英伟达携Vertiv与优步拓宽物理AI生态软硬件基建构筑全栈护城河

浑水摸鱼清源

AI投资正从单纯的数据中心转向机器人、电信基站与自动驾驶等物理基础设施，掌控不可替代硬件与聚合平台的关键供应商将锁定长期收益。

AI-assisted summary of 【Ticker Symbol YOU】英伟达CEO黄仁勋在GTC柏林的演讲让我震撼 — original source. Check the original for full context.

Source published: 2026-10-27

Key ideas

英伟达在GTC柏林大会展示AI正从纯数据中心向工厂机器人、基站、自动驾驶和具身智能等物理基础设施全面扩散。 (0:00)
1吉瓦规模的AI工厂成本高达500亿至600亿美元（约50至60美元/瓦），英伟达推出集成504颗GPU与液冷的预制DSX pod，将部署周期从数月压缩至数周。 (1:05)
电信网络拥有约100吉瓦电力容量，英伟达携手诺基亚推进AI-RAN软件化改造，将基站转化为兼顾通信与低延迟AI推理的分布式边缘算力网。 (8:20)
基于“卖水人”平台逻辑，作者提出25%英伟达搭配Vertiv、Coherent、Cognex、Nokia及Uber各15%的六资产配置模型，规避单一模型或造车竞争风险。 (15:35)

Detailed notes

## Background Published on Bilibili by the channel 浑水摸鱼清源 on October 27, 2026, this video translates and breaks down an analysis by Alex from Ticker Symbol YOU regarding NVIDIA CEO Jensen Huang's keynote at GTC Berlin. According to the page description, the presentation focuses on physical AI infrastructure beyond the data center—highlighting capital-intensive AI factories, core supply chain partners like Vertiv and Coherent, industrial robotics powered by Jetson Thor and Cognex, and the transformation of cellular networks into AI-RAN.

## TL;DR Former MIT AI researcher Alex argues that Wall Street misunderstands artificial intelligence by focusing strictly on NVIDIA's current data center revenue while missing the massive wave of physical AI and industrial infrastructure. Drawing from Jensen Huang's GTC Berlin keynote, Alex details the capital mechanics of 1-gigawatt AI factories, the upstream 'three-computer' robotics architecture, the conversion of Western telecom grids into 100-gigawatt AI-RAN networks, and autonomous mobility partnerships with Uber. Rather than speculating on individual model developers or robot makers, he recommends a pragmatic picks-and-shovels investment strategy anchored by NVIDIA (25%) and balanced across critical infrastructure leaders Vertiv, Coherent, Cognex, Nokia, and Uber (15% each).

## Atmosphere & style The presentation carries the authoritative, methodical demeanor of a seasoned electrical engineer delivering a masterclass in technology investing. Alex cuts briskly through financial headlines and market hype, grounding every claim in hard physical realities like gigawatts, thermal management, cabling distance, and compute throughput. It feels intellectually rigorous yet accessible, building a clear emotional arc from contrarian skepticism toward Wall Street into a high-conviction, empowering investment thesis for patient long-term allocators.

## Bulletin - [00:00](https://www.bilibili.com/video/BV12sn967EjM?t=0s) 英伟达在GTC柏林大会展示AI正从纯数据中心向工厂机器人、基站、自动驾驶和具身智能等物理基础设施全面扩散。 - [01:05](https://www.bilibili.com/video/BV12sn967EjM?t=65s) 1吉瓦规模的AI工厂成本高达500亿至600亿美元（约50至60美元/瓦），英伟达推出集成504颗GPU与液冷的预制DSX pod，将部署周期从数月压缩至数周。 - [08:20](https://www.bilibili.com/video/BV12sn967EjM?t=500s) 电信网络拥有约100吉瓦电力容量，英伟达携手诺基亚推进AI-RAN软件化改造，将基站转化为兼顾通信与低延迟AI推理的分布式边缘算力网。 - [15:35](https://www.bilibili.com/video/BV12sn967EjM?t=935s) 基于“卖水人”平台逻辑，作者提出25%英伟达搭配Vertiv、Coherent、Cognex、Nokia及Uber各15%的六资产配置模型，规避单一模型或造车竞争风险。

## Key takeaways - [00:00](https://www.bilibili.com/video/BV12sn967EjM?t=0s) 华尔街过度紧盯英伟达超92%的数据中心营收占比，低估了AI在物理基础设施和边缘场景中的指数级扩张潜力。 - [01:05](https://www.bilibili.com/video/BV12sn967EjM?t=65s) AI工厂的核心商业逻辑在于尽早开始生产tokens，一旦延误3个月可能导致高达100亿美元的潜在收入损失。 - [01:55](https://www.bilibili.com/video/BV12sn967EjM?t=115s) DSX pod整合7个机架的Vera Rubin系统及25 exaflops算力，工厂预制化测试大幅降低现场施工与排线调试难度。 - [02:36](https://www.bilibili.com/video/BV12sn967EjM?t=156s) Vertiv深度绑定英伟达，承担DSX pod组装与AI工厂供电散热设计，单季营收增长24%并上调全年预期至140亿美元。 - [03:14](https://www.bilibili.com/video/BV12sn967EjM?t=194s) Coherent获得英伟达20亿美元投资与长期大单，为共封装光学交换机供应激光组件，前瞻市盈率约为22倍。 - [05:27](https://www.bilibili.com/video/BV12sn967EjM?t=327s) 工业机器人正从繁琐的手动编程转向“三电脑”架构：数据中心训练模型、数字孪生仿真技能、边缘电脑端侧执行。 - [06:37](https://www.bilibili.com/video/BV12sn967EjM?t=397s) 西门子实践表明，引入数字孪生可将工厂工程工作量减少85%，并使产品上市周期缩短40%。 - [07:00](https://www.bilibili.com/video/BV12sn967EjM?t=420s) 康耐视（Cognex）推出搭载Jetson（157 TOPS）的工业视觉控制器，以零负债和稳健现金流卡位机器视觉与AI质检。 - [08:20](https://www.bilibili.com/video/BV12sn967EjM?t=500s) 西方电信基站拥有10万个站点和100吉瓦功率，AI-RAN技术将其专用射频硬件替换为运行在通用加速芯片上的软件栈。 - [09:20](https://www.bilibili.com/video/BV12sn967EjM?t=560s) 英伟达向诺基亚注资10亿美元，诺基亚目标是将频谱效率翻倍，并大幅受益于AI数据中心光纤与IP网络需求爆发。 - [10:00](https://www.bilibili.com/video/BV12sn967EjM?t=600s) 自动驾驶迎来“ChatGPT时刻”，英伟达推出开源视觉语言动作模型Alpamayo与Hyperion平台，覆盖全球逾14%汽车产能合作伙伴。 - [10:30](https://www.bilibili.com/video/BV12sn967EjM?t=630s) 英伟达联手Uber计划到2028年在全球28座城市部署Robotaxi，由Uber充当需求分发与运营调度层。 - [11:58](https://www.bilibili.com/video/BV12sn967EjM?t=718s) 相比前瞻市盈率高达175倍且利润率承压的特斯拉，Uber以17倍前瞻估值聚合超30家自动驾驶伙伴，具备更优风险收益比。 - [13:17](https://www.bilibili.com/video/BV12sn967EjM?t=797s) 英伟达与Agile Robots联合发布Agile 1SN人形机器人参考平台（4.5英尺高、31个自由度），开箱即用降低研发门槛。 - [14:10](https://www.bilibili.com/video/BV12sn967EjM?t=850s) 新一代Blackwell架构Jetson Thor芯片算力较Orin提升7.5倍、能效提高3.5倍，开发者套件起售价3500美元。 - [15:35](https://www.bilibili.com/video/BV12sn967EjM?t=935s) 作者建议采取“1+5”组合投资：25%英伟达作为核心基石，15%分别配置Vertiv、Coherent、Cognex、Nokia与Uber，锁定各垂直产业链咽喉要道。

## Detailed walkthrough ### [00:00](https://www.bilibili.com/video/BV12sn967EjM?t=0s) Wall Street's Blind Spot and the Reality of AI Factories

Alex, an electrical engineer and former AI researcher at MIT for eight years, opens with a blunt critique of conventional market analysis: “I'm convinced that Wall Street still doesn't understand AI.” The core problem is that finding great investments requires deeply understanding a company's underlying physical products, not merely scanning backward-looking financial figures—especially in exponential growth markets. On paper, NVIDIA appears almost exclusively to be a data center business: last quarter, the company brought in $96.2 billion in revenue, with $89 billion (over 92%) coming straight out of data centers, while PCs, gaming consoles, robotics, and automotive combined accounted for less than 8%. Yet at GTC Berlin, CEO Jensen Huang delivered a keynote that pointed toward an entirely different transformation. While headlines fixated on data centers, Alex highlights that the real disruption is unfolding across everyday infrastructure—factory robots, cellular towers, robotaxis, and humanoids.

Building these systems requires colossal capital. Jensen Huang noted on stage that a 1-gigawatt AI factory costs roughly $50 billion, while NVIDIA's official blog estimates the figure closer to $60 billion. Because one gigawatt equals one billion watts, building out capacity translates directly to about $50 to $60 per watt—a clear baseline metric for comparing capital expenditure across the sector. Crucially, only a portion of that budget flows to NVIDIA itself; the remainder pays for real estate, buildings, power distribution, and liquid cooling infrastructure. Execution speed is everything: if a 1-gigawatt AI factory generates $40 billion annually once operational, a mere 3-month construction or commissioning delay destroys approximately $10 billion in potential revenue. As Alex emphasizes: “Remember, an AI factory makes money when it starts producing tokens, not when construction is complete.”

To compress these costly build schedules, NVIDIA unveiled the DSX pod. Rather than forcing specialized engineering crews to spend months on site manually unboxing, mounting, plumbing, and testing discrete components, the DSX pod integrates seven racks of Vera Rubin systems—encompassing 504 GPUs and delivering around 25 exaflops of AI compute—fully liquid-cooled and pre-wired with over 10 miles of internal cabling. Fabricated and pre-tested entirely in the factory before shipping, it cuts field deployment time from months down to a few weeks, with the first production units scheduled to ship early next year.

```mermaid flowchart TD accTitle: AI Factory Deployment Bottleneck and Solution accDescr: Comparison showing the old multi-month manual data center deployment versus NVIDIA's pre-fabricated DSX pod approach. subgraph Traditional["Traditional Buildout (Months of Delays)"] direction TB T1["Raw Land & Facility Construction"] --> T2["On-Site Hardware Delivery"] T2 --> T3["Manual Racking & Cabling (10+ Miles)"] T3 --> T4["On-Site Fluid & Thermal Testing"] T4 --> T5["Delayed Token Generation (~$10B Lost / Qtr)"] end

subgraph DSXPod["NVIDIA DSX Pod Solution (Weeks to Deploy)"] direction TB D1["Factory Pre-Fabrication (7 Racks / 504 GPUs)"] --> D2["Integrated Liquid Cooling & Testing"] D2 --> D3["Plug-and-Play Site Delivery (Early Next Year)"] D3 --> D4["Rapid 25 Exaflops Token Output"] end ```

[02:36](https://www.bilibili.com/video/BV12sn967EjM?t=156s) Pure-Play Infrastructure Picks: Vertiv and Coherent

To capitalize on the industrial buildout of AI factories, Alex highlights two critical suppliers directly plugged into NVIDIA's supply chain: Vertiv (ticker: VRT) and Coherent (ticker: COHR). Vertiv provides critical power and thermal management equipment and was highlighted on two of NVIDIA's key partner registries at GTC Berlin: as an official manufacturing partner assembling the DSX pods, and as a core engineering partner designing power and cooling architectures for the broader AI factory footprints. In its most recent reported quarter, Vertiv expanded revenue by 24% year-over-year to $3.3 billion while raising its full-year revenue projection to roughly $14 billion.

On the high-speed networking layer, the 10 miles of interconnect within these systems rely on NVIDIA Spectrum-X Ethernet switches, with the latest generation transitioning to co-packaged optics (CPO) rather than pluggable optical transceivers. Jensen Huang showcased one of these switches during the Berlin keynote, confirming they are already in full volume production. Coherent supplies the specialized lasers and photonic components that drive these optical engines. Back in March, NVIDIA solidified this relationship by investing $2 billion directly into Coherent alongside a multi-billion-dollar long-term procurement contract. Coherent's most recent quarter demonstrated a 34% year-over-year revenue increase to slightly above $2 billion. Despite this acceleration, Coherent trades at roughly 22 times forward earnings, giving it a valuation multiple lower than mega-caps like Alphabet, Amazon, or Microsoft.

- **Vertiv (VRT)**: Earnings report scheduled for October 28th; full-year guidance raised to ~$14 billion. - **Coherent (COHR)**: Earnings report scheduled for November 4th; 34% YoY growth, supported by a $2B NVIDIA equity investment and multi-billion-dollar purchase agreement.

[05:00](https://www.bilibili.com/video/BV12sn967EjM?t=300s) Sponsor Integration: Liquid

Alex transitions briefly to the video's sponsor, Liquid, addressing the reality that major generative AI players like OpenAI and Anthropic remain privately held. Liquid allows traders to gain exposure to pre-IPO asset valuations via perpetual contracts. Branded as the "everything exchange," the platform supports 24/7 trading across equities, crypto, commodities, foreign exchange, and prediction markets, enabling investors to react instantly to earnings releases or weekend macroeconomic developments. Liquid also integrates CoInvest, an AI trading assistant that ties directly into Anthropic's Claude and OpenAI's ChatGPT to analyze market drivers and construct trade tickets without autonomous execution, keeping user confirmation mandatory. Since its launch the prior year, Liquid has cleared over $6 billion in trading volume.

[05:27](https://www.bilibili.com/video/BV12sn967EjM?t=327s) The "Three-Computer" Architecture and Cognex's Vision AI

Turning to industrial operations, Alex points out that factories represent the primary proving ground for physical AI: there are currently more than 5 million industrial robots operating globally, with over 600,000 commissioned in the past year alone. Historically, almost every unit has been laboriously hand-coded to execute rigid, repetitive tasks, requiring physical downtime and manual fine-tuning across the line whenever a product redesign occurs. NVIDIA's strategy flips this paradigm by shifting robotic training out of the factory floor into an upstream digital workflow.

According to Jensen Huang, true robotic automation requires a coordinated three-computer architecture: 1. **The Data Center Computer**: Trains the foundation models and high-level artificial intelligence. 2. **The Digital Twin Computer (Simulation)**: Runs physically accurate, real-time virtual simulations (such as NVIDIA Omniverse) where the robot's neural network can practice and master complex manipulation tasks risk-free. 3. **The Embedded Edge Computer**: Runs directly inside the physical robot body (such as NVIDIA Jetson modules) to execute inference and real-time control in the physical plant.

Alex references Siemens, which reported that deploying full-scale digital twins slashes factory engineering workload by up to 85% while cutting new product time-to-market by 40%. The broader industry shift is definitive: “The big idea here is that every industrial company will eventually become a robotics company.”

To invest in this transition, Alex singles out Cognex (ticker: CGNX). Founded in 1981 to build optical character recognition systems for industrial parts, Cognex has evolved into a premier provider of industrial machine vision, smart sensors, and vision-AI software that directs automated production lines across automotive, electronics, and consumer packaged goods. Cognex released an advanced industrial vision controller powered by the NVIDIA Jetson platform capable of delivering up to 157 trillion operations per second (TOPS), alongside its OneVision software suite currently utilized by hundreds of industrial accounts. Financially, Cognex posted record quarterly revenue of $291 million (up 17% year-over-year), while diluted earnings per share surged by 80%. Cognex carries a pristine balance sheet featuring $755 million in liquid cash and investments with zero long-term debt, projecting strong double-digit top-line expansion for the year ahead of its November 4th earnings release.

```mermaid flowchart LR accTitle: NVIDIA Three-Computer Robotics Architecture accDescr: The workflow across data center training, digital twin simulation, and embedded runtime execution. subgraph Computer1["Computer 1: Data Center"] C1["AI Foundation Model Training"] end

subgraph Computer2["Computer 2: Simulation"] C2["Digital Twin Environment (Physics-Accurate Sim)"] end

subgraph Computer3["Computer 3: Physical Edge"] C3["Embedded Robot Controller (NVIDIA Jetson / 157 TOPS)"] end

C1 -->|"Transfers Neural Nets"| C2 C2 -->|"Validates Skills & Workflows"| C3 C3 -.->|"Feeds Real-World Edge Data"| C1 ```

[08:20](https://www.bilibili.com/video/BV12sn967EjM?t=500s) Cellular Infrastructure: Turning 100 GW of Telecom into AI-RAN

The final transformation Alex details is the repurposing of global telecom networks into distributed AI engines. Across the Western world, cellular telecommunications providers operate approximately 100,000 physical cell sites representing roughly 100 gigawatts of aggregate grid power capacity. Putting that figure in perspective, Amazon Web Services added 3.9 gigawatts of power capacity over the past year—which AWS noted was more new power than any other enterprise on Earth secured. Repurposing 100 gigawatts of already electrified, real-estate-cleared edge infrastructure represents a massive leap forward.

This transition is known as AI-RAN (AI Radio Access Network). Today, cellular towers rely on bespoke, single-purpose application-specific hardware dedicated entirely to processing 5G radio frequency waveforms. AI-RAN replaces this rigid equipment with software-defined radio stacks running on top of NVIDIA's accelerated edge computing nodes. Under this setup, identical server silicon dynamically manages mobile calls and processes low-latency local AI inference workloads simultaneously. Nokia is spearheading this push with a stated engineering goal to double cellular spectral efficiency—moving twice the throughput over existing radio frequency bands—with live pilot programs beginning late this year and broader commercial rollout slated for 2027. Consequently, Alex highlights Nokia (ticker: NOK), pointing to NVIDIA's direct $1 billion investment into Nokia equity at $6.50 per share as validation of the platform's potential.

[10:00](https://www.bilibili.com/video/BV12sn967EjM?t=600s) Autonomous Driving: NVIDIA's Alpamayo, Hyperion, and the Uber Partnership

Before diving into automotive autonomy, Alex finishes evaluating Nokia’s telecommunications turnaround, noting that its fastest-growing segment stems directly from AI data centers. Nokia's sales to cloud and AI customers more than doubled last quarter, with its optical and IP networking division projected to expand by at least 25% in the coming year. Yet cell towers represent only one piece of everyday infrastructure being upgraded by physical AI; the vehicles on the road are undergoing a parallel transformation. At CES earlier in the year, NVIDIA introduced Alpamayo, an open-source family of generative vision-language-action AI models designed to ingest camera feeds, plot safe driving trajectories, and articulate the rationale behind every driving choice. Automakers utilize Alpamayo to train distilled, localized driving models directly on vehicle hardware. After experiencing a demonstration firsthand inside an Alpamayo-powered Mercedes at NVIDIA GTC, Alex emphatically endorses Jensen Huang’s assessment: “And after riding in a Mercedes that runs Alpamayo at NVIDIA GTC and seeing it for myself, I agree with Jensen when he says the chat GPT moment for self-driving cars has arrived.”

NVIDIA capitalized on this inflection point by announcing an expansive commercial alliance with Uber to deploy autonomous robo-taxis built on NVIDIA's full-stack driving architecture and routed through Uber's rideshare ecosystem. The initial rollout begins in Los Angeles and San Francisco during the first half of the rollout schedule, targeting commercial driverless services across 28 global metropolitan markets by 2028. Automakers including BYD, Geely, Isuzu, and Nissan are already constructing robo-taxi-compliant production vehicles around NVIDIA Hyperion—a standardized hardware-software suite comprising sensor arrays, high-performance compute modules, and certified safety operating software. Combined, those four manufacturers accounted for more than one out of every nine vehicles sold globally last year. At GTC Berlin, momentum expanded further as Volvo, Mercedes-Benz, and Jaguar Land Rover (JLR) committed to Hyperion compatibility, capturing an incremental 3% of global automotive volume.

```mermaid flowchart TD accTitle: NVIDIA Autonomous Vehicle Ecosystem accDescr: Diagram illustrating how foundational AI models, hardware platforms, and ride-hailing networks connect to global automakers.

subgraph Intelligence ["AI Foundation & Compute"] A["Alpamayo Models (Vision & Reasoning)"] --> B["NVIDIA Hyperion (Sensors, Compute & OS)"] end

subgraph Manufacturing ["OEM Vehicle Fleet (14%+ Global Sales)"] B --> C["Mass-Market OEMs (BYD, Geely, Isuzu, Nissan)"] B --> D["Premium OEMs (Mercedes, Volvo, JLR)"] end

subgraph Distribution ["Global Fleet Deployment"] C --> E["Uber Ride-Hailing Network (28 Cities by 2028)"] D --> E end ```

- **Market Reach of Hyperion Partners**: The initial quartet of BYD, Geely, Isuzu, and Nissan represents over 11% (one in nine) of global vehicle deliveries, expanding past 14% with the additions of Mercedes-Benz, Volvo, and JLR. - **Rollout Timeline**: Driverless operations launch commercially in Los Angeles and San Francisco before scaling to 28 international cities by 2028. - **Jensen Huang’s Core Premise**: “As Jensen put it, RoboTaxis are the first mass-market application of robotics and physical AI, and Nvidia's bet is that almost every carmaker will want to plug into one shared platform instead of having to develop and test their own.”

---

[11:58](https://www.bilibili.com/video/BV12sn967EjM?t=718s) The Investment Case: Why Uber Beats Tesla in Autonomous Mobility

Evaluating how investors should position themselves for the autonomous mobility shift, Alex argues that the winning equity play is Uber rather than standalone hardware manufacturers or Tesla. Rather than gambling on which vehicle OEM engineers the premier autonomous stack, Uber operates as the indispensable demand and logistics layer; self-driving fleets ultimately require passenger volume, and Uber already maintains integration agreements with more than 30 separate autonomous vehicle developers. Uber’s fundamentals reflect operational discipline: revenue rose 18% over the past year, while operating margins expanded sequentially to exceed 12%. Because Uber takes a platform fee on rides hailed through its application regardless of who manufactures the car, it monetizes autonomy friction-free.

In stark contrast, Alex rejects Tesla as an autonomous vehicle investment based on both current valuation and operational trajectory. While Tesla trades at an eye-watering forward price-to-earnings multiple of 175 times, its top-line revenue stalled flat in the most recent quarter and operating margins have experienced multi-year compression. Alex highlights the valuation discrepancy by pointing out that an investor could aggregate the forward price-to-earnings ratios of Apple, Microsoft, Amazon, Alphabet, Meta, and NVIDIA combined, and Tesla’s standalone valuation multiple would still exceed all six mega-cap tech leaders put together—despite Tesla displaying the weakest growth rates and lowest margins of the peer group.

- **Valuation Comparison**: Uber trades at approximately 17 times forward earnings (less than 10% of Tesla’s multiple), whereas Tesla trades at 175 times forward earnings. - **Utilization Scale**: Uber completes more trip bookings and deliveries in a single day than Waymo logs across an entire operating year, while Tesla does not officially disclose autonomous ride volume. - **The Aggregator Advantage**: By integrating over 30 self-driving partners, Uber extracts cash flow from every completed robo-taxi fare without absorbing vehicle depreciation or manufacturing capital expenditures.

---

[13:17](https://www.bilibili.com/video/BV12sn967EjM?t=797s) Humanoid Robotics: Agile 1SN, Jetson Thor, and GR00T

The keynote shifted decisively toward embodiment, with Jensen Huang asserting that humanoid robots represent the eventual highest-volume physical robotic form factor because our built world—from staircases and door handles to assembly lines—was designed exclusively around human dimensions. To lower the barrier for robotics researchers and enterprises, NVIDIA introduced a turnkey reference platform co-engineered with German robotics firm Agile Robots: the Agile 1SN. Standing 4.5 feet tall with 31 independent degrees of freedom (joints), the robot comes fully pre-assembled to bypass the years of bespoke mechatronic engineering that typically paralyze research labs.

```mermaid flowchart LR accTitle: NVIDIA Humanoid Robotics Stack accDescr: Flow diagram showing how Jetson Thor hardware and GR00T foundational software integrate into reference platforms and OEM designs.

subgraph NVIDIA ["NVIDIA Core Stack"] JT["Jetson Thor Computer (Blackwell Architecture)"] GR["Project GR00T (Pre-Trained Robot Brain)"] end

subgraph Integration ["Reference Platform"] JT --> A1["Agile 1SN Humanoid (4.5 ft, 31 DOF)"] GR --> A1 end

subgraph Ecosystem ["Commercial Humanoid OEMs"] JT -.-> BD["Boston Dynamics"] JT -.-> FIG["Figure"] JT -.-> AR["Agility Robotics"] JT -.-> OX["1X & OpenAI (Evaluating)"] end ```

The computing architecture inside the Agile 1SN centers on NVIDIA’s newly unveiled Jetson Thor embedded processor and Project GR00T multimodal models. Built upon the Blackwell data center GPU microarchitecture, Jetson Thor executes inference directly on the robot rather than relying on high-latency cloud connections. GR00T acts as a foundation model—an off-the-shelf "robot brain" that enterprises fine-tune for specific manipulation and locomotion workflows. Alex points out that platform ownership is where asymmetric value accrues, observing: “Uber, the world's largest taxi company, doesn't own any vehicles, and Airbnb became one of the biggest places to book a stay without owning any real estate. I think the same thing will happen in robotics, where the biggest winner might not make any robots themselves.”

- **Jetson Thor Performance Leap**: Delivers up to 7.5 times the AI computational throughput of its predecessor, the Jetson Orin, while operating with 3.5 times greater energy efficiency. - **Developer Accessibility**: The Jetson Thor development kit starts at $3,500, drastically reducing experimentation costs. - **Industry Buy-In**: Major robotics innovators including Boston Dynamics, Figure, and Agility Robotics are building on Jetson Thor, with 1X and OpenAI actively evaluating the hardware platform. - **Hardware Availability**: The Agile 1SN reference platform begins commercial shipments early next year ready out of the box.

---

[15:35](https://www.bilibili.com/video/BV12sn967EjM?t=935s) Portfolio Blueprint: How to Invest Across the AI Stack

Alex synthesizes the technological breakthroughs from Jensen Huang’s GTC Berlin keynote into a pragmatic, balanced investment framework. With over 92% of NVIDIA’s present revenue generated by data centers, data infrastructure remains the principal revenue engine today, but physical AI and edge deployments represent the next massive economic wave. Rather than speculating on early-stage, single-purpose hardware vendors or unproven AI startups, Alex recommends allocating across the indispensable platform layers: power and optics for the data center AI factories, vision and cellular connectivity for physical automation, ride-hailing networks for mobility, and NVIDIA as the foundational operating ecosystem across simulation, training, and real-time physical inference.

```mermaid flowchart TD accTitle: Recommended 6-Stock AI Allocation accDescr: Structural breakdown of the recommended investment allocation across AI factories, edge infrastructure, consumer aggregation, and core compute.

subgraph Portfolio ["AI Hardware & Infrastructure Allocation"] NVDA["NVIDIA (NVDA): 25% Core Ecosystem & Compute"] VRT["Vertiv (VRT): 15% AI Factory Power & Cooling"] COHR["Coherent (COHR): 15% Transceivers & Optical Networking"] CGNX["Cognex (CGNX): 15% Factory Vision & Robotics"] NOK["Nokia (NOK): 15% AI-RAN & Edge Telephony"] UBER["Uber (UBER): 15% Autonomous Fleet Orchestration"] end

NVDA --- VRT NVDA --- COHR NVDA --- CGNX NVDA --- NOK NVDA --- UBER ```

Alex outlines his specific capital distribution: maintaining a core anchor position of 25% in NVIDIA, while assigning equal 15% weightings to the remaining five supporting leaders—Vertiv (VRT), Coherent (COHR), Cognex (CGNX), Nokia (NOK), and Uber (UBER). He underscores that Nokia, Uber, and NVIDIA all currently trade below the broader S&P 500 forward price-to-earnings average of 19. Five of the six companies are slated to report quarterly financial results within weeks, providing immediate catalysts. Emphasizing the fundamental thesis of platform investing, Alex summarizes why owning the picks-and-shovels ecosystem delivers enduring capital appreciation: “Every one of these companies also has a direct connection to Nvidia and owns a layer of the stack that everyone else relies on, so it doesn't matter which AI Lab trains the best model, whose robots roll off the assembly line first, or which carmaker builds the best robo-taxi. These companies all still get paid, making them a great way to get rich without getting lucky.” He signs off with his channel motto, reminding viewers that their personal growth remains their highest-yielding asset.

- **Core Position (25%)**: NVIDIA (NVDA) — Serves as the operating system and silicon backbone for model training, Omniverse physical simulation, and on-robot edge inference. - **Data Center Infrastructure (15% each)**: Vertiv (power management and liquid cooling) and Coherent (optical transceivers and photonics). - **Physical & Edge Infrastructure (15% each)**: Cognex (AI-powered industrial machine vision) and Nokia (AI-RAN telecom networks and data center optical routing). - **Mobility Platform (15%)**: Uber (UBER) — The demand aggregator that monetizes autonomous robo-taxi deployments across all partner fleets. - **Valuation Advantage**: NVIDIA trades around 17 times forward earnings alongside Nokia and Uber, all sitting below the S&P 500's benchmark multiple of 19.

## Notable quotes > [00:00](https://www.bilibili.com/video/BV12sn967EjM?t=0s) Remember, an AI factory makes money when it starts producing tokens, not when construction is complete. — Alex

> [05:27](https://www.bilibili.com/video/BV12sn967EjM?t=327s) The big idea here is that every industrial company will eventually become a robotics company. — Alex

> [00:00](https://www.bilibili.com/video/BV12sn967EjM?t=0s) I'm convinced that Wall Street still doesn't understand AI. — Alex

## Mentioned Alex, Jensen Huang, 英伟达 (NVIDIA), Vertiv (VRT), Coherent (COHR), Cognex (CGNX), Nokia (NOK), Uber (UBER), Tesla, DSX pod, Vera Rubin, Spectrum-X, Jetson Thor, Jetson Orin, Omniverse, Project GR00T, Alpamayo, Hyperion, Agile Robots, Agile 1SN, Liquid, Siemens, T-Mobile, Amazon Web Services (AWS), Alphabet, Microsoft, Boston Dynamics, Figure, Agility Robotics, 1X, OpenAI, Anthropic, Waymo

## Video description 【Ticker Symbol YOU】英伟达CEO黄仁勋在GTC柏林的演讲让我震撼 NVIDIA CEO Jensen Huang Left Me SPEECHLESS at GTC Berlin 机器人算力是这场的重头：新一代机器人芯片 Jetson Thor 较上代 Orin 性能最高提升 75 倍、能效高 35 倍，全球工厂已有超 500 万台工业机器人在运行，黄仁勋据此称机器人时代已经到来。

｜AI工厂：基建与效率提升英伟达CEO黄仁勋在GTC大会上指出，一个1吉瓦的AI工厂成本约为500亿至600亿美元。其中，英伟达的DSX pod旨在解决AI工厂部署效率问题。DSX pod集成了504颗GPU，具备约25 exaflops的AI计算能力，采用液冷技术，通过超过10英里的电缆连接。该系统可在工厂预制和测试，将部署时间从数月缩短至数周，首批DSX pod预计明年初发货。AI工厂的盈利关键在于开始生产tokens，而非完成建设。

｜AI工厂核心供应商：VRT与COHR Vertiv（VRT）是英伟达的重要合作伙伴，为DSX pod提供电源和冷却系统，并参与英伟达AI工厂设计。Vertiv上季度销售额同比增长24%至33亿美元，全年营收预测上调至约140亿美元。Coherent（COHR）则提供英伟达Spectrum X以太网所需的激光和光学组件，其最新交换机采用共封装光学技术。英伟达在3月向Coherent投资20亿美元，并签订了数十亿美元的采购协议。Coherent上季度营收同比增长34%至略超20亿美元，目前市盈率约为Alphabet、亚马逊和微软的22倍。Vertiv和Coherent分别计划于10月28日和11月4日发布财报。

｜工业机器人：从手动编程到AI驱动全球工业机器人数量已超500万台，去年新增60万台。传统机器人多为手动编程，新产线部署需逐一调试。英伟达致力于将机器人训练从工厂车间前置，通过数据中心训练AI、数字孪生模拟环境学习技能，以及机器人内部计算机执行任务。西门子表示，数字孪生可将工厂工程工作量减少85%，产品上市时间缩短40%。Cognex（CGNX）作为机器视觉系统供应商，其产品包括摄像头和AI软件，用于指导生产线机器人。Cognex今年推出了搭载英伟达Jetson的视觉控制器，每秒可执行157万亿次AI操作。上季度Cognex营收达到创纪录的2.91亿美元，同比增长17%，每股收益增长80%。公司拥有7.55亿美元现金和投资，无债务，预计今年营收将实现两位数增长。Cognex将于11月4日发布财报。

｜蜂窝网络：向分布式AI网络转型黄仁勋指出，西方世界的电信网络拥有约10万个站点，总计约100吉瓦的功率容量，有望成为一个巨大的分布式AI网络。相比之下，亚马逊网络服务去年新增3.9吉瓦电力容量。AI RAN（AI无线接入网络）理念是将蜂窝站点的基础设施从专用硬件转变为在英伟达加速计算机上运行的软件，使其能同时处理通话和AI工作负载。本月，英伟达、T-Mobile和诺基亚在华盛顿的一个商用蜂窝站点完成了全球首次AI RAN通话。诺基亚的目标是实现两倍的频谱效率。

接下来建议的操作指引： 1. Vertiv（VRT）10月28日发布财报，关注上季度33亿美元销售额（同比+24%）之后，全年约140亿美元营收指引的兑现情况。 2. Coherent（COHR）11月4日发布财报，关注英伟达20亿美元投资叠加数十亿美元采购协议，能否延续上季度20亿美元（同比+34%）的光组件营收增速。 3. Cognex（CGNX）11月4日发布财报，关注搭载英伟达Jetson的AI视觉控制器放量，验证上季度2.91亿美元创纪录营收（同比+17%）与两位数全年增长预期。

[0:00] 华尔街对AI的理解不足 [0:36] 英伟达GTC大会亮点：数据中心之外的AI应用 [1:05] AI工厂成本与效率：1吉瓦工厂500-600亿美元 [1:55] 英伟达DSX pod：加速AI工厂部署 [2:36] Vertiv（VRT）：AI工厂电源与冷却系统供应商 [3:14] Coherent（COHR）：英伟达Spectrum X以太网光学组件供应商 [3:55] Coherent财务数据与英伟达投资 [4:17] Vertiv和Coherent财报预告 [5:27] 工业机器人：AI驱动的未来 [6:00] 英伟达的机器人三电脑架构 [6:37] 西门子：数字孪生提升工厂效率 [7:00] Cognex（CGNX）：机器视觉与AI软件 [7:50] Cognex财务数据与产品亮点 [8:10] Cognex财报预告 [8:20] 蜂窝网络：100吉瓦的分布式AI网络潜力

[Source](https://www.bilibili.com/video/BV12sn967EjM)

中文简报

英伟达携Vertiv与Uber拓宽物理AI生态，以软硬件基建筑牢全栈护城河 AI投资正从单纯的数据中心转向机器人、电信基站与自动驾驶等物理基础设施，掌控不可替代硬件与聚合平台的关键供应商将锁定长期收益。前麻省理工学院AI研究员Alex认为，华尔街仅关注英伟达当前的数据中心营收，却忽略了物理AI与工业基础设施的巨大浪潮，是对人工智能的误读。结合黄仁勋在GTC柏林大会的主题演讲，Alex深入剖析了1吉瓦AI工厂的资本运作机制、机器人领域的上游“三电脑”架构、西方电信网络向100吉瓦AI-RAN网络的改造，以及与Uber在自动驾驶出行方面的合作。相比于押注个别模型开发商或机器人制造商，他建议采取务实的“卖水人”投资策略：以英伟达为核心持仓（25%），并均衡配置于关键基础设施领军企业Vertiv、Coherent、Cognex、Nokia和Uber（各占15%）。英伟达在GTC柏林大会展示AI正从纯数据中心向工厂机器人、基站、自动驾驶和具身智能等物理基础设施全面扩散。 (0:00) 1吉瓦规模的AI工厂成本高达500亿至600亿美元（约50至60美元/瓦），英伟达推出集成504颗GPU与液冷的预制DSX pod，将部署周期从数月压缩至数周。 (1:05) 电信网络拥有约100吉瓦电力容量，英伟达携手诺基亚推进AI-RAN软件化改造，将基站转化为兼顾通信与低延迟AI推理的分布式边缘算力网。 (8:20) 基于“卖水人”平台逻辑，作者提出25%英伟达搭配Vertiv、Coherent、Cognex、Nokia及Uber各15%的六资产配置模型，规避单一模型或造车竞争风险。 (15:35)

Reading summary

根据电气工程师及前MIT研究员Alex对英伟达GTC柏林主题演讲的解读，华尔街目前过度局限于数据中心财务数字，而低估了物理AI在电信、工业制造与移动出行等实体基础设施中的扩散潜力；投资者应关注围绕英伟达技术底座的各关键平台层与供应链伙伴，而非单纯押注单点硬件或未经证实的终端厂商。 AI工厂建造成本与预制化方案 (0:00) 英伟达上季度总营收达962亿美元，其中数据中心收入890亿美元（占比超92%），反映出算力基建目前的庞大规模。黄仁勋在GTC柏林指出，一座1吉瓦（1 GW）的AI工厂造价约500亿至600亿美元，即基建资本开支约为每瓦50至60美元。由于1 GW工厂全面运转后每年预计产生约400亿美元营收，工期若延误3个月可能造成约100亿美元的潜在收入损失，因此工程交付速度至关重要。针对传统现场装机需要数月手动布线与调试的瓶颈，英伟达发布了DSX pod。该方案在出厂前将7个Vera Rubin机架（包含504个GPU，提供约25 exaflops算力）与超过10英里的内部线缆及液冷系统完全预制并测试完毕，可将现场部署时间从数月压缩至数周，首批设备预计明年初发货。配套产业链中，维谛技术（Vertiv）作为官方制造与工程伙伴，提供电力及液冷架构，其最新季度营收增长24%至33亿美元；高意（Coherent）则为内部光互联供应激光器与光子组件，获得英伟达20亿美元股权投资与长期采购合同，最新季度营收增长34%。工业制造与蜂窝网络的边缘算力重构 (5:27) 在工业自动化领域，全球现有超过500万台工业机器人，黄仁勋提出了包含“数据中心模型训练、数字孪生（如Omniverse）仿真演练、边缘嵌入式计算（如Jetson平台）物理控制”的三电脑协同架构。该架构旨在替代传统逐行人工编程的调试流程；西门子数据表明，数字孪生可削减高达85%的工厂工程工作量并缩短40%的新品上市周期。康耐视（Cognex）推出了基于Jetson平台（算力达157 TOPS）的工业视觉控制器与OneVision软件，其最新季度营收达2.91亿美元，持有7.55亿美元现金且无长期债务。在电信基础设施端，西方电信运营商拥有约10万个物理基站，聚合电网容量约100吉瓦，规模远超AWS去年新增的3.9吉瓦电量。通过AI-RAN（AI无线电接入网）架构，英伟达推动将传统单用途5G射频硬件重构为软件定义边缘计算节点，使基站既能处理通话又能分流低延迟推理。诺基亚正在主导这一转型，目标是将射频频谱效率翻倍，英伟达已对其投资10亿美元，诺基亚计划于今年底开展试点并于2027年实现全面商业化部署。自动驾驶平台化与出行网络竞争 (10:00) 在汽车与自动驾驶领域，英伟达推出开源视觉-语言-动作模型家族Alpamayo，车企可用其提炼并在车端部署本地化驾驶模型。针对商业落地，英伟达与优步（Uber）达成全栈软硬件合作，计划在洛杉矶与旧金山率先启动Robotaxi车队运营，并在2028年前扩展至全球28个大都市。目前，基于英伟达Hyperion标准硬件套件造车的车企包括比亚迪、吉利、五十铃与日产（合计占全球汽车销量11%以上），随后沃尔沃、梅赛德斯-奔驰与捷豹路虎也宣布兼容Hyperion，累计覆盖全球汽车销量的14%以上。 Alex在投资逻辑上倾向于优步而非特斯拉。他认为优步作为需求和物流分发层，已集成超过30家自动驾驶开发商，无需承担车辆制造成本与折旧即可抽取平台费用，其营收同比增长18%，营业利润率超12%，预期市盈率约为17倍。相比之下，特斯拉预期市盈率高达175倍，其单体估值倍数超过苹果、微软、亚马逊、Alphabet、Meta与英伟达的总和，但最新财季营收停滞且营业利润率持续承压。人形机器人标准平台与全栈组合配置 (13:17) 黄仁勋认为人形机器人最终将成为出货量最大的物理机器人形态，英伟达联合德国Agile Robots推出了Agile 1SN交钥匙参考平台。该机器人高4.5英尺，具备31个自由度，核心计算基于Jetson Thor芯片与GR00T多模态基础模型。相较于前代Orin，基于Blackwell架构的Jetson Thor计算吞吐量提升至7.5倍，能效提升3.5倍，开发套件起售价3500美元，波士顿动力、Figure及Agility Robotics均已采用该硬件。结合上述工业与实体演进，Alex提出了覆盖AI全栈的六资产投资组合配置：以英伟达（NVDA）作为核心底座配置25%仓位；其余75%资金等权（各15%）分配至数据中心基建标的维谛技术（VRT）与高意（COHR）、工业视觉与边缘网络标的康耐视（CGNX）与诺基亚（NOK），以及出行聚合网络优步（UBER）。该策略的核心理念是持有整个技术栈中不可或缺的平台层与基础设施，从而规避单一硬件或模型竞赛的研发波动风险。该论述的核心前提在于物理AI能够顺利从概念验证迈向商业化量产，但电信网络全面向AI-RAN转型、人形机器人落地以及大规模Robotaxi商业运营仍高度依赖2027至2028年的时间表兑现，期间各垂直领域车企与工业客户的实际资本开支意愿仍是关键考验。

精炼阅读摘要

电气工程师兼前MIT研究员Alex指出，华尔街仍将英伟达单纯视为数据中心企业，而忽视了其向物理AI与工业基础设施的深度渗透；投资者应聚焦于英伟达及其关键供应链与聚合平台构成的核心底座。 AI工厂建造成本与模块化部署方案 (0:00) 黄仁勋在GTC柏林演讲中透露，建造一座1吉瓦（GW）规模的AI工厂成本约为500亿至600亿美元，折合每瓦资本开支达50至60美元。在这类资本密集型项目中，除了芯片采购外，大量预算流向地产、电力分配与液冷设施。Alex强调，工厂一旦投产年产值可达400亿美元，这意味着哪怕延期投产3个月，也会造成约100亿美元的潜在代币产出损失。为了缩短现场耗时数月的组装排线流程，英伟达推出了DSX pod，将包含504颗GPU、可输出约25百亿亿次（exaflops）算力的7个机架在出厂前完成全液冷和内嵌线缆装配，预计明年初出货，可将现场部署压缩至数周。在此硬件链条中，维谛技术（Vertiv）负责电源与液冷架构设计及DSX pod制造，其季度收入同比增长24%至33亿美元；高意（Coherent）则为Spectrum-X以太网交换机的共封装光学（CPO）技术供应激光器与光子器件，并在英伟达20亿美元股权投资及长期采购订单支持下实现了34%的季度营收增长。工业三电脑架构与电信网络重塑 (5:27) 在工业机器人领域，全球现有超过500万台工业机器人长期依赖耗时的人工硬编码。黄仁勋提出的“三电脑”架构试图颠覆该现状：第一台数据中心电脑训练基础大模型；第二台数字孪生电脑（基于Omniverse）进行高精度物理仿真以演练操作技能；第三台嵌入式边缘电脑（基于Jetson）在工厂车间执行实时推理。西门子数据显示，全面部署数字孪生能使工厂工程工作量减少多达85%，并将新产品上市周期缩短40%。康耐视（Cognex）推出了基于Jetson平台算力达157 TOPS的视觉控制器与OneVision软件，季度营收达2.91亿美元，手握7.55亿美元现金且无长期债务。与此同时，全球蜂窝电信基站所拥有的约100吉瓦并网电力容量正在被探索转型为AI无线接入网（AI-RAN）。诺基亚正在推进软件定义方案，目标是在同一硬件上并行处理通信波形与低延迟本地AI推理，力求将频谱效率翻倍；英伟达已对其投资10亿美元，诺基亚计划于今年底开展实地试点并于2027年铺开商业部署。自动驾驶生态结盟与Uber商业模式 (10:00) 在自动驾驶领域，英伟达发布了端到端开源视觉-语言-动作模型家族Alpamayo，车企可借此在车载硬件上蒸馏训练出具备解释能力的驾驶模型。此外，基于传感器、计算单元与安全操作系统的Hyperion平台已吸引比亚迪、吉利、五十铃、日产、奔驰、沃尔沃及捷豹路虎加入，覆盖全球逾14%的汽车销量。英伟达进一步与Uber达成商业联盟，计划从洛杉矶和旧金山起步，至2028年在全球28个大都市部署自动驾驶出租车（Robotaxi）。Alex认为，在移动出行投资逻辑上，Uber相较于特斯拉更具优势：特斯拉远期市盈率高达175倍且营收增长停滞、利润率承压；而Uber远期市盈率约为17倍，年营收增长18%，营业利润率超12%。Uber已接入超过30家自动驾驶开发商，能够凭借聚合网络在免除重资产车辆折旧和制造成本的前提下，从每一次自动驾驶出行中抽取平台佣金。人形机器人标准平台与全栈配置组合 (13:17) 黄仁勋预测，由于人类现实环境是围绕人体工学构建的，人形机器人终将成为产量最高的物理形态。为降低研发门槛，英伟达联手德国敏捷机器人（Agile Robots）推出了高4.5英尺、拥有31个自由度的Agile 1SN整机参考平台，定于明年初交付。该平台搭载基于Blackwell架构的Jetson Thor嵌入式处理器（算力为Orin的7.5倍，能效提升3.5倍，开发套件售价3500美元）以及GR00T多模态通用“机器人大脑”，波士顿动力、Figure与Agility Robotics等均已接入该技术体系。基于此，Alex构建了一个跨越数据中心、边缘通信与物理场景的六股配置方案：以英伟达作为核心底座配置25%仓位；维谛技术、高意、康耐视、诺基亚和Uber各占15%。该逻辑强调持有不可替代的平台底层与关键基建，从而摆脱单一大模型优劣或特定硬件量产成败的单一风险。该投资框架高度依赖英伟达软硬件生态的持续统治力，以及AI-RAN与自动驾驶出租车等技术路线在2027至2028年间能否如期兑现商业回报；传统车企与电信运营商对统一平台的实际采纳意愿仍有待时间验证。

Source quotes

[0:00] Alex: Remember, an AI factory makes money when it starts producing tokens, not when construction is complete. [5:27] Alex: The big idea here is that every industrial company will eventually become a robotics company. [0:00] Alex: I'm convinced that Wall Street still doesn't understand AI.

Original transcript

[0:00] I'm convinced that Wall Street still doesn't understand AI. That's because the key to finding great stocks is understanding a company's products, not just their profits, especially in high-growth markets. My name is Alex, and I spent 8 years as an electrical engineer and AI researcher at MIT. And after NVIDIA's latest keynote, I'm sure that the AI revolution is far from over. So let me show you what just happened and and how I'm investing in it. Your time is valuable, so let's get right into it. If you just look at NVIDIA's financials, it's a data center company. Last quarter, NVIDIA made $96.2 billion in revenue, and 89 billion of that came from data centers. That's over 92%. Everything else, from PCs and game consoles, to robots and automotive, added up to less than 8%. but their products tell a different story. Last week, NVIDIA CEO Jensen Huang gave a keynote at GTC Berlin, and most of the headlines were about data centers. the parts I found most exciting were basically everything else. Even the cell towers and factory robots that we have today are about to get a whole lot smarter. So, that's what I want to focus on in this video. The latest news on AI factories and the companies building them, how AI impacts the things we already rely on today like factory robots and cell towers, the markets of the future like robotaxis and humanoid robots, and of course, which stocks I'm buying as a result. Let's start with data centers, since they generate the biggest returns today. These systems are gigantic. A 1 gigawatt data center, 1 gigawatt AI factory, costs about $50 billion. In his keynote, Jensen said that a 1 gigawatt AI factory costs about $50 billion and NVIDIA's blog puts the price even higher, at about $60 billion. A gigawatt is a billion watts of power, so every watt of data center capacity costs about $50 to $60 to build. That's a good metric for investors to compare between companies and track over time. Only a part of that money goes to NVIDIA. The rest pays for things like land, buildings, power, and cooling. those pieces has to be finalized for the factory to make money. Using Jensen's own example, if a 1 gigawatt data center brings in $40 billion a year, then a 3-month delay costs about $10 billion in lost revenue. That's the problem their new DSX pod is trying to solve. Each pod holds 7 racks worth of Vera Rubens. That's 504 GPUs and about 25 exaflops of AI computing, all liquid cooled and connected by over 10 miles of cable. That usually takes specialized data center crews months to unpack, install, and install. connect and test on site. but the DSX pod can be built and tested at the factory and then shipped already ready to plug in. Nvidia says this cuts deployment times from months down to weeks, and the first DSX pods are scheduled to ship early next year. Remember, an AI factory makes money when it starts producing tokens, not when construction is complete. My first stock pick here is Vertiv, ticker symbol VRT. Vertiv makes power and cooling systems and it's on two of Nvidia's most important partner lists from GTC Berlin. They're one of the companies building DSX pods. and one of the power and cooling partners for NVIDIA's AI factory designs. Last quarter, Vertiv's sales grew by 24% year-over-year to $3.3 billion, and they raised their full-year revenue forecast to about $14 billion. And on the networking side, I'm buying Coherent, ticker symbol COHR. Those 10 miles of cable are NVIDIA Spectrum X Ethernet, and their newest switches use co-package optics. instead of being separate modules that plug into external ports. Jensen showed one on stage and said that they're already in full volume production. Coherent makes the lasers and optical components for these systems, and back in March, Nvidia invested $2 billion in Coherent and signed a multi-billion dollar purchase agreement as well. Last quarter, coherence revenue grew by 34% year-over-year. to just over $2 billion. And it currently trades at about 22 times forward earnings, giving it a lower forward P/E ratio than Alphabet, Amazon and Microsoft today. Vertiv reports earnings on October 28th and Coherent reports on November 4th. So let me know in the comments if you want me to make a video breaking those down for you. By the way, OpenAI and Anthropic are two of the biggest AI companies in the world. but they're still private. That's where Liquid comes in, the sponsor of this video. Liquid uses perpetual contracts to track almost any asset's price, which means you can trade around the price action of pre-IPO companies. But Liquid goes way beyond private companies. [5:00] calls itself the everything exchange, because you can trade stocks, crypto, commodities, foreign exchange, and even prediction markets, all on one platform. And on Liquid, all of these markets trade 24 hours a day, 7 days a week, so you can react to earnings and breaking news even at night or over the weekend. And then there's CoInvest, Liquid's AI assistant, which connects directly with Claude and ChatGPT so I can research a market, Understand what's moving it. analyze trades, and even prepare orders right inside the conversation. Nothing executes automatically. I still review and confirm every trade. That combination of pre-IPO markets, 24-7 trading, and AI investing is why Liquid is growing so fast. Since launching last year, Liquid has handled over $6 billion in trading volume and raised funding from top investors in the space. So if you want to trade almost any asset 24-7 and integrate investing directly into Quad or Chat GPT, check out Liquid with my link below today. Alright, data centers are really just the real-world impact is everywhere else. There are now over 5 million industrial robots working in factories around the world, and more than 600,000 of them were installed just last year. Most of these robots are programmed by hand to do a single job, and setting up a new production line means individually tuning and testing every robot on a factory floor. NVIDIA wants to move that work off of the factory floor and earlier into the robot's lifecycle. Jensen says that every robot actually needs three computers to run, one in the data center to train its AI, a second one to run a digital twin, which is a physically accurate simulation of its environment where it can learn and practice skills, and a third computer inside the robot itself. When I was at GTC back in March, some of the world's biggest robot makers talked about building simulation software directly into the tools they use to set up factory lines and put Jetson modules robot controllers. And last week, Siemens said that standing up digital twins can cut factory engineering efforts by up to 85% and get new products to market up to 40% faster. The big idea here is that every industrial company will eventually become a robotics company. When it comes to factory robots, check out Cognex, ticker symbol CGMX. Cognex was founded in 1981 read the letters and numbers printed on products and parts. Today, they make machine vision systems, cameras and AI software that inspect parts and guide robots on production lines in a wide variety of industries, like automotive, consumer electronics, and packaged goods. This year, Cognex launched a vision controller powered by NVIDIA's Jetson that delivers up to 157 trillion And their new OneVision platform is already being used by hundreds of customers building AI vision applications. Last quarter, Cognex hit record revenues of $291 million, which was up 17% year-over-year. but their earnings per share jumped by 80%. They have $755 million in cash and investments, no debt, and expect strong double-digit revenue growth this year. Cognex is about to report earnings on November 4th. But another one that every single one of us relies on is the Cell Network. Last week, Jensen pointed out that telecom networks have about 100,000 sites across the Western world, which add up to about 100 gigawatts. That could all become one giant distributed AI network. For context, Amazon Web Services added 3.9 gigawatts of power capacity last year. and they say that was more new capacity than any other company on Earth. So turning 100 gigawatts of already up and running telecom networks into distributed AI capacity is a really big deal. The idea is called AI-RAN, or AI Radio Access Network. The infrastructure at each cell site that connects to your phone runs on specialized hardware today, built for 5G radio signals. But with AI-RAN, it runs as software on NVIDIA's accelerated computers, so the same hardware can handle our calls and AI workloads at the same time. Nokia's goal is to achieve two times the spectral efficiency. Or, said in English, they want to move twice as much data through the same bands of airwaves, with pilot deployments starting at the end of this year, and commercial availability in 2027. That's why my pick here is Nokia stock, ticker symbol NOK. NVIDIA invested a billion dollars in Nokia stock at $650. [10:00] dollars per share last year. and it trades at around $8 today. Nokia's fastest growth is coming from AI data centers. Last quarter, their sales to AI and cloud customers more than doubled, and they expect their optical and IP networking business to grow by 25% or more next year. But cell towers aren't the only thing we rely on that's getting smarter thanks to AI. So let's talk about the cars we drive next. When I was at CES back in January, Nvidia released Alpamayo, a family of open source AI models that watch video from a car's cameras, decide where it should drive, and explain their reasoning along the way. carmakers use Alpamayo to train the smaller self-driving models that actually run in their cars. And after riding in a Mercedes that runs Alpamayo at NVIDIA GTC and seeing it for myself, I agree with Jensen when he says the chat GPT moment for self-driving cars has arrived. That's also when NVIDIA and Uber announced plans to launch robo-taxis that run on NVIDIA's self-driving stack and Uber's ride-hailing network. Starting in LA and San Francisco in the first half and scaling to driverless services in 28 cities around the world by 2028. BYD, Geely, Isuzu, and Nissan are all building robo-taxi-ready cars on the Hyperion platform. That's NVIDIA's full package of sensors, computers, and safety software that automakers can build on to have cars drive themselves. These four brands account for over one out of every nine new vehicles sold on Earth last year, and at GTC Berlin, Volvo, Mercedes, and JLR all signed on to be Hyperion compatible as well. That's another roughly 3% of all new vehicles sold last year. As Jensen put it, RoboTaxis are the first mass-market application of robotics and physical AI, and Nvidia's bet is that almost every carmaker will want to plug into one shared platform instead of having to develop and test their own. If you agree with that logic, the stock to buy is Uber. Regardless of who wins the self-driving race, RoboTaxis still need riders and Uber already works with more than 30 self-driving partners. Uber's revenues grew by 18% last year. Their operating margins have climbed every single year to over 12%, and they earn money on robo-taxi rides no matter which carmaker comes out on top so long as the rides themselves get booked on Uber. Let me tell you why I didn't pick Tesla. Tesla's revenues were flat last quarter. Their operating margins have been shrinking for years, and the stock trades at an astronomical 175 times forward earnings. That means you can add up the forward P.E. ratios of Apple, Microsoft, Amazon, Alphabet, Meta, and NVIDIA, and Tesla would still be more expensive than all six companies combined, while also having the slowest growth and lowest margins by far. Uber stock trades at less than 10% of Tesla's multiple, at 17 times forward earnings, more trips in a single day than Waymo does in an entire year, especially when you include food deliveries. Tesla doesn't currently report their number of rides. And that brings me to humanoid robots. During GTC Berlin, Jensen said that humanoid robots are likely going to be the highest volume configuration of robot. And that makes sense, since the modern world is built for humans by humans, so human-shaped robots will have an easier time operating in it. And to speed the industry up, Nvidia designed a reference robot alongside a German company called Agile Robots that builds and sells it. It's called the Agile 1SN. It stands about 4.5 feet tall, has 31 independent joints, and runs on Nvidia's Jetson Thor computer and Groot robotic AI models. Jetson Thor is a small computer that goes inside the robot built on the same Blackwell chip design as Nvidia's data center GPUs. and waiting on the cloud, and Groot is is NVIDIA's family of open models specifically for humanoid robots. Think of it as a pre-trained robot brain that companies can then fine-tune for their own robots and jobs. The Agile 1SN ships early next year, pre-assembled and ready to go out of the box, which lets researchers skip months or even years building their own robots from scratch. But here's the thing: Uber, the world's largest taxi company, doesn't own any vehicles, and Airbnb became one of the biggest places to book a stay without owning any real estate. I think the same thing will happen in robotics. where the biggest winner might not make any robots themselves. Companies like Uber, Airbnb, and Nvidia win by owning the platform that everyone else plugs into. That's why I think Nvidia is actually the best stock for humanoid robots. Most companies working on them are still private, and many of the big ones are already building on Nvidia's Jetson Thor. [15:00] including Boston Dynamics, figure and agility robotics. OneX and OpenAI are evaluating it too. Jetson Thor delivers up to 7.5 times the AI computing of NVIDIA's previous robot chip, called Jetson Oran, while being 3.5 times more energy efficient and the developer kit starts at just $3,500. That means the platform's barrier to entry is very low. Nvidia itself trades at about 17 times forward earnings, which is cheaper than the overall S&P 500. So let me put everything together and show you exactly how I'd split my money between these stocks. And if you feel I've earned it, consider hitting the like button and subscribing to the channel. That really helps, and it lets me know to make more content like this. Thanks, now let's talk about my top stocks after Jensen's keynote at NVIDIA GTC Berlin. Over 92% of NVIDIA's revenue still comes from data centers today, and that's where most of the money is, which is why Vertiv and Coherent are my picks for the AI factory build-out. For the infrastructure we already rely on every day, like factory robots to make our goods, and cell networks to keep us all connected, I like Cognex and Nokia. taxis plug into, and NVIDIA is the platform that robots will use for simulation, training, and real-time inference. If I had to split my money between them, I'd keep Nvidia as the core 25% position, since they own the ecosystems that everyone else is building on, and then I'd put 15% into the other 5 stocks. Nokia, Uber and Nvidia all trade below the S&P 500's forward PE ratio of 19, and 5 of these 6 companies report earnings in the next few weeks. Every one of these companies also has a direct connection to Nvidia and owns a layer of the stack that everyone else relies on, so it doesn't matter which AI Lab trains the best model, or the best model. whose robots roll off the assembly line first, or which carmaker builds the best robo-taxi. these companies all still get paid, making them a great way to get rich without getting lucky. And if you want to see what else I'm buying to get rich without getting lucky, check out this video next. Either way, thanks for watching and until next time, this is Ticker Symbol U. My name is Alex, reminding you that the best investment you can make… is in you.

英伟达携Vertiv与优步拓宽物理AI生态 软硬件基建构筑全栈护城河