Nvidia Pairs With Microsoft to Pivot Silicon Focus Toward AI Agents

Nvidia联手微软将芯片重心转向AI智能体 Nvidia正刻意降低对停滞不前、以人为中心的个人计算市场的投入优先级，以构建专为自主数字实体量身定制的垄断性闭环生态系统。 本视频记录了NVIDIA从以人为中心的个人计算，大幅转向由自主数字智能体驱动、面向企业的“AI工厂”范式。它详细介绍了在与微软的限制性闭环联盟中开发的新硬件架构，例如服务器级的Vera CPU，以及贴近消费端的RTX Spark和N1X笔记本平台。此外，报道还将企业乐观的AI叙事与科技行业的大规模裁员进行了对比，并调查了围绕Supermicro黑市走私突击搜查以及随后中国芯片禁令所引发的、不断升级的法律与地缘政治后果。 NVIDIA 正系统性地将业务重心从以人类为中心的个人计算转向面向自主数字智能体的“AI 工厂”，黄仁勋直言客户不再想买普通电脑，而是瞄准数以百亿计的数字智能体市场以打破人类潜在用户规模的上限。 (0:00) 黄仁勋极力倡导扩大能源生产并呼吁“多生孩子”，声称 AI 劳动力仅消耗电力从而能提升效率，并将代币化计算与能源基础设施深度绑定，使 NVIDIA 加速向新型公用事业基础设施公司转型。 (2:26) 针对 AI 导致的行业失业潮争议，黄仁勋虽坚称 AI 会创造更多岗位，但 Meta、Cisco、HP 等主要科技企业已直接将大规模裁员归因于转向 AI 领域的资本支出。 (5:01) NVIDIA 发布了配备 88 个专有 Olympus 核心并支持高达 1.5TB 板载焊接内存的服务器级 Vera CPU，同时联手微软推出基于 ARM 架构的 N1X 及 RTX Spark 方案，将大语言模型深度整合进 Windows 核心生态，形成高度闭环。 (10:57) 在台当局突袭走私窝点并查扣 50 台涉嫌违规转运至中国大陆的超微（Supermicro）服务器之际，黄仁勋高调展开台湾夜市公关之旅，并试图将 NVIDIA 与走私行为划清界限；同时中国官方已对降规版 RTX 5090 D V2 等芯片出台进口禁令。 (20:02)

Gamers Nexus reports that NVIDIA is systematically redirecting its core business away from consumer personal computing and toward enterprise 'AI factories' built to serve autonomous digital agents. This transformation involves heavy infrastructure demands, closer operating-system integration with Microsoft, and escalating geopolitical friction surrounding export restrictions and illicit chip smuggling. Hardware Re-architected for Autonomous Digital Agents (0:00) NVIDIA CEO Jensen Huang announced that the company's enterprise clients are moving past standard computers to construct massive 'AI factories.' Arguing that the economically relevant human addressable market is constrained to roughly 1 billion users, Huang framed artificial digital agents as an infinitely scalable customer base that requires high-bandwidth, continuous compute. To power these autonomous workloads, NVIDIA unveiled enterprise silicon like the standalone Vera CPU, engineered with 88 Olympus cores, 176 threads, PCIe Gen 6 support, and up to 1.5 terabytes of soldered LPDDR5X memory—a physical design that also prevents decommissioned enterprise silicon from filtering into secondary consumer markets. Energy Demands, Utility Economics, and Workforce Contraction (2:26) NVIDIA's evolving operational thesis positions the company as an infrastructure utility where power converts into tokens, and tokens convert directly into revenue. Drawing comparisons to OpenAI CEO Sam Altman's assertions regarding the inefficiencies of human training cycles, Huang argued that electricity-fed artificial labor holds distinct advantages over food-dependent human labor, while calling for aggressive expansion of power grids and traditional energy extraction. Simultaneously, Huang pushed back against the narrative that AI adoption is driving unemployment, calling the connection an excuse used by corporate management. However, this assertion conflicts with widespread workforce reductions across major tech firms citing capital reallocation toward AI, including Meta cutting 10% of its workforce, Cisco eliminating 4,000 roles, HP displacing between 4,000 and 6,000 workers, Pinterest cutting 15%, and Block dismissing 4,000 employees. Tightening the PC Ecosystem with Microsoft (14:19) NVIDIA and Microsoft announced an initiative to inject large language models directly into the Windows operating system, an architectural evolution Huang likened to the historical introduction of DirectX. The companies revealed the 'RTX Spark' platform, which combines a Blackwell-based graphics processor, a 20-core Grace CPU built with MediaTek, and 128 gigabytes of unified memory to support background AI workloads that exceed conventional consumer requirements. Host Steve Burke highlighted that this ARM-based N1X architecture creates a closed-loop platform where NVIDIA controls the processor silicon and Microsoft controls the operating system, threatening to sideline traditional component manufacturers. While presented onstage as an imminent shift, partner leaks suggest consumer laptops featuring N1X processors may face delays until January 2027. Smuggling Crackdowns and Geopolitical Backlash (20:02) NVIDIA's public events took place amid heightened legal and regulatory scrutiny over the diversion of restricted hardware into China. Taiwanese authorities arrested three individuals accused of forging shipping documents to route 50 Supermicro AI servers to China via Japan, following a U.S. Department of Justice indictment against Supermicro co-founder Wally Liao over an alleged billion-dollar smuggling scheme. While Huang publicly emphasized that NVIDIA requires business partners to follow export laws, Gamers Nexus noted NVIDIA's history of downplaying black-market activity despite close corporate ties. Concurrently, geopolitical friction has intensified. Beijing banned imports of the NVIDIA GeForce RTX 5090 D V2—a stripped-down model engineered specifically to comply with U.S. export thresholds—and blocked domestic tech giants from purchasing newly approved H200 chips, responding partly to an agreement granting Washington a 25% financial cut of Chinese H200 sales. Gamers Nexus concludes that NVIDIA's aggressive reprioritization toward AI agents leaves regular consumers marginalized, binding the company's future to utility-scale energy consumption, tight corporate alliances, and escalating legal disputes across international borders.

Gamers Nexus 指出，NVIDIA 正系统性地将业务重心从以人类为中心的传统个人电脑转向“AI 工厂”与自主数字代理，试图突破有限的人类市场规模。与此同时，这一激进转型伴随着高企的能源诉求、科技行业的裁员潮、与微软日益封闭的软硬件生态绑定，以及持续发酵的跨境走私和地缘政治审查。 算力范式重构与 AI 代理扩张 (0:00) 在近期举办的 GTC 与 Computex 活动中，NVIDIA 首席执行官黄仁勋明确表示其核心客户已不再倾向于购买传统计算机，而是转向构建大型“AI 工厂”。他认为此前所有 CPU 架构皆是为人类终端用户设计，然而全球具备经济价值的人口规模上限仅约 10 亿人，个人电脑年销量因而面临停滞；若将目标转向自主运行的数字代理（AI agents），硬件市场的潜在用户规模将跃升至数十亿级。黄仁勋预测，未来的数字代理数量将远超人类，且对计算响应有着极低的容忍度，迫使底层架构从根本上转向服务自动化软件。 能源博弈与技术失业现实 (2:26) 支撑这一庞大算力体系的代价是巨大的能源开销。黄仁勋将人力劳动与 AI 自动劳动的生产成本进行对照，称培养一名劳动力需耗费约 21 年的食物与生活资源，而 AI 仅消耗电力；为此他公开呼吁大幅扩充电力基础设施与传统能源开采。在面对 AI 引发失业的公众关切时，黄仁勋辩称将大规模裁员归咎于 AI 过于轻率，并主张技术正在催生新企业与新岗位。但媒体指出，现实中企业界出现了广泛的“AI 洗白”式裁员：Meta 削减 10% 员工以对冲 AI 投资，Cisco 裁撤 4,000 人，HP 裁员 4,000 至 6,000 人，Atlassian、Block 及 Pinterest 亦分别缩减 1,600、4,000 人与约 15% 人员编制，均将精简人力归因于向 AI 倾斜资源或追求效率提升。 企业级 Vera 与消费级软硬件闭环 (10:57) 在硬件端，NVIDIA 推出了专为高带宽服务器集群设计的独立 Vera CPU，集成 88 个自研 Olympus 核心、176 线程、164MB 统一 L3 缓存，支持高达 1.5TB 的板载焊接式 LPDDR5X 内存以及 PCIe Gen 6 标准，其高带宽设计直接封死了硬件流入二手消费市场的渠道。面向桌面与移动生态，NVIDIA 与微软深度捆绑，黄仁勋甚至将大语言模型类比为现代版 DirectX。双方联合推出针对工作站与 PC 的 RTX Spark 及 N1X 架构；其中 RTX Spark 配备基于 Blackwell 架构的 6,144 个 CUDA 核心及 128GB 统一内存，为后台常驻的个人代理预留了远超人类软件所需的显存空间。但媒体也援引业内消息提醒，相关 ARM 笔电可能延期至 2027 年初方可交付，且 NVIDIA 掌控硅片核心、微软主导操作系统的紧密协作，极可能进一步挤压传统硬件合作伙伴的生存空间。 走私丑闻与跨国地缘政治夹击 (20:02) 伴随商业版图扩张的是严峻的法律与地缘政治挑战。近期，台湾当局对多处涉嫌非法转运芯片的窝点展开突击搜查，检方扣押了 50 台通过虚假报关、原定经由日本转运至受限中国市场的 Supermicro 高端服务器；此前，Supermicro 联合创始人 Wally Liao 亦于 2026 年 3 月因涉案金额达数十亿美元的走私网络遭到美国司法部起诉。黄仁勋虽在公关巡回中表态要求合作伙伴严格守法，但难掩行业黑市的严峻现实。在国际贸易层面，芯片管制博弈进一步激化：即便美方放行上一代 H200 芯片，中国政府仍全面禁止了 NVIDIA 针对出口管制特别降级的 GeForce RTX 5090 D V2，并限制阿里巴巴、腾讯等国内科技巨头采购 H200 与 H20，以反制美方抽取 25% 销售分成的安排并保护本土半导体产业。 Gamers Nexus 总结认为，NVIDIA 的长期路线已经从满足传统电脑玩家和普通消费者，演变为脱离人类日常需求的垄断性算力巨头。通过人工凭空塑造“AI 代理”这一全新买家群体，NVIDIA 试图维持无休止的硬件扩张，但这种战略转型同时也深陷于能源紧缺、劳动替代争议、走私诉讼与国际贸易摩擦的重重阴影之中。

[0:00] Jensen Huang: don't want to buy computers. [0:00] Jensen Huang: only 1 billion of us [2:26] Jensen Huang: make babies [2:26] Jensen Huang: Because human labor needs rice. But, auto AI labor needs electricity. [2:26] Jensen Huang: drill baby drill [2:26] Jensen Huang: I also hate elevators. I hate the concept of elevators [8:19] Jensen Huang: It's okay to clap. [14:19] Jensen Huang: Large language models, in a lot of ways, is the modern version of DirectX. [17:47] Jensen Huang: the most amazing chip the world's ever built. [17:47] Steve Burke: the slop and the garbage. [20:02] Jensen Huang: Gou ma? [20:02] Jensen Huang: Well, we're very we're rigorous in explaining to all of our partners um all of the world's laws and regulations and uh we uh insist that they're compliant with laws and regulations. Uh ultimately, Supermicro has to run their own company, just as everybody has to run their own company. And I hope that they will they will um uh enhance and improve their regulation compliance and uh avoid that from happening in the future. [20:02] Jensen Huang: We are not dinosaurs [20:02] Jensen Huang: We are not a car. [20:02] Jensen Huang: We never said we were going to invest 100 billion dollars in one round [26:02] Supermicro: We are proud to have worked closely with Taiwanese authorities on the recent event, helping to prevent the illicit diversion of our highly sought-after systems into the restricted China market. Our collaboration with authorities in Taiwan resulted in the arrest of three suspects and the seizure of 50 servers that had been deceptively acquired after being sold by Supermicro to an authorized reseller. [28:19] Black-market broker: castrated version of the card.

[0:00] But the new operating system [0:02] is of course the old operating system [0:04] plus [0:06] large language models. [0:08] Large language models in a lot of ways [0:10] is the modern version of DirectX. [0:13] >> And for all the young ladies in the [0:14] room, this way you can make babies. [0:17] >> I always bring a bottle of whiskey. [0:20] It's okay to clap. [0:21] >> Then you need to make babies. [0:23] >> The new double-07 game, I'm looking [0:24] forward to playing it. [0:26] I look a little bit like him and we're [0:27] working with so many companies, Cadence [0:29] and CrowdStrike and the Soul and [0:31] Palantir. [0:32] >> Make a baby. [0:33] >> If you can hire a software engineer [0:35] and you could generate [0:37] 9 trillion dollars worth of [0:39] productive work, why wouldn't you want [0:41] to hire more software engineers? This is [0:43] going to show up in our economy somehow [0:45] soon. [0:46] >> To be fair, he didn't specify if it show [0:49] up in the economy [0:51] in a good way. [0:53] A global economic crisis would also show [0:57] up in the economy. [1:00] Nvidia's keynote at Computex or GTC has [1:03] concretely stated what we've been [1:04] inferring for a while now. [1:06] >> But ultimately, our customers, our [1:09] partners [1:10] don't want to buy computers. [1:13] They want to build AI factories. All of [1:16] the CPUs until now [1:19] were created for people. [1:21] We were the users. [1:24] We were the users. We were the renters. [1:27] This is CPU [1:29] for agents. [1:31] All the CPUs of the past we built for [1:33] humans. [1:35] This CPU is built for agents. And these [1:38] agents are going to be using the CPUs [1:40] with very little patience. There'll be a [1:43] lot more agents than there are people [1:45] and then there will the agents are very [1:47] impatient. [1:48] >> Nvidia has realized that the total [1:50] adjustable market of humans is low. [1:52] >> We created [music] CPUs [1:55] for humans in the past. [1:57] And humans, there only 1 billion of us. [2:00] >> Now, some of you might be objecting to [2:02] that, saying, "Technically, there's 8.3 [2:05] billion humans, Jensen, but he's talking [2:07] about the ones who matter to him. Jensen [2:09] needs to increase the TAM beyond the 1 [2:11] billion relevant humans and find a way [2:13] to sell billions more computers to [2:15] somebody. And to do that, he's going to [2:17] sell them for use by agents, as they [2:20] call them, [2:22] to get more than the couple hundred [2:23] million PC sales per year and grow into [2:25] the billions. But, Jensen's also going [2:28] to need something else for his AI [2:30] dreams, and that [2:32] is babies [2:33] to feed the AI factories. [2:35] The good news, [2:36] Jensen Huang is just the aphrodisiac we [2:39] need to make them. [2:40] >> Speaking of babies, and for all the [2:42] young ladies in the room, this way, you [2:44] can make babies in addition to the four [2:48] young people who got married, [2:51] they need to get they need to make [2:53] babies, make a baby. Speaking of babies, [2:55] uh and then and then speaking of babies, [2:58] >> Speaking of babies. [2:59] >> But, babies require energy, and energy [3:02] requires food, and Sam Altman already [3:04] told us that humans inefficiently use [3:08] food and energy to train for 21 years or [3:11] so to eventually do work for the rich. [3:15] >> One [clears throat] of the things that [3:16] is always unfair in this comparison is [3:18] people talk about how much energy it [3:20] takes to train an AI model [3:23] relative to how much it cost a human to [3:25] do one inference query. [3:27] But, it also takes a lot of energy to [3:28] train a human. It takes like 20 years of [3:30] life and all of the food you eat during [3:32] that time. Probably, AI has already [3:34] caught up on an energy efficiency basis [3:36] measured that way. [3:37] >> And Jensen agrees. [3:39] >> Because [3:40] human labor needs rice. [3:44] But, auto AI labor needs electricity. [3:48] >> Jonathan Swift would be so proud to see [3:51] his modest proposal realized. Except [3:53] it's the [3:54] the data centers he [3:55] the humans instead, but the point [3:57] stands. Jensen though also needs the [3:59] energy for his data centers. [4:01] >> Uh we could use more energy in Taiwan. [4:03] You need energy. [4:05] Energy growth [4:06] >> Taiwan needs more energy. [4:08] >> Drill baby drill. We need more energy. I [4:10] think that we all recognize that. We [4:12] need more land power and shell. [4:14] >> AI [4:15] >> is [4:16] the world's best opportunity to [4:18] modernize the power grid. [4:20] >> We vilified energy for so long. [4:23] President Trump's taking his neck out [4:25] and making taking it on the chin and [4:27] helping this helping [music] the country [4:30] realize that energy is necessary for our [4:32] growth is one of the the [4:34] really the one of the greatest things [4:35] he's done right off the bat. [4:37] >> Jensen does have other plans for the [4:38] energy shortage though. [4:40] >> I also hate elevators. I hate the [4:43] concept of elevators. [4:46] >> That's right. [4:47] Elevators. Look at all these buttons. [4:50] Inefficient. [4:53] Pitiful. [4:55] You disgust me. [4:58] >> I hate the concept of elevators. [5:01] >> And Jensen has had a busy trip in [5:03] Taiwan. When he wasn't busy vandalizing [5:05] bathrooms or practicing his blue steel, [5:08] Jensen was on the interview circuit and [5:10] spoke about how AI wasn't actually [5:12] causing layoffs. [5:13] >> AI is creating so many jobs. [5:17] AI is creating so many new companies. [5:20] >> I think the the narrative [music] [5:22] that connects AI to job loss [5:25] for many of the CEOs that are doing it, [5:28] um it is just too lazy. AI has just [5:30] arrived. How's it possible they're [5:31] already losing jobs? [5:32] >> Huang's comments follow many companies [5:34] blaming AI for layoffs or AI washing [5:37] their layoffs. Nvidia GPU hoarder Meta [5:39] laid off 10% of its workers to offset [5:41] quote other investments and quote, [5:43] namely AI. Cisco cut 4,000 employees [5:46] even though it insists it will be a [5:48] winner in the AI era. HP will reduce [5:50] head count by 4 to 6,000 people to [5:53] invest in AI adoption. Atlassian cut [5:55] 1,600 people to invest in AI. Block laid [5:58] off 4,000 people due to gains from AI [6:00] and said companies would follow suit. [6:02] Even Pinterest reduced its workforce by [6:04] around 15% to prioritize its resources [6:07] in AI, and the list goes on. But Jensen [6:09] can't be blamed for using his [6:11] imagination. [6:12] >> There's an optimistic future [6:14] that I that I imagine. [6:17] >> [music] [6:17] >> And facts would support my imagination [6:21] of that future. [6:22] >> And that makes sense, because to be [6:23] blind to the damage caused by the AI [6:25] rollout, you'd either need to have a [6:27] good imagination [6:29] or [6:31] be drunk. [6:32] >> We had a We had some scotch. [6:34] >> I brought him some scotch. [6:35] >> I always bring a bottle of whiskey. [6:37] >> But you know I love whiskey. [6:38] >> Cheers. Cheers. Cheers. [6:45] >> Oh, not bad. I think we're going to [6:46] enjoy it. Let's Let's finish the rest of [6:48] it. [6:52] >> [music] [6:54] >> Why not both? We brought you this video [6:56] with the GN store on [6:57] store.gamersnexus.net. [6:59] Like with our GPU shredder shirt, which [7:01] has gotten me looks all over Taipei [7:02] because Nvidia is that recognizable [7:04] here. And although it once represented [7:06] the paper launch of the 50 series, now [7:09] it represents [7:10] what we should do with AI, which is feed [7:12] it into a wood chipper. [7:15] Kind of like how we've fed everything [7:17] into the wood chipper that is AI. In [7:18] addition to these super comfortable [7:20] cotton t-shirts, we also bring our [7:22] coverage to you with our gold halo GPU [7:24] VRAM pint glasses that we just restocked [7:27] after months of being unavailable. And [7:29] no, the irony isn't lost on us that our [7:32] glassware, we literally printed VRAM on, [7:34] was also unavailable for months. The [7:37] glasses have an ultra-fine print detail, [7:39] the gold halo, and go great with [7:41] Jensen's favorite pastime. [7:42] >> But you know I love whiskey. [7:44] >> Our store and the products on it fund [7:45] all of GN's travel because we don't [7:47] accept any travel compensation, [7:49] accommodations, or plane tickets. And [7:51] although uncommon, we also don't accept [7:53] money to visit booths at trade shows. [7:55] This is possible because of your ongoing [7:57] support. So, if you like our independent [7:59] reporting and information gathering, [8:01] consider grabbing something on the GN [8:03] store like our soldering and project [8:05] mats, copper-plated stainless steel mule [8:07] mugs, or GPU VRAM pint glasses. If you [8:09] already have our gear, we also have a [8:12] donate button in the bottom left of the [8:13] store page to give us the greatest cut [8:16] of funding support. Thanks for your [8:17] help. Now, back to it. NVIDIA's Jensen [8:19] Huang says that tokens in data centers [8:21] equal revenue and energy equals tokens. [8:23] Therefore, energy equals revenue. [8:25] They're on their path to becoming a [8:27] utility company, which maybe tracks with [8:29] their plans to be an infrastructure [8:31] company. [8:32] >> The tokens are now in extraordinary [8:35] demand. Because if you could do this, [8:37] you're going to want to produce more of [8:38] it. And because tokens are now [8:40] profitable units. [8:42] Tokens are now prof- prob- [8:45] >> And it's clear that the audience [8:46] absolutely loved Jensen's AI speech. [8:50] >> Do you guys What do you guys think? It's [8:51] pretty amazing, right? [8:53] This is [8:54] NVIDIA Vera. [8:57] What do you think? [8:58] It's okay to clap. [9:00] >> NVIDIA technically had news for its [9:02] keynote that we'll go over like its [9:03] plans to supplant computing for humans [9:05] with computing for AI agents instead. [9:07] And no, that is not hyperbole. He [9:10] literally said that like a dozen times [9:13] throughout the keynote. He would not [9:14] stop saying that it's all about PCs for [9:17] agents now. They also had news on the [9:19] Vera CPU deployment and continued Vera [9:21] Rubin rollout. Also, sort of news on N1 [9:24] and N1X laptops that use NVIDIA CPUs, [9:27] information that GN gathered [9:29] independently from NVIDIA partners. [9:31] NVIDIA also talked about plans for [9:33] investing in what it claims will be [9:35] Yeah, [9:35] billion investment into Taiwan. [9:38] Hopefully with more reliability than its [9:40] investment into Open AI. The company [9:42] also announced its plan to quote [9:44] reinvent Windows PCs for the age of [9:46] personal AI and quote with Microsoft [9:49] soft which it called the quote world's [9:52] at first Windows at PCs purpose built [9:54] for personal agents and quote then [9:56] detailed specs that exceed what the [9:58] actual users typically have access to [10:00] prioritizing their so-called agents for [10:03] the 128 GB of memory instead. [10:05] >> There's lots to talk about today. So, [10:07] Nvidia had sort of the pre-Computex show [10:09] where Jensen went on a tour around [10:11] Taipei meeting with different business [10:12] partners, eating a lot of food, [10:15] eating more food, drinking a lot, and [10:18] and then drinking some more. But, they [10:20] also have their GTC event which happens [10:23] to be at the exact same time as Computex [10:27] which feels just sort of like a [10:28] microcosm for what Nvidia is doing to [10:31] consumers right now. Computex, the [10:33] consumer hardware show, is facing [10:35] basically a hostile takeover from [10:39] Nvidia's GTC which is happening at the [10:41] exact same time. [10:43] >> With blackjack and hookers. [10:45] >> We've had to relocate. We made a grave [10:46] mistake and there was an elevator [10:48] somewhat near our previous location. [10:51] I could feel it in the spirit world as [10:53] Jensen screamed at me through the doors. [10:57] Nvidia had some actual news at the show [10:59] and that news mostly revolved around its [11:01] Vera CPU and its N1 or N1X solutions for [11:05] mobile devices alongside others. Here's [11:08] what Jensen had to say. [11:09] >> CPUs built for the age of AI. [11:12] All the CPUs of the past we built for [11:14] humans. [11:16] This CPU is built for agents. [11:18] This is [11:20] Nvidia Vera. [11:22] What do you think? [11:23] >> Vera spec sheet lists 88 Nvidia Olympus [11:26] cores and 176 threads using what Nvidia [11:29] calls spatial multi-threading. It has [11:31] 164 megabytes of unified L3 cache, up to [11:35] 1.2 terabytes per second aggregate [11:37] memory bandwidth, up to 1.5 terabytes of [11:40] LPDDR5X memory for capacity, and a 250 [11:44] to 450 watt TDP. Vera will utilize a [11:48] NVLink chip-to-chip technology, which [11:50] Nvidia explains stating this. [11:52] >> And memory coherent NVLink chip-to-chip [11:55] connects GPUs [music] [11:56] directly to the fabric. [11:58] Beyond GPUs, NVLink chip-to-chip can [12:01] [music] scale Vera up to multiple [12:03] sockets, enabling massive bandwidth [12:05] between CPUs. [12:07] >> Jensen also added, [12:08] >> It's the first one to be PCI Express Gen [12:11] 6. [12:13] It is also [12:14] the first one to have LPDDR5 [12:18] with 1.2 terabytes per second. [12:20] >> Explaining the decision behind 88 cores, [12:22] Jensen stated, [12:24] >> The number of CPU cores, the number of [12:26] CPUs is going to be quite high. [12:29] And the reason for that is very simple. [12:32] We created CPUs [12:35] for humans in the past. [12:37] And humans, they're only 1 billion of [12:39] us. [12:41] There will be billions of agents, and [12:45] these agents are going to be using the [12:47] CPUs with very little patience. [12:50] >> Jensen also stated what he described as [12:53] Vera's four properties. [12:55] >> These four properties, instructions per [12:57] clock or single-threaded performance, [13:00] bandwidth per core, [13:02] the total bandwidth around the chip and [13:04] inside the chip, [13:06] and energy efficiency [13:08] defines Vera. [13:09] >> In Nvidia's first-party performance [13:11] comparison, the company claimed Vera [13:12] using LPDDR5X multiplied the bandwidth [13:15] per core against an unnamed x86 CPU [13:19] using DDR5. Nvidia claims Vera is, [13:22] quote, "now in full production," [13:23] unquote, and intends to deploy the CPU [13:25] in, quote, "standalone Vera servers and [13:27] video Vera Rubin systems and Vera [13:29] BlueField 4 STX AI storage platforms and [13:32] quote. Each standalone Vera server [13:34] consists of 256 Vera CPUs. Maybe the big [13:37] takeaway here is the extreme amount of [13:39] memory capacity. This is not abnormal [13:41] for Nvidia server solutions. They've [13:43] done this for a while. Actually, even [13:44] the DPUs use memory. And so, when you [13:47] look at where does all the memory go, [13:49] this is part of it. HBM's a lot of it, [13:51] especially because the yields are such [13:52] that if you lose one HBM die, you lose [13:55] potentially multiple DRAM dies depending [13:58] on how far they are in the process. [14:00] But, this is where a lot of that goes [14:01] and this is also not something that can [14:02] just be pulled and sold into the second [14:05] hand market later. If it's LPDDR5X [14:07] soldered to a board or on a card that's [14:09] not suitable for, say, your typical [14:11] desktop system. But, they also had other [14:13] news, which was with Microsoft. [14:16] I said it right that time. [14:18] They'll be proud of me. Nvidia also had [14:20] news that positioned its [14:22] consumer adjacent announcements, where [14:25] in collaboration with Microsoft for [14:26] Windows, Nvidia unveiled its new RTX [14:29] Spark. Nvidia and its partners heavily [14:31] teased the announcement, each tweeting a [14:34] new era of PC along with the coordinates [14:37] of Taipei Music Center, where the [14:38] keynote was held. On stage, Jensen [14:41] attempted to clarify the statement, [14:42] stating, [14:43] >> 40 years later, [14:46] Microsoft and Nvidia [14:48] are going to reinvent the PC. [14:51] This is going to be the new PC. [14:54] >> And we don't know what that means. But, [14:55] Jensen doubled down on LLMs, comparing [14:58] them to DirectX. [14:59] >> But, the new operating system [15:02] is, of course, the old operating system [15:04] plus [15:05] large language models. [15:07] Large language models, in a lot of ways, [15:10] is the modern version of DirectX. [15:13] >> Regarding the specs, the RTX Spark will [15:14] feature a Blackwell RTX GPU with 6144 [15:17] CUDA cores, a claimed one petaflop of [15:20] FP4 AI performance, as they call it, and [15:22] a 20-core Grace CPU built with MediaTek [15:25] connected using Nvidia's NVLink C2C for [15:28] chip-to-chip interconnect, as well as [15:31] 128 GB of unified LPDDR5X [15:34] for the memory. [15:36] Cool. [15:38] You know, I've never thought about using [15:40] a car on a sidewalk, have you? [15:42] >> Lots in the park. [15:43] >> Yeah. [15:44] >> Not like elevators, right? [15:46] >> No, not like elevators. Yeah. The 6144 [15:49] CUDA cores would be the same as found on [15:51] an RTX 5070, but they're not going to be [15:53] comparable directly. The memory [15:55] arrangement is way different. The power [15:56] budget is different. So, it won't [15:58] perform like a 5070. Jensen showed two [16:00] RTX Spark Windows laptop models on [16:02] stage. He said this. [16:04] >> Here it is. [16:07] Of course, [16:08] I got to show you the most beautiful [16:09] part, which is video games. [16:12] It is It's also the closest to our [16:13] heart. [16:14] >> A little late for us to believe that [16:16] one, but Jensen also showed off an RTX [16:18] Spark desktop variant. [16:20] >> Connected to your whole house, [16:22] connected to your laptop, [16:24] connected to your display, all the [16:26] cameras, your your dryer, your water [16:30] cooler, your water heater, your [16:32] everything, whatever you want. Your [16:34] security system, all connected to this, [16:36] and this becomes your personal AI. [16:39] >> Jensen also unveiled the so-called DGX [16:41] Station for Windows. The system specs [16:44] list a GB200 Ultra desktop superchip and [16:47] a 72-core Grace CPU connected via [16:49] Nvidia's new C2C interconnect. [16:52] Additionally, Nvidia claims, quote, "It [16:54] features up to 748 GB of coherent memory [16:58] and up to 20 petaflops of FP4 [17:00] performance and can be paired with an [17:02] Nvidia RTX Pro 6000 Blackwell [17:04] workstation GPU." End quote. Closing out [17:06] the Windows announcements, Jensen [17:08] clarified that Nvidia intended on the [17:10] partnership remaining long-term, [17:12] stating, quote, [17:14] >> This is a brand new product family for [17:15] us. [17:17] Every single generation of architecture, [17:20] we will have a desktop, a laptop, a [17:23] workstation. [17:24] >> Nvidia also briefly showed its N1 and [17:27] N1X laptop solutions. They didn't really [17:29] give detail on this. There are leaks, [17:30] which I think Jensen acknowledged on [17:32] stage. Uh the company showed these a few [17:34] times over the years to press and [17:36] partners, but now they are finally [17:38] moving to launch it, it seems like, and [17:40] here's what they said. [17:41] >> I have too many things in my pocket. [17:47] Okay. [17:48] All right. This is the most amazing chip [17:50] the world's ever built. [17:52] This [17:53] is the N1X that we built in partnership [17:56] with MediaTek. I think I saw I saw Rick [17:59] earlier. This is N1X. This is a [18:01] beautiful chip. And this is this is a a [18:05] a chip that frankly would take 33 years [18:08] to build. And the reason for that is [18:10] because 100% of Nvidia's software stack [18:14] runs here. Microsoft and Nvidia [18:16] meticulously [18:17] optimized everything so that this [18:19] computer literally runs everything the [18:23] world has ever created plus [18:27] it now runs agents. [18:29] >> Some of Nvidia's partners will have N1 [18:30] and N1X laptops to show soon, if not at [18:33] this event. We might be able to see [18:34] some. Uh we were hearing from a couple [18:36] of Nvidia's partners that they won't be [18:38] ready to ship the N1X until closer to [18:40] January 2027. I don't know if that was [18:42] that specific partner or not, but this [18:44] will be an Armat solution, and it really [18:47] sets Nvidia and Microsoft up to be [18:49] more of sort of a closed loop. [18:52] It's just shrinking that circle down a [18:53] little bit to inflate it more between [18:55] the two of them, where they start to [18:57] block out other partners. Cuz now Nvidia [18:58] will do the CPU, they do the GPU, [19:00] Microsoft does the slop and the garbage, [19:03] and between the two of them they have a [19:05] laptop that might work better than the [19:07] one that we're editing this video on, [19:09] which I just bought for over $5,000, has [19:11] a defective 3 and 1/2 mm jack, sometimes [19:14] doesn't turn on, and has at least one [19:16] other problem that I can't remember now, [19:18] but that's for the best because my blood [19:19] pressure's high enough already. [19:22] Hopefully, they make something better [19:24] than that. It's a brand new. [19:26] It's from Asus. Nvidia didn't share [19:28] specs though for its laptop solutions. [19:30] VideoCards has published some rumored [19:32] ones. VideoCards lists the N1X as a [19:34] 20-core as a 10 + 10 build out using, [19:37] they say, Cortex-X9-25 [19:39] and A725 cores. Per VideoCards, it also [19:42] has a 6144 CUDA core arrangement. [19:45] VideoCards likewise lists a 9 + 9 and a [19:47] 40 SM config, stating that, quote, "Both [19:50] N1X chips are designed for a 45 W to 80 [19:52] W power envelope," end quote. And [19:54] rumored specs for the N1 without the X [19:56] have it at 8 + 4 for the CPU, 7 + 3 for [19:59] an alternative CPU with a 2560 or 2048 [20:03] CUDA core layout. Nvidia's PR tour has [20:05] been in full swing this past week though [20:06] as Jensen's friend, Wally Liao of [20:09] Supermicro, faced indictments on [20:11] allegations and charges from the US [20:13] government of GPU smuggling. And also, [20:16] just before Jensen arrived here, the [20:18] Taiwanese government raided a dozen [20:20] locations also for allegations of GPU [20:24] smuggling. And they were seeking [20:26] fugitives at one point in relation to [20:28] that story. Nvidia CEO and man who eats [20:32] corn dogs the long way, Jensen Huang, [20:35] also went back to his roots as a Denny's [20:37] waiter just before GTC, giving out food [20:40] to everyone. For every restaurant Jensen [20:42] visited, he was swarmed by press, [20:44] business partners, security, and [20:46] paparazzi. And for the inconvenience of [20:49] packing a restaurant full of business [20:51] partners, press, and paparazzi, Jensen [20:54] Huang generously tipped the restaurant [20:57] staff. Just have to do the math really [21:00] quick. [21:01] 3/10 billions of his wealth. [21:05] It It sounds like more than it is. [21:09] 2,000 Taiwan dollars from what we could [21:11] see or about 60 bucks US, establishing [21:14] Jensen as the big spender of the [21:17] evening. That's not to mention his [21:18] generosity of handing a street vendor [21:21] another $30 US and asking, literally in [21:24] Chinese, "Gou ma?" which just means, "Is [21:27] it enough?" [21:28] >> Gou ma? Good. My My [21:32] >> A man worth almost 200 billion dollars [21:35] saying, "I'll buy for everybody." And [21:37] and yes, food is very cheap here. I [21:40] think there were probably more than $30 [21:42] worth of people buying food, but it it's [21:45] it's okay. I mean, he's lost [21:47] The man's stepping over hundred-dollar [21:49] bills in the street. He's probably if we [21:51] had ten-thousand-dollar bills, he would [21:52] step over those too. So, but Jensen had [21:54] an image to maintain here, treating his [21:56] gargling gaggle to street food. But as [21:59] for the news, it really doesn't seem [22:00] like a coincidence that the Taiwanese [22:02] government orchestrated a bust of GPU [22:05] smugglers right as Nvidia's CEO arrived. [22:08] And right after Jensen scrambled to fly [22:11] to China with Trump to meet with China's [22:15] own government officials. Just before [22:17] Jensen was swarmed by reporters in [22:19] Taiwan, Taiwanese authorities raided 12 [22:21] locations and started looking for [22:22] fugitives accused of smuggling [22:24] Supermicro's AI servers with Nvidia GPUs [22:27] in them to China. After months of [22:29] denying any evidence of chip smuggling, [22:31] which GN has itself proven at this [22:33] point, Jensen came close to admitting [22:35] its existence. [22:36] >> Well, we're very we're rigorous in [22:38] explaining to all of our partners [22:41] um all of the world's laws and [22:42] regulations and uh we uh insist that [22:45] they're compliant with laws and [22:47] regulations. Uh ultimately, Supermicro [22:49] has to run their own company, just as [22:51] everybody has to run their own company. [22:53] And I hope that they will they will um [22:55] uh enhance and improve their regulation [22:58] compliance and uh avoid that from [23:00] happening in the future. [23:02] >> But Jensen did have important news. He [23:04] has continued narrowing down [23:06] what the it is Nvidia actually is. [23:09] >> Because we are not dinosaurs. [23:13] Nvidia, we're not dinosaurs. [23:16] >> We are not We're not a car. [23:19] We are not a car. [23:20] >> It's like never have I ever except for [23:22] billionaires. Never have I ever been a [23:24] dinosaur. Never have I ever been a car. [23:27] Never have I ever been a loser. I am not [23:30] a loser. We are not a loser. We're not [23:33] losers. [23:36] >> I don't You're not talking to somebody [23:37] who woke up a loser. [23:39] And that loser attitude [23:41] >> Add to it, never have I ever smuggled [23:43] GPUs. While on his Taiwan tour as a [23:46] budding food influencer or food [23:47] fluencer, Jensen talked extensively [23:50] about Nvidia's investment in Taiwan. At [23:52] an employee event, Jensen said Nvidia [23:54] would spend $150 billion in Taiwan every [23:56] year. [23:58] >> 4 years ago, 5 years ago, Nvidia was [24:01] spending about [24:03] 10 [24:04] 15 billion dollars a year in Taiwan. [24:08] Now, [24:09] we're spending 100 going to 150 billion [24:13] dollars in Taiwan each year. [24:16] >> And that's not counting the big spender [24:18] Jensen is for giving these out so [24:20] recklessly. Hey, you got to be careful, [24:22] Jensen. You give too many of these out [24:24] and you'll bankrupt the company. I I [24:26] just even I even uh-oh. Oh, I [24:29] actually need those. I I can't step the [24:32] I can't step over those. I'm not This is [24:35] not This is not one [24:38] Sorry, three 10 billions of my wealth. [24:40] This is a [24:41] This is far greater. This This might be [24:43] like a couple percentage points, [24:44] actually. So, I just I'm going to [24:47] I'm just going to keep those safe. We'll [24:50] We'll just kind of get back to the news. [24:52] So, $150 billion spent. [24:55] That's a lot of money. It's like I don't [24:57] know, it's like half of their revenue [25:00] for the year, more or less. But but we [25:03] know Jensen's good for it. [25:04] >> Nvidia is not going to invest as much as [25:08] 100 billion in Open AI. [25:11] >> We never We never said we were going to [25:13] invest 100 billion dollars in one round. [25:15] That never was said. [25:17] >> As CNBC pointed out, Nvidia's forecasted [25:19] spend of 150 billion would far exceed [25:21] what Nvidia makes in one quarter. Nvidia [25:23] made 81.6 billion dollars in revenue [25:25] last quarter. Nvidia has also made other [25:28] commitments for that revenue. If you can [25:30] still take their word for commitments. [25:32] Normally, you might not need to question [25:33] this, but if you're Sam Altman, [25:36] you might question it. Jensen also [25:37] detailed the new Nvidia campus in [25:39] Taiwan, known as Constellation, which [25:42] will be home to about 4,000 employees. [25:44] Nvidia said that once the site is up and [25:46] running, it would be, quote, one of the [25:48] largest AI research and development [25:50] hubs, end quote, in the Asia Pacific [25:52] region. While all this has been [25:53] happening, the political backdrop has [25:55] continued to be about GPU smuggling and [25:58] the GPU black market. Jensen has [26:00] repeatedly denied allegations of chip [26:02] diversion. [26:02] >> And there's no evidence of any AI [26:05] chip diversion. Nvidia's chips are not [26:08] little tiny potato chips. Old people [26:10] talk about smuggling, but it's [26:11] incredibly hard to do. And as we chase [26:14] down every opportunity, every [26:15] >> Nvidia's statements have called [26:16] smuggling a, quote, non-starter, end [26:18] quote, or likened smuggling to tall [26:21] tales. But our own documentary from a [26:23] year ago continues to be proven correct. [26:25] Taiwanese officials are, quote, seeking [26:27] to detain three individuals for forging [26:29] documents to export Nvidia Corporation [26:32] AI chips to China, end quote, per GNBF [26:35] at Bloomberg. Taiwan accused the three [26:38] suspected smugglers of making, quote, [26:40] fraudulent declarations, end quote, for [26:42] approximately 50 AI servers manufactured [26:45] by Super Micro. Several days later, [26:47] prosecutors arrested three individuals [26:49] involved in the case. Per Tom's Hardware [26:51] via Bloomberg, the chip bootleggers plan [26:53] to ship servers to Japan before [26:56] exporting them to China. Reporters asked [26:58] Jensen about the Supermicro smuggling [27:00] case right when he got to Taiwan ahead [27:02] of GTC and Computex. [27:04] >> Ultimately, Supermicro has to run their [27:06] own company just as everybody has to run [27:08] their own company. [27:09] >> At this point, he has repeatedly gotten [27:11] what he requested. [27:13] >> Every so often somebody says, you know, [27:15] these GPUs are being smuggled. I really [27:17] would love to see it. [27:18] >> In March, the US Department of Justice [27:20] indicted Supermicro co-founder Wally [27:22] Liao and two others for chip smuggling [27:24] in an elaborate billion-dollar scheme [27:26] involving decoy servers and hair dryers, [27:28] something we already detailed in a deep [27:30] dive video that will link below. Jensen [27:32] was pictured with Wally Liao several [27:34] days before the US government charged [27:36] Liao with the crime. Just days ago, [27:38] before GTC, Supermicro commented on the [27:41] new raids by Taiwanese police. Quote, [27:44] "We are proud to have worked closely [27:45] with Taiwanese authorities on the recent [27:47] event, helping to prevent the illicit [27:49] diversion of our highly sought-after [27:52] systems into the restricted China [27:53] market. Our collaboration with [27:54] authorities in Taiwan resulted in the [27:56] arrest of three suspects and the seizure [27:58] of 50 servers that had been deceptively [28:00] acquired after being sold by Supermicro [28:03] to an authorized reseller." End quote. [28:05] >> Do you want it? I have nothing to hide. [28:06] Do you want to search my car or house or [28:09] warehouse [28:11] filled with dummy servers that [28:13] hair dryers? [28:16] Have we moved to the dummy servers and [28:17] the hair dryers? You can search the [28:19] warehouse if you want to. [28:20] >> While the US clears Nvidia's last [28:22] generation H200 AI chips for export to [28:24] China, and while Taiwan arrest people [28:26] for fraud, China is responding by this [28:29] past couple weeks banning more Nvidia [28:31] chips for import into China. [28:36] We don't want your stinking chips [28:38] anyway, I guess is where we're at at [28:39] this point as China tries to bolster its [28:42] own now growing and in some cases [28:44] booming chip development industry. [28:47] According to the Financial Times, the [28:48] Chinese government has now banned the [28:50] RTX 5090 D V2, which in our GPU black [28:53] market smuggling documentary that we [28:55] published last year, we'll link below, [28:57] one of the sellers of that series of [29:00] GPUs called it a castrated version of [29:04] the card. But even with its BGA balls [29:06] cut off, the modified version of the [29:08] 5090 that Nvidia created for China to [29:10] comply with the US government's export [29:12] controls is no longer allowed in China [29:15] because of China this time. China's bans [29:17] go beyond the 5090 D V2 balls GPU. FT [29:21] also reported that China blocked sales [29:23] for Nvidia's chips like the freshly [29:25] approved H200 and H20 specifically to [29:28] customers like Alibaba and Tencent. In [29:31] January, the US government and Nvidia [29:33] agreed to a deal that would give the US [29:34] government a 25% cut from Nvidia's H200 [29:37] sales to Chinese customers, which we [29:39] think is likely why the Chinese [29:41] government doesn't want to allow the [29:42] sale. This comes at a time when Dell, a [29:45] close partner of Nvidia's and a recent [29:47] stock purchase of Trump who bought [29:49] somewhere between one and five million [29:51] dollars of Dell stock, has just been [29:52] given a nearly ten billion dollar [29:54] defense contract. Sometimes when I'm [29:56] pulling on my bootstraps, I think to [29:58] myself, [30:00] if only I knew about the next ten [30:02] billion dollar technology defense [30:03] contract deal, I [30:05] could make millions of dollars on that [30:07] company's stock. I [30:09] I don't know why I haven't tried that [30:11] yet. Just didn't really occur to me. [30:16] The contract includes a quote suite of [30:18] software to the US military, unquote, [30:20] and references the quote joint [30:22] war-fighting cloud capability contract, [30:25] unquote again. [30:26] The war-fighting cloud capability [30:29] contract, and Dell is a major customer [30:32] of both Nvidia and Microslopsoft. [30:37] Micros [30:38] slopsoft. Microsoft. [30:41] We do have a micro slob t-shirt on the [30:42] store though. [30:44] Which you can buy it at help us do stuff [30:46] like this. [30:48] posting except verbally in Donald [30:50] Bridge in Taiwan. Really the big [30:52] takeaway here for the show is something [30:53] we've been talking about a little bit, [30:55] but it's just becoming more obvious. So [30:59] uh [30:59] it it's kind of interesting. They they [31:01] actually they did catch me by surprise a [31:02] little bit a couple of months ago when [31:03] they started making the push towards [31:04] this idea of AI agents using PCs. And we [31:08] kind of knew they were reducing their [31:10] focus on consumer PCs. [31:12] What was interesting or surprising [31:14] though was this concept of like what if [31:16] we create the people out of thin air to [31:19] sell the computers to instead? Uh that [31:21] was not really I didn't think it would [31:22] go that direction, but that's really [31:24] what they're pushing here. I mean [31:26] I can't tell you how many times [31:27] throughout the keynote, which is very [31:29] long, he said that they [31:32] there will be more AI agents using [31:34] computers than people using computers, [31:36] especially if you do the math at 1 [31:38] billion humans. Like [31:40] >> Jensen Huang [31:42] >> He's a little [31:43] little out of touch, but [31:45] that's it for the news. There wasn't a [31:47] ton of sort of hard news that's [31:48] interesting with specs and [31:51] you know, just objective information. [31:53] Certainly not much for consumers. We [31:54] gave you all the stuff that's [31:55] interesting. A lot of more concerning [31:58] movements from Nvidia, especially uh on [32:01] the data center rollout, the sort of [32:03] political side we've been talking about, [32:06] and that'll cover it for now for Nvidia. [32:08] So subscribe for more, go to [32:09] store.gamersnexus.net to support us [32:11] always.