内容来源：https://aibusiness.com/generative-ai/nvidia-intros-new-ai-chips-and-open-models

内容总结：

在面临空前规模的市场竞争压力下，人工智能巨头英伟达于本周一发布了六款全新AI芯片及一系列开放模型，展现出其巩固行业领先地位的决心。此次在拉斯维加斯国际消费电子展上推出的Rubin平台，集成了包括Vera CPU、Rubin GPU在内的六款芯片，构成一套完整的AI超级计算机系统。同时发布的开放模型涵盖多个领域：Nemotron系列新增多智能体构建模型，Cosmos模型套件推出面向人形机器人及物理AI应用的“世界基础模型”，并升级了用于自动驾驶的Alpamayo视觉语言推理模型。

行业分析师指出，此次密集发布凸显英伟达正从纯GPU供应商向全栈解决方案提供商转型。高德纳分析师奇拉格·德凯特强调：“英伟达试图传递的核心信息是——AI竞争已超越单纯的GPU性能比拼，进入以‘AI工厂’为标志的基础设施整体架构竞赛。”通过将CPU、GPU、DPU等异构芯片整合至Rubin平台，英伟达旨在为客户提供涵盖模型训练、大规模部署及智能体技术落地的端到端解决方案。

值得关注的是，此次发布的开放模型呈现出高度专业化特征。奥姆迪亚分析师马克·贝丘指出，这种针对特定场景的开放模型策略不同于常见的通用模型路径，能让企业更快地将AI应用于医疗、自动驾驶等垂直领域。富图拉姆研究所分析师布拉德利·希明认为，这印证了2026年AI模型“加速落地”的行业趋势，表明英伟达正从“前沿模型创造者”转向“应用智能构建者”。

然而挑战依然存在。尽管开放模型提供了权重参数和训练方案，但目前企业仍更倾向于使用专有模型。德凯特同时警告，英伟达在AI基础设施领域的深度创新，可能使企业在享受技术红利的同时，面临难以摆脱的供应商依赖风险。在AMD、英特尔、高通等竞争对手加速追赶的背景下，英伟达能否通过这套“全栈组合拳”持续引领生态发展，仍有待市场检验。

中文翻译：

由谷歌云赞助
如何选择首个生成式AI应用场景
要开始应用生成式AI，首先应关注能够优化人类信息交互体验的领域。

英伟达近日推出的新型芯片与模型，既展现了其在AI市场的创新实力，也向客户揭示了避免过度依赖单一供应商的挑战。面对空前激烈的市场竞争，这家AI巨头于本周一发布了六款新型AI芯片及全新开放模型，彰显其保持行业领先地位的决心。

在拉斯维加斯国际消费电子展上，这家AI软硬件提供商推出了包含六款新型芯片的英伟达鲁宾平台，这些芯片共同构成了一套AI超级计算机系统。新发布的开放生成式AI模型包括Nemotron系列中更多用于构建智能体的模型，以及专为人形机器人与实体AI应用设计的Cosmos模型套件中的新型世界基础模型，该套件同时支持合成数据生成。

英伟达还展示了去年12月初发布的自动驾驶汽车模型Alpamayo。高德纳分析师奇拉格·德卡特指出，这一系列发布表明英伟达不仅致力于打造从芯片到软件的全栈解决方案，更试图赋能第三方开发自主的全栈产品。

"他们试图强调，AI已不仅是GPU的竞赛，"德卡特解释道。他提及广泛用于生成式AI模型训练与推理的图形处理器芯片时表示："竞争焦点已超越GPU芯片本身，实质是AI超级计算机的较量。"

英伟达通过鲁宾平台整合多类型AI芯片，证明其不仅是GPU供应商。该平台包含英伟达Vera CPU、鲁宾GPU、NVLink 6交换机、ConnectX-9超级网卡、BlueField-4 DPU及Spectrum 6以太网交换机。

作为广泛应用的黑威尔平台继任者，鲁宾平台采用英伟达NVLink互连技术与多种变压器技术，在智能体AI、高级推理及专家混合模型扩展方面较前代显著提速。德卡特认为，该平台旨在引导客户与受众超越GPU视角，将底层基础设施整体视为"AI工厂"。

"英伟达强调的核心在于，无论进行模型训练还是大规模部署模型——无论是直接应用还是作为智能体技术战略的组成部分——底层基础设施都将面临AI工厂级规模的挑战，"德卡特补充说，这正是英伟达希望为数据中心运营商、超大规模厂商及企业客户解决的问题。

"AI已不仅是简单的设备问题，而是涉及多维度、多形态的复杂体系，"德卡特进一步指出。这种超越GPU的AI视野正是英伟达区别于AMD、英特尔和高通等竞争对手的关键。

"多数竞争者仍难以企及其水平，"他表示，"虽然正在逐步追赶，但尚未达到同等高度。"

此次新模型发布距离Nemotron 3开放模型家族亮相不足一月，该系列专为构建多智能体系统设计。英伟达介绍，Nemotron语音模型包含新型自动语音识别系统，能为实时字幕与语音AI应用提供低延迟识别。同时，Nemotron RAG技术新增嵌入与重排视觉语言模型，英伟达还同步发布了配套数据集、训练资源及架构蓝图。

除Nemotron系列外，英伟达通过Cosmos Reason 2、Cosmos Transfer 2.5及Cosmos Predict扩展了世界基础模型产品线。其中Reason 2作为视觉语言模型，赋能机器人与AI智能体感知物理世界；Transfer 2.5与Predict 2.5则能生成跨环境条件的合成视频。Alpamayo 1模型则是面向自动驾驶汽车的推理视觉语言模型。

Informa TechTarget旗下Omdia分析师马克·贝丘指出，虽然英伟达并非首家发布开放模型的厂商，但其明确标注各模型专属用途的做法独具特色。"这种专业化开放模型的路径并不常见，"贝丘表示，"但确实能帮助客户更快部署应用。"

Futurum市场研究公司分析师布拉德利·希明认为，这种专业化趋势印证了该公司对2026年的预测——相较于通用模型，专业化AI模型将加速落地。"英伟达的发布正是例证，他们针对具体痛点提供解决方案。"

"这些模型正被应用于医疗、自动驾驶等特定领域及企业具体场景，"希明补充道，"他们的目标不仅是成为顶尖前沿模型制造商，更要成为最佳应用智能制造商。"

然而贝丘指出，尽管这些模型完全开放且英伟达公布了权重参数与构建方案，企业落地仍存障碍。"目前企业使用专有模型的比例仍远高于开源模型。"

德卡特提出另一重挑战：英伟达在模型与AI基础设施市场的创新，可能加剧企业对供应商的依赖。"新型芯片与模型既彰显创新实力，也向客户揭示了避免供应商锁定的难题。"

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Both the new chips and models demonstrate how Nvidia is innovating within the AI market, while also highlighting for customers the challenge of avoiding dependency on the vendor.
As its market dominance is being challenged by a host of competitors on a scale not seen before, Nvidia on Monday rolled out six new AI chips and new open models, which appear to show the AI giant is determined to stay ahead of the field.
At the CES consumer electronics show in Las Vegas, the AI hardware and software provider released the Nvidia Rubin platform, a system comprising the six new chips that collectively form an AI supercomputer. The new open generative AI models include more agent-building models in the Nemotron family and new World Foundation Models in the Cosmos model suite, designed for use with humanoid robots and other physical AI applications, as well as for generating synthetic data.
Nvidia also showcased a model that powers autonomous vehicles, Nvidia Alpamayo, which the vendor previously released, in early December.
The slew of releases demonstrates how Nvidia is not only pursuing a full-stack approach from chips to software but also looking to enable third parties to develop their own full-stack products, said Chirag Dekate, an analyst at Gartner.
"What they're trying to highlight here is that AI is no longer just a GPU game," Dekate said, referring to the ubiquitous graphics processing unit chips that power training and inference for generative AI models. "It is no longer about the GPU chip; it is actually an AI supercomputer.
One way Nvidia is showing that it is not solely a GPU vendor is by combining diverse types of AI chips within the Rubin platform. The components include the Nvidia Vera CPU, Nvidia Rubin GPU, Nvidia NVLink 6 Switch, Nvidia ConnectX-9 SuperNIC, Nvidia BlueField-4 DPU, and Nvidia Spectrum 6 Ethernet Switch.
The Rubin platform is the successor to the widely used Nvidia Blackwell platform. The Rubin platform uses Nvidia's NVLink interconnect technology and various transformer technologies to accelerate agentic AI, advanced reasoning, and the scale of mixture-of-experts models, compared to Blackwell.
With the platform, Nvidia is trying to inspire its customers and audience to look beyond just GPUs and see the whole underlying infrastructure component as more of an AI factory, Dekate said.
"What Nvidia is trying to highlight is whether you're trying to solve a problem in the context of model training, or if you're trying to deploy models at scale, either directly or as part of your agent tech strategy, the underlying infrastructure is likely going to be an AI factory scale problem," Dekate said. He added that this is a problem that Nvidia wants to address for data center operators, hyperscalers, and enterprise clients.
"AI is no longer just a small, simple device issue; it is actually multifaceted and multi-form factor," Dekate continued.
This focus on AI as more than just a GPU is part of what differentiates Nvidia from its competitors, he added. Competitors include AMD, Intel and Qualcomm.
"Many of the competitors struggle to meet them," he said. "They're starting to get there, but they're not there yet."
The new models arrive less than a month after the release of the Nemotron 3 family of open models, designed to build and implement multi-agent systems. They include Nemotron speech, which includes a new automatic speech recognition model that provides real-time, low-latency speech recognition for live caption and speech AI applications, Nvidia said. Also, Nemotron RAG technology has new embed and rerank vision language models. Nvidia also released datasets, training resources and blueprints for the models.
In addition to Nemotron, Nvidia expanded the World Foundation Model line with Cosmos Reason 2, Cosmos Transfer 2.5, and Cosmos Predict. Cosmos Reason 2 is a vision language model that enables robots and AI agents to interact with and understand the physical world. Transfer 2.5 and Predict 2.5 generate synthetic videos across different environments and conditions.
The Alpamayo 1 model is a reasoning vision language model for autonomous vehicles.
While Nvidia is not the first vendor to release open models, the way it specifies what each model is for is unique, said Mark Beccue, an analyst at Omdia, a division of Informa TechTarget.
"This is a little different," Beccue said, specialized open models are not a common approach. However, specializing open models makes sense because it enables customers with open models to start using them faster, Beccue said.
The specialization of the open models confirms one of the trends the Futurum market research firm has identified for 2026, which is faster implementation of AI models, as opposed to generalized models, said Bradley Shimmin, an analyst at Futurum.
"You can see that in what Nvidia is rolling out," Shimmin said. "They're tackling particular problems."
"They're applying those models within specific domains like healthcare, autonomous vehicles and very specific use cases in the enterprises," Shimmin added. "What they're doing is not just trying to be the best frontier model maker, but to be the best applied intelligence maker."
However, despite these models being open and Nvidia releasing the weights and recipes for them, enterprise adoption is still a challenge, Beccue said.
"Companies are still using proprietary models more than they are using open source right now," he said.
Another challenge is that Nvidia's innovation in the model and AI infrastructure market will make it harder for enterprises to avoid being dependent on the vendor, Dekate said.