NVIDIA Emerges as the Biggest Winner in the AI Era: How Long Can Its Dominance Last? (AI Translation)
Listen to the full version
文｜财新周刊 翟少辉
By Caixin Weekly's Zhai Shaohui
美国当地时间3月18日，英伟达一年一度的GPU（图形处理器）技术大会（GTC）时隔五年重回线下，线下参会者从2019年的约8000人翻倍至1.6万人，线上参会者高达约30万人。“以前只有苹果的发布会能有如此派头。”一名常驻硅谷的科技行业人士对财新感叹道，如今英伟达已是硅谷“顶流”。
On March 18, local time in the United States, Nvidia's annual GPU (Graphics Processing Unit) Technology Conference (GTC) returned to an in-person format for the first time in five years. The number of attendees on-site doubled from approximately 8,000 in 2019 to 16,000, with an additional roughly 300,000 participants online. "In the past, only Apple's product launches could achieve such grandeur," a tech industry insider based in Silicon Valley told Caixin, noting that Nvidia has now become a top trendsetter in Silicon Valley.
因为报名情况超预期，英伟达不得不在位于圣何塞会议中心的GTC主会场之外，单独为英伟达创始人兼CEO黄仁勋另寻演讲场馆，最终落定SAP中心体育馆，这里是北美职业冰球联盟（NHL）球队圣何塞鲨鱼队的主场，也是硅谷最主要的演唱会举办地之一。“我希望你们意识到，这不是音乐会，而是一场开发者大会。”当地时间13时，黄仁勋身着标志性的黑色皮衣登台，在长达两个小时的演讲中不断抛出重磅技术和产品，现场掌声热烈应和；数个小时后，他的演讲内容和视频就被解读、裁切成各种版本，在中国的社交网络广泛流转。
Due to an unexpectedly high number of registrations, Nvidia was compelled to find an alternative venue for its founder and CEO, Jensen Huang, outside the main GTC venue at the San Jose Convention Center. The final choice was the SAP Center Arena, home to the NHL's San Jose Sharks and one of Silicon Valley's primary concert venues. "I hope you realize that this is not a concert but a developers' conference," Huang stated as he took the stage in his signature black leather jacket at 1 p.m. local time. Throughout his two-hour speech, he unveiled significant new technologies and products, eliciting enthusiastic applause from the audience. Hours later, his presentation and videos were interpreted and edited into various versions, circulating widely on Chinese social media platforms.
此次英伟达发布的新一代AI芯片架构Blackwell，性能为前一代Hopper架构的2.5—5倍。相较以往架构迭代时注重单颗GPU性能增幅，Blackwell的重心落在“超级芯片”、平台系统层面，更加强调组合拳效果，并在软件层面升级策略，力图打造服务各行各业的“AI代工厂”（AI Foundry）。英伟达还亲自下场，强化其元宇宙平台Omniverse在工业、仓储等领域的应用，打造人工智能（AI）的垂直落地场景。比如，以Omniverse的虚拟仿真能力为中层支点，携上层基础模型和底层芯片，入局近年火热的人形机器人赛道。
NVIDIA has unveiled its latest AI chip architecture, Blackwell, boasting a performance that is 2.5 to 5 times greater than its predecessor, the Hopper architecture. Unlike previous iterations which focused on enhancing the performance of individual GPUs, Blackwell shifts the focus towards "super chips" and platform-level systems, emphasizing a combined effect. It also upgrades its strategy at the software level in an effort to establish an "AI Foundry" that serves various industries. NVIDIA is also taking a hands-on approach to strengthen the application of its metaverse platform Omniverse in areas such as industry and warehousing, aiming to create vertical scenarios for artificial intelligence (AI) implementation. For example, leveraging Omniverse's virtual simulation capabilities as a mid-layer support, combined with upper-layer base models and lower-layer chips, NVIDIA is entering the highly popular field of humanoid robots in recent years.
- DIGEST HUB
Digest Hub
Back
- Nvidia's annual GPU Technology Conference (GTC) in 2023 marked a significant return to in-person attendance, with the unveiling of its new AI chip architecture, Blackwell, which significantly outperforms its predecessor. The company's market capitalization is nearing $2.3 trillion, surpassing major tech giants and reflecting its dominant position in the AI chip market with an 85% share.
- Amidst a global surge in demand for generative AI following the launch of ChatGPT by OpenAI, Nvidia has become essential for large model training, with no substitutes for its high-end GPUs. This demand has led to a remarkable financial performance for Nvidia, with revenues reaching $60.922 billion and net profits hitting $29.76 billion for the fiscal year ending January 28, 2024.
- Despite Nvidia's dominance, competitors like AMD and Intel are making strides with new chips and ecosystems to challenge Nvidia's market position. Additionally, geopolitical tensions have impacted Nvidia's operations in China, prompting Chinese tech companies to invest in domestic AI chip solutions like Huawei's Ascend series.
Nvidia's annual GPU Technology Conference (GTC) in the United States marked a significant return to an in-person format after five years, drawing double the attendees from 2019 and highlighting Nvidia's prominence in Silicon Valley [para. 1]. The conference showcased the unveiling of Nvidia's new AI chip architecture, Blackwell, which significantly outperforms its predecessor and shifts focus towards "super chips" and platform-level systems. This move is part of Nvidia's broader strategy to establish an "AI Foundry" catering to various industries and strengthen its metaverse platform Omniverse for AI implementation [para. 2].
The market has responded positively to Nvidia's innovations, with major financial institutions raising their price targets for the company. Nvidia's market capitalization has soared, surpassing major tech giants and positioning it as a leading force in the tech industry. This surge comes as generative AI gains momentum globally, with Nvidia emerging as a crucial player due to its indispensable high-end GPUs for large model training [para. 3][para. 4].
Nvidia dominates the AI chip market share in 2023, capturing 85% of it. Its revenues and net profits have seen astronomical increases year-over-year, reflecting its strong performance amidst a booming demand for AI computing power. Despite this success, there are questions about the valuation rationality and future sustainability of Nvidia’s market position [para. 5][para. 6].
Competitors like AMD and Intel are stepping up their efforts to challenge Nvidia by launching new chips and developing platforms aimed at countering Nvidia's CUDA software ecosystem. New entrants like Cerebras and Groq are also making bold claims about their AI chips' capabilities. Amidst this competition, some of Nvidia’s customers are contemplating developing their own solutions to mitigate costs and supply shortages [para. 7][para. 8].
In response to these challenges, especially from companies developing proprietary AI chips, Nvidia is accelerating its product release cycle and focusing on maintaining its software ecosystem advantage. However, U.S. government chip control policies have posed significant hurdles for Nvidia in China, prompting it to develop specially tailored versions of its GPUs for the Chinese market [para. 9][para. 10].
Despite these challenges, industry insiders suggest that entities investing heavily in R&D stand a better chance against Nvidia’s dominance. Counterpoint predicts that by 2024, Nvidia’s data center business revenue will exceed $72 billion [para. 11]. Jensen Huang aims to position Nvidia as the "TSMC of the AI Industry," leveraging breakthroughs such as ChatGPT and advancements in deep learning supercomputers like DGX-1 [para. 12].
Nvidia’s newly released Blackwell architecture represents a leap forward in efficiency for training large models with fewer resources required compared to previous generations. The company is also focusing on providing comprehensive AI cloud services and enterprise-level generative AI microservices to simplify deploying AI systems for customers [para. 13].
Amid global competition and technological conflicts compounded by geopolitical tensions, China seeks domestic alternatives like Huawei's Ascend chips due to U.S. export bans on high-end GPUs. This shift is driven by both necessity due to regulatory constraints and a desire for more autonomy in critical technology sectors [para. 14][para. 15].
As competition intensifies both globally and within China’s burgeoning tech landscape, companies are exploring various strategies—from developing proprietary technologies to forming strategic partnerships—to secure their place in the rapidly evolving world of artificial intelligence and computational power bases [para. 16].
- 2006：
- Jensen Huang launched CUDA, a unified computing architecture for his GPUs.
- 2012：
- A team from the University of Toronto used an NVIDIA gaming GPU to train their neural network, AlexNet, firmly linking GPUs with AI as its perfect application.
- 2016：
- NVIDIA introduced its first deep learning supercomputer, DGX.
- 2019：
- NVIDIA unveiled its latest AI chip architecture, Blackwell, boasting a performance 2.5 to 5 times greater than its predecessor, the Hopper architecture.
- November 2022：
- The U.S. technology company OpenAI released its chatbot ChatGPT, sparking a global sensation and initiating a new wave of generative AI.
- December 2023：
- AMD released its MI300 series chips, directly challenging Nvidia's flagship product at the time, the H100.
- March 18, 2024：
- Nvidia's annual GPU (Graphics Processing Unit) Technology Conference (GTC) made a return to an in-person format for the first time in five years.
- March 19, 2024：
- Cerebras, an American AI chip company founded in 2016, held its AI Day event just a ten-minute walk from the GTC venue.
- PODCAST
- MOST POPULAR
Sign in with Google
Sign in with Facebook
Sign in with 财新