多個垂直領域擁有超過 130 種設計。

[18:02] For largecale edge business, our customers need the best technology and chipsets, easy to use reference designs, and appropriate software stacks.
  對於大規模的邊緣業務，我們的客戶需要最好的技術和晶片組、易於使用的參考設計以及適當的軟體堆疊。

[18:14] And at Intel, we've done all of that.
  而在英特爾，我們已經完成了這一切。

[18:19] We have over 4,000 edge ecosystem partners deploying into such verticals such as manufacturing, robotics, retail, and many more.
  我們擁有超過 4,000 家邊緣生態系統合作夥伴，部署到製造、機器人、零售等垂直領域，以及更多領域。

[18:29] For those of you here at Computex, you can see some of that at the pavilion.
  對於在 Computex 現場的各位，您可以在展館看到其中一些。

[18:38] Given that capability, given the IP,
  考慮到該能力，考慮到 IP，

[18:40] Given the chipsets, given the scale that we have, there's a massive opportunity ahead of us across many segments of physical AI.
  鑑於晶片組和我們的規模，我們在實體人工智慧的許多領域都有巨大的機會。

[18:49] It's projected to be a 25 trillion market by 2050, and it will leverage all of our scale in the PC ecosystem.
  預計到 2050 年將成為一個 25 兆美元的市場，並將利用我們在 PC 生態系統中的所有規模。

[19:01] Physical AI form factors will take shape across key industries, as you see behind me.
  實體人工智慧的形態將在關鍵行業中成形，正如你們在我身後所見。

[19:09] We will continue growing these markets with the same strategy of leading IP and chipsets, complete reference platforms of enduser hardware and applicable software stacks enabling our customers to expand into new physical AI form factors and applications.
  我們將繼續以領先的知識產權和晶片組、完整的終端用戶硬體參考平台以及適用的軟體堆疊的相同策略來發展這些市場，使我們的客戶能夠擴展到新的實體人工智慧形態和應用。

[19:28] And indeed, this will be our future.
  確實，這將是我們的未來。

[19:32] Now back to you, Lipu. Thank you.
  現在回到你，Lipu。謝謝。

[19:41] Thank you.
  謝謝你。

[19:41] Thank you.
  謝謝你。

[19:41] Thank you.
  謝謝你。

[19:46] Uh, thanks, Alex.
  呃，謝謝你，Alex。

[19:48] AI is profoundly impacting the way we use our devices.
  人工智能正在深刻地影響我們使用設備的方式。

[19:54] A major focus area for us is the use of AI on device.
  我們的一個主要關注領域是在設備上使用人工智能。

[20:01] Together with partners, we are at the forefront of advancing intelligence.
  與合作夥伴一起，我們走在推動智能發展的前沿。

[20:07] To tell you more about it, let me welcome on stage my close friend and founder CEO of Perplexity, Aravvin.
  為了向您介紹更多關於它的信息，請允許我歡迎我的摯友兼 Perplexity 的創始人兼首席執行官 Aravvin 上台。

[20:26] Well, Aravin, welcome.
  嗯，Aravin，歡迎你。

[20:26] You and I have been talking about hybrid compute for a while.
  你和我已經談論混合計算有一段時間了。

[20:33] The reason why are clear you know the privacy cost performance and let's talk about how to make this
  原因很清楚，你知道隱私、成本、性能，讓我們來談談如何實現這一點

[20:44] work. Yeah. So in February we launched

[20:48] perplexity computer. Computer is an AI

[20:51] operating system. It creates a team of

[20:55] agents,

[20:56] uses up to 20 different AI models, and

[21:00] it orchestrates across models, tools,

[21:04] and files in one single system.

[21:07] The agent harness inside computer is

[21:11] model agnostic.

[21:13] Perfectly balancing intelligence,

[21:16] accuracy, privacy, and cost is the

[21:20] orchestration problem it solves. And so

[21:23] this allows you to run smaller models

[21:27] locally on the Intel Core Ultra Series 3

[21:31] GPU. And so for the first time ever, we

[21:35] work together to create hybrid agenic

[21:39] inference. And so what we are showing

[21:42] today is just the start. Hybrid agentic

[21:46] inference is how we maximize token value

[21:51] per watt per user.

[21:55] >> So should we show them how it works?

[21:58] >> Yep.

[22:03] >> All right. So here it is.

[22:06] Let's say I'm an associate at a private

[22:09] equity firm and um I'm working on

[22:12] something that has a confidential

[22:14] project code name project falcon. Here's

[22:17] the query.

[22:21] So, think of it as me trying to

[22:24] understand if a certain private company

[22:28] is worth $1.1 billion and I'm feeding it

[22:32] confidential deal materials.

[22:35] The work begins on the laptop. It sees

[22:39] that project falcon has private dealroom

[22:43] files

[22:45] and an NDA, a local leverage buyout

[22:48] financial model, a whiteboard diagram

[22:52] and bilingual transcripts that are very

[22:54] confidential. You don't want these

[22:56] materials to be shipped to the server.

[22:58] So what the local model does on the core

[23:02] ultra series 3 is it first decides this

[23:05] is all very important work and shouldn't

[23:08] be sent to the server. It reads the

[23:10] files classifies what is sensitive and

[23:12] what is not and then computer decides

[23:15] what should leave the device and what

[23:17] shouldn't and each of these things is

[23:20] done with local AI.

[23:24] The orchestrator can spin up additional

[23:26] agents as necessary.

[23:29] And so if you need a research agent to

[23:32] bring in outside file materials against

[23:36] local model without exposing any private

[23:39] files, that's what you want in the

[23:41] hybrid system. And so computer arc acts

[23:44] as one single system, brings all inputs

[23:47] and outputs together. And so let's

[23:49] actually skip and see what the actual

[23:52] result would be.

[23:59] All right. So the result is a document,

[24:01] a research report and sporting data. And

[24:04] it's being created by agents on large

[24:08] cloud-based models keeping your

[24:10] sensitive information only on your

[24:13] device. And so all your local device

[24:16] models will take care of the private

[24:18] device and the server side models will

[24:20] take care of other things through hybrid

[24:23] inference orchestration.

[24:24] >> This is the architecture we both believe

[24:26] in and the future is more compute in the

[24:30] data center and more compute on the

[24:33] local machine.

[24:36] And so I think of this as a big

[24:39] milestone for engineering on both the

[24:42] agent harness AI side as well as the

[24:45] chip side. And so um it's been really

[24:48] fun to partner with you and Intel on

[24:50] this. So thank you so much Libu.

[24:52] >> Definitely. Thank you Aravvin and

[24:54] looking forward to continue partnership.

[24:56] Thank you. [applause]

[25:02] We talk a lot about the exciting new

[25:04] developments in the PC, edge and

[25:08] physical AI space. I want to take a few

[25:11] minutes to talk about the foundational

[25:13] IP that power all these advancements.

[25:18] Let us talk about x86.

[25:23] When most people think of generalpurpose

[25:26] computing, they think x86

[25:30] and that is a good reason for that. X86

[25:34] is architecture that has power data

[25:38] center for nearly five decades

[25:44] and the leadership continues.

[25:47] uh according to the IDC expect eight out

[25:50] of the 10 servers installed through 2030

[25:56] to be x86based

[25:59] powering modern computing from

[26:01] foundational to emerging intelligent use

[26:05] cases.

[26:08] Intel pioneer most of the breakthrough

[26:11] architectural innovation

[26:14] that have enhanced 886 over the last

[26:18] four decades starting with the 8086

[26:23] that become the foundation of modern

[26:26] computing.

[26:27] Uh if you can see the chart today we

[26:30] have two flagship CPU cores

[26:33] PC and ECores.

[26:36] One optimized for performance,

[26:39] one the others is for efficiency.

[26:45] These are Intel most advanced CPU cores

[26:49] with the accelerator building

[26:52] built in spec uh specularly for

[26:55] foundational workloads like security.

[26:59] Our x86 coursees power our PC client

[27:04] edge portfolio and also power our data

[27:07] center and AI portfolio.

[27:10] Under my leadership, we are committed to

[27:14] building the best CPU cores in the world

[27:18] and we will enhance ensure that the most

[27:22] compute intensive workload run best on

[27:26] 886 x86.

[27:29] Next, let us talk. Thank you.

[27:35] Now let us talk about how x86 is

[27:39] enabling foundational data centers.

[27:43] To tell you more about it, let me invite

[27:46] on to the stage Kavoke.

[27:53] [applause]

[27:56] Thank you. Thank you. Thank you, Lipu.

[28:00] Wow. It's so great to be here. uh

[28:03] specifically in this point in our

[28:06] history, global history, collective

[28:08] history and be at Computex with a blue

[28:12] badge. So I'm very happy, I'm very

[28:13] humbled to be here to share with you

[28:15] some of the innovations that we have. So

[28:18] uh let's [snorts] talk a bit about uh

[28:21] see what this AI thing is about. So when

[28:24] we say foundational,

[28:26] we mean the workloads that keep the

[28:28] world running. So currently we have data

[28:31] centers and there's a number of items

[28:34] and workloads and entities that run on

[28:36] these data centers. So uh for example we

[28:40] have 5G networks that uh keep us

[28:43] connected. We have databases that keep

[28:46] our data safe. We have cloud services

[28:49] that power our daily lives. So we expect

[28:52] and demand for these workloads to to

[28:55] grow in size and capacity between uh now

[28:58] and 2030 from 80 gawatt to about 100

[29:02] gawatt and uh yeah most of you involved

[29:05] in this uh domain understand the the the

[29:08] the extent of this type of an expansion.

[29:12] These workloads are broad. They are

[29:14] mission critical. So attention special

[29:17] attention has to be taken uh when

[29:19] running them but also they require

[29:22] performance, efficiency, security and

[29:25] resiliency. And we can't emphasize

[29:27] enough uh all these four factors.

[29:31] That is why we are excited to have Intel

[29:33] Xeon 6 Plus introduced at Computex this

[29:37] week.

[29:39] It has 288 eores, a massive 576 mgabyte

[29:45] of L3 cache

[29:47] built with our Intel 18A technology. And

[29:50] we can't emphasize enough the the value

[29:53] of uh Intel technology that brings to

[29:56] data center products. But most

[29:58] importantly, it delivers

[30:01] efficiency and density which enables our

[30:04] partners to save uh very precious real

[30:07] estate, have more compact servers and

[30:11] the racks.

[30:14] So this is a leadership compute for the

[30:16] next era of cloud and network

[30:19] infrastructure.

[30:23] So Zeon 6 Plus launches with the

[30:25] strength of our ecosystem that's been

[30:27] built over decades and decades of data

[30:29] center development both from a hardware

[30:33] but also from a software and

[30:34] infrastructure perspective.

[30:37] Moreover, our ODM partners are bringing

[30:40] Zeon 6 plus solutions to the market

[30:42] today.

[30:44] So these range from full rack scale

[30:47] deployments to server level designs.

[30:52] Xeon 6 Plus joins [sighs]

[30:55] our lineup of data center processors

[30:58] next to our already launch Xeon 6 based

[31:01] on peores.

[31:03] Both of these uh category and class of

[31:06] uh solutions delivers new performance

[31:08] and choice for all the enterprises whose

[31:11] infrastructure backbone is built on x86

[31:15] and zeon.

[31:17] This is critical for enterprises that

[31:19] need to increasingly balance preparing

[31:22] for AI workloads but at the same time

[31:26] running their day-to-day mission

[31:27] critical applications.

[31:31] So let's switch gears and talk about how

[31:34] Intel was certifying the deployment of

[31:37] intelligence at scale.

[31:40] It's undeniable that enterprise

[31:41] infrastructure today will have to evolve

[31:44] to keep up with the AI demand.

[31:47] Recent research forecasts that AI

[31:49] inference workloads are expected to

[31:52] become 40% of all data center power

[31:55] demand and much more than uh they are

[31:57] today.

[32:00] So we have these two paradigms where we

[32:03] have the foundational data centers keep

[32:05] on running their traditional workloads

[32:07] but at the same time they have to figure

[32:09] out ways of building their

[32:10] infrastructures to serve intelligence at

[32:13] scale and this is where Intel and Xeon 6

[32:16] plus come in.

[32:24] Now up to now training split the data

[32:27] center into two. So on one hand we have

[32:30] CPUled enterprise infrastructure

[32:34] and the other hand we have GPU heavy AI

[32:36] factories and that was very clear divide

[32:39] for a while right and we've all been

[32:41] accustomed to that uh that reality

[32:44] but as agentic AI moves into real

[32:46] workflows data tools governance the

[32:49] needs change the next wave is not just

[32:52] about training models it is about

[32:54] putting AI to work so let's look at why

[32:58] Agentic AI changes the infrastructure

[33:00] equation.

[33:03] The way AI inference works is

[33:04] straightforward. We take a prompt, it

[33:08] gets fed into an LLM where it spends

[33:10] most time reasoning about the prompt. I

[33:12] we've all seen this. We've done this

[33:13] thousands of times. And out comes an

[33:16] answer.

[33:17] In this case, a lot of time is spent

[33:19] computing the large language model which

[33:21] is mostly GPU and compute intensive.

[33:28] Now the way agentic AI works is

[33:29] radically different. It's given goals

[33:33] rather than prompts. So we all seen the

[33:36] uh the different types of loop that

[33:38] people are running on this agentic AI.

[33:40] It's also very iterative in nature but

[33:42] also prompted by automation

[33:45] and thinking, planning, acting and

[33:47] reflecting are a natural way of these

[33:50] agents interacting with us.

[33:53] as it works, it uses tools, reads and

[33:56] writes files, checks rules and other

[33:59] aspect that were, you know, in the

[34:00] traditional realm of CPUs and x86.

[34:05] And then for each step, the type of

[34:07] underlying compute needs is very

[34:10] different and we'll show that in a bit.

[34:14] This is particularly important as agents

[34:16] scale up their work spawning new agents

[34:19] that work concurrently and the category

[34:22] and the complexity of agents are going

[34:24] to be very different depending on the

[34:26] complexity of the work.

[34:29] That's the main reason that there's such

[34:31] a rapid increase in CPU demand for

[34:34] Aentic AI. The CPU orchestrates the

[34:37] show.

[34:39] Now what we're seeing is we're also

[34:41] seeing the balance and the ratio of uh

[34:44] one CPU to 8GPU and more is uh is coming

[34:47] much closer to parity. So let's take a

[34:50] look at the real example. John,

[34:53] >> thanks Kaborg. You talked about how

[34:55] Agentic AI is changing the compute

[34:57] requirements. Let's take a look at a

[34:58] real example. I have a traditional AI

[35:01] inference setup on the lefth hand side

[35:02] of the screen. Let's send a request.

[35:05] write a Python function that calls an

[35:07] OpenAI compatible chat completions API.

[35:10] The model gets the response, generates

[35:13] code, and sends the request back. Take a

[35:16] look at the slider on the top of the

[35:17] screen. GPU dominates nearly 7 to1 GPU

[35:21] heavy.

[35:23] In contrast, let's take a look at an

[35:25] Aentic AI system. Across the top, look

[35:28] at the pipeline stages. Green is GPU

[35:31] work, blue is CPU work. Linting is

[35:34] happening on our Xeon 6 plus processor

[35:37] with effic efficiency cores. Web fetch

[35:40] and compile is happening on our Xeon 6

[35:43] performance cores

[35:45] and unit testing is coming back and

[35:47] running on our Xeon 6 plus effic

[35:49] efficiency cores.

[35:51] The right class CPU for each stage of

[35:54] the pipeline. Take a look at the slider

[35:56] across the top again. We're near par but

[35:58] CPU heavy this time.

[36:02] What's this look like when we multiply

[36:04] that by millions of queries a day?

[36:09] As you mentioned, each Xeon 6 Plus

[36:11] processor has up to 288 cores. That's

[36:14] 576 cores per two socket server. When we

[36:20] look at that from a rack scale

[36:21] perspective,

[36:23] that gives us over 36,000 cores per 32

[36:26] years of compute space.

[36:29] >> Thank you, John. Wow, this is pretty

[36:31] amazing and some data to ponder on.

[36:34] [applause]

[36:38] By far the the density of CPU we showed

[36:42] is is has the highest density per rack

[36:45] ever. But also looking at the number of

[36:48] agents and these are the new metrics

[36:50] that are emerging, we can safely say

[36:53] that that particular rack can run up to

[36:56] 150,000 agents. So good news to all the

[37:00] CIOS in the audience. Now your very

[37:02] expensive GPUs can be can see more

[37:05] utilization because of uh our solutions.

[37:11] Now both Zeon 6 with pores and ecores

[37:15] are built on for intelligence at scale.

[37:18] There are different cores of course but

[37:20] we've seen the workloads that require

[37:22] very high performance cores pushing the

[37:24] frequencies but also there's a need for

[37:27] very high density power efficient cores.

[37:30] So we've seen all the workloads we've

[37:32] run all the analysis and we are

[37:34] delivering these solutions uh to to all

[37:36] of you.

[37:39] Now

[37:40] having said that we are working with our

[37:44] customers and partners to make sure that

[37:46] each solution is uh tailored to to your

[37:49] needs. So

[37:51] I'd like to welcome Liu back on stage to

[37:54] talk about the server and rack sale

[37:56] solutions that our partners are working

[37:58] on. Thank you.

[38:00] [applause]

[38:04] >> Thank you my friend.

[38:05] >> Thank you. Thank you.

[38:10] Uh thank you Kavoke. It is great to see

[38:14] the momentum uh in the data center.

[38:18] As we look forward to see that is for

[38:22] intelligence at scale. Discrete compute

[38:26] alone is not enough. Our customer are

[38:29] asking us to think of system level to

[38:33] help them serve real agentic workloads

[38:36] at scale. It push us to rethink how we

[38:41] deliver our compute beyond the socket

[38:45] and to the rack.

[38:48] That is why we start the initiative

[38:51] called Rex scale blueprints

[38:55] working with ecosystem partners to

[38:57] develop Rex scale blueprints built on

[39:02] open standards. So customer can rapidly

[39:06] scale their intelligent infrastructure

[39:09] with confident without proprietary

[39:12] workarounds.

[39:15] Behind me as you can see two examples of

[39:18] these blueprints. One is for agentic

[39:22] performance based on Intel Xeon 6 with

[39:26] pores.

[39:28] The other is agent density with the

[39:32] Intel Xeon 6 Plus with ECOS. We are

[39:37] working closely with our partners

[39:39] ecosystem including Foxcom sanova to

[39:44] expand our Rex scale offering. Let me

[39:47] call on stage one of our partners chief

[39:51] product officer of Foxcom Jerry Xiao to

[39:56] talk about how we partners on Rick scale

[39:59] solution.

[40:03] >> Thank you. [applause]

[40:05] Thank you, Jerry.

[40:08] >> Um, I'm so excited to be here today.

[40:11] Wonderful product and amazing event.

[40:17] Jerry, Intel and Foxcom have been

[40:19] working together to many decades and

[40:23] Foxcom has been instrumental in driving

[40:26] technology innovation in Taiwan and

[40:29] around the world.

[40:32] >> Yeah, that's right, Abu. Um I'm proud

[40:36] the work we have done together from AI

[40:39] servers to data centers

[40:42] and to age computing all together and

[40:46] today we're excited to announce the next

[40:49] step in our partnership. Intel and

[40:53] Foxcom are working together to develop

[40:55] Rex scale products built upon Intel Xeon

[41:00] processors. Together we will focus on

[41:03] exploring the development integrations

[41:07] and commercialization of differentiated

[41:11] uh rack scale AI infrastructure solution

[41:16] leveraging comp complementaryary

[41:18] architecture to address diverse AI

[41:21] workload requirements.

[41:23] Yeah, together we will continue to

[41:26] deepen and expand our partnership

[41:29] unlocking new opportunities ahead.

[41:31] Through this collaboration, we will

[41:34] deliver system level AI solution to our

[41:37] joint customers in airballing more

[41:40] integrated and scalable computing

[41:43] environments. This makes it marks an

[41:47] important step ahead and we look forward

[41:51] unveiling more in the near future. Thank

[41:53] you Liu.

[41:55] >> Today is an exciting uh milestone for

[41:59] our continual partnership uh with

[42:02] Foxcom. Jerry, thank you for joining us.

[42:05] >> Fantastic. Thank you for having me.

[42:08] [applause]

[42:15] Thank you to the many partners in the

[42:18] audience today that is helping to bring

[42:22] this rack scale vision to life providing

[42:25] choice throughout uh through the

[42:28] ecosystem power. Uh we do not believe

[42:31] one size fit all approach for

[42:34] intelligent centers. Each enterprise

[42:38] will run unique workloads. So their

[42:42] infrastructure needs will also need to

[42:45] be unique and purpose-built as you can

[42:48] see from the screen here.

[42:51] Just look at the server in front of me.

[42:54] Uh this is whole series of partnership

[42:56] we have.

[43:12] Intel is working with a lot of partners

[43:16] to provide service rack scale solution

[43:20] designed to fit your existing

[43:23] infrastructure ready for AI at scale. As

[43:27] you can tell in front in front of you,

[43:31] we see token usage exploding.

[43:35] Agent now consume 1,000 time more tokens

[43:40] than single event reasoning.

[43:43] In addition to building the best CPUs,

[43:47] it is critical that we deliver compute

[43:50] solutions optimized for token

[43:53] consumption and token generation.

[43:58] The bottom line, AI at scale will

[44:01] require hetogeneous computing. To this

[44:04] end, Intel recently announced a

[44:06] partnership with Sonova.

[44:09] To talk more about this, let me call to

[44:12] the stage founder CEO of Samberonova,

[44:15] Rodrigo Leong.

[44:20] [applause and music]

[44:24] Rodrigo, welcome to joining me today.

[44:27] >> Thank you, Lupu.

[44:28] >> And over the next few months, we have

[44:31] announced few updates on our joint

[44:35] development partnership. Can we talk a

[44:37] little bit more about the work that

[44:39] Intel and Samanova are doing together?

[44:42] >> Absolutely. We've been busy. Earlier

[44:45] this year, we announced a multi-year

[44:48] collaboration to deliver high

[44:50] performance, costefficient AI inference

[44:53] solutions based on Xeon infrastructure.

[44:56] We've been building something really

[44:58] special. Excited to show you today.

[45:03] This is the This is the SM50

[45:06] sambar we announced earlier this year.

[45:09] Rack scale AI infrastructure built for

[45:12] agentic workloads.

[45:15] >> [applause]

[45:18] >> It uses Intel Xeon 6 processors with

[45:21] Sonova SM50 RDUs and shipping to

[45:25] customers later this year. Today, we're

[45:28] also excited to demonstrate the world's

[45:31] first heterogeneous disagregated

[45:33] inference using Sonova's RDU with

[45:37] Intel's CPU and Nvidia GPUs.

[45:41] What you're about to see is the same

[45:42] prompt, the same model running side by

[45:46] side, two different stacks.