Nvidia’s AI advantage is moving beyond the GPU

英伟达的 AI 优势正超越 GPU 本身

Before this week, the dominant story about Nvidia went something like this: For the first few years of the AI boom, Nvidia was the only source for state-of-the-art GPUs, which became immensely profitable as the industry scaled out. In the last few years, hyperscalers like Amazon and Google have started building their own chips, and Nvidia is no longer the only game in town, leading many investors to wonder how durable its advantage really is. 在本周之前，关于英伟达的主流叙事大致如下：在 AI 热潮的最初几年，英伟达是尖端 GPU 的唯一来源，随着行业规模的扩大，这些 GPU 变得极其有利可图。在过去几年里，亚马逊和谷歌等超大规模云服务商开始制造自己的芯片，英伟达不再是市场上的唯一玩家，这导致许多投资者怀疑其优势究竟能持续多久。

It’s a compelling story, and mostly true. After growing its market cap 10x between the start of 2023 and mid-2025, Nvidia shares have been on a more modest trajectory for the past year, driven by concerns about GPU competition. A new narrative has taken shape since the company’s earnings on Wednesday and investors are starting to realize that Nvidia’s advantage goes far beyond GPUs. 这是一个引人入胜的故事，而且基本属实。在 2023 年初至 2025 年年中市值增长 10 倍之后，受 GPU 竞争担忧的影响，英伟达股价在过去一年中走势较为平缓。然而，自周三公司发布财报以来，一种新的叙事已经形成，投资者开始意识到英伟达的优势远不止于 GPU。

As AI’s compute grows into the gigawatt scale, orchestration has become an increasingly complex task. Not surprisingly, Nvidia has built much of the state-of-the-art hardware needed to handle it, giving the company a huge advantage in the systems that surround the GPU even as it sees increased competition on the GPUs themselves. For all the talk of compute as a commodity, it’s still incredibly difficult to operate a megascale data center at peak efficiency — and as deployments get bigger and faster, that challenge is only growing. 随着 AI 计算规模增长至吉瓦（GW）级别，任务编排已成为一项日益复杂的任务。不出所料，英伟达构建了处理这些任务所需的大部分尖端硬件，这使其在 GPU 周边的系统领域占据了巨大优势，即使在 GPU 本身面临日益激烈的竞争时也是如此。尽管人们常说计算能力已成为一种商品，但要以最高效率运行超大规模数据中心仍然极其困难——而且随着部署规模越来越大、速度越来越快，这一挑战只会与日俱增。

Rack by Rack

机架与机架之间

You can see some of this just by looking at the details of what Nvidia is actually selling. The company is currently rolling out its Vera Rubin architecture, which pairs the Rubin GPU with a collection of other units, including the Vera CPU, the Groq 3 LPX inference accelerator and similar racks for storage and networking. 只需看看英伟达实际销售的产品细节，就能看出端倪。该公司目前正在推出其 Vera Rubin 架构，该架构将 Rubin GPU 与一系列其他单元配对，包括 Vera CPU、Groq 3 LPX 推理加速器以及用于存储和网络的类似机架。

Over the past week, I’ve been talking to folks at Nvidia about what those systems actually do, and the results have been surprising. Like the Rubin GPU itself, they’re extremely specialized systems, but instead of churning through tokens, they’re making sure everything outside the GPU works as efficiently as possible. If the GPU is the engine, these are the rest of the car. 在过去的一周里，我一直在与英伟达的工作人员交流这些系统的实际功能，结果令人惊讶。就像 Rubin GPU 本身一样，它们是非常专业的系统，但它们不是在处理 Token，而是确保 GPU 之外的一切都能尽可能高效地运行。如果说 GPU 是引擎，那么这些系统就是汽车的其他部分。

The Vera CPU in particular is focused on the problem of orchestrating data. “Vera is important because there’s only so much memory that you can put in a single server or any sort of compute platform,” Jason Hardy, Nvidia’s VP of storage technology, told me. As data centers have scaled up computing power, memory capacity has scaled up too, which is why companies like Micron have gotten rich in the second wave of the infrastructure boom. 特别是 Vera CPU，它专注于数据编排问题。英伟达存储技术副总裁 Jason Hardy 告诉我：“Vera 之所以重要，是因为单个服务器或任何计算平台所能容纳的内存是有限的。”随着数据中心计算能力的提升，内存容量也随之扩大，这就是为什么像美光（Micron）这样的公司能在基础设施繁荣的第二波浪潮中获利丰厚的原因。

But getting that data to the GPU at the right time isn’t straightforward — and as companies look to drive tokens-per-watt lower and lower, they’re realizing how important that kind of traffic direction is. “We saw upwards of 3x improvement in these operations, where the Vera CPU is allowing for acceleration,” Hardy said. “So now we can use our flash to its fullest potential, because we can get all that performance out of it without bottlenecking.” 但在正确的时间将数据传输到 GPU 并非易事——随着公司寻求不断降低每瓦特 Token 的能耗，他们意识到这种流量调度有多么重要。Hardy 说：“我们看到这些操作的性能提升了 3 倍以上，这得益于 Vera CPU 带来的加速。因此，我们现在可以充分发挥闪存的潜力，因为我们可以在不造成瓶颈的情况下榨干其所有性能。”

You can see versions of the same problem outside of Nvidia. When OpenAI developed its Jalapeño chip, a major focus was avoiding these challenges entirely by minimizing the amount of data that needs to be moved around. “We designed Jalapeño to minimize data movement and communication delays,” the company said in a blog post earlier this month. “Its large domain allows the entire workload to remain within one connected system, minimizing data movement and helping the complete request stay fast and efficient from beginning to end.” 在英伟达之外，你也能看到类似的问题。当 OpenAI 开发其 Jalapeño 芯片时，一个主要重点是通过最大限度地减少需要移动的数据量来完全避免这些挑战。该公司在本月初的一篇博客文章中表示：“我们设计 Jalapeño 是为了最大限度地减少数据移动和通信延迟。其巨大的域允许整个工作负载保留在一个连接的系统中，从而最大限度地减少数据移动，并帮助整个请求从头到尾保持快速和高效。”

It’s a different approach, avoiding data movement entirely by conducting a workload within one integrated chip. But the overall logic is the same, increasing efficiency with smarter traffic control instead of just more processor cycles. That in turn opens up a whole new layer of infrastructure for companies to compete over. 这是一种不同的方法，通过在单个集成芯片内执行工作负载来完全避免数据移动。但总体逻辑是一样的，即通过更智能的流量控制而不是仅仅增加处理器周期来提高效率。这反过来又为公司开辟了一个全新的基础设施竞争层面。

This new focus on data orchestration isn’t automatically a win for Nvidia. The company will have to compete with rival chipmakers and hyperscalers just as it has with GPUs. But the competition has moved to a new layer, where building a rival GPU matters less than being able to make the entire system work efficiently. And at least in the early stages, Nvidia looks to have a commanding lead. 这种对数据编排的新关注并不能自动保证英伟达的胜利。该公司将不得不像在 GPU 领域一样，与竞争对手芯片制造商和超大规模云服务商展开竞争。但竞争已经转移到了一个新的层面，在这个层面上，制造一个竞争对手的 GPU 远不如能够让整个系统高效运行重要。至少在早期阶段，英伟达似乎拥有绝对的领先优势。