字号·· | 护眼
连线

英伟达在黄仁勋风波之前就已陷入困境Nvidia Was Already in Trouble Before the Jensen Huang Mess

点「原文对照」整页切到原文,或双击某段只看那段的原文。

Nvidia 在很大程度上得益于其最近被罢免的 CEO 黄仁勋,在推动 AI 淘金热方面发挥了关键作用。黄仁勋擅长预见技术变革,比大多数人更早投资用于训练 AI 的芯片。

Nvidia played a key role enabling the AI gold rush thanks in large part to its recently ousted CEO Jensen Huang. Huang was a master of foreseeing technological changes, investing in chips to train AI earlier than most.

但就在他离职之际,公司面临着一份日益增长的竞争对手名单,这些对手拥有一系列新的先进技术,正紧追不舍。对于如今由 CFO 科莱特·克雷斯掌舵的 Nvidia 来说,时机并不理想。

But just as he’s leaving, the company has a growing list of competitors with a range of new, advanced technology breathing down its neck. For Nvidia, where CFO Colette Kress has now taken the reins, the timing isn’t great.

“今天售出的一切都是两到三年前设计的,”一家大型 AI 公司的 CEO 表示,该人士因不愿损害商业关系,接受《连线》采访时要求匿名。这位 CEO 称,克雷斯或其永久继任者将不得不规划芯片行业两年后将走的路线。竞争对手已经开始行动。

“Everything sold today was designed two-to-three years ago,” says the CEO of a large AI company, who spoke to WIRED on condition of anonymity to avoid hurting business relationships. Kress, or whoever replaces her permanently, will have to chart the course that the chip industry will take two years from now, the CEO says. Already, competitors have started to move in.

根据 Silicon Analysis 的数据,Nvidia 占据 AI 加速器市场 80% 的销售额,处于主导地位。老对手 AMD 正在全力打造有竞争力的 AI 芯片。其市场份额已从 2022 年的仅 2% 爬升至 2026 年的 7% 至 10% 之间。

Nvidia dominates the market for AI accelerators, capturing 80 percent of sales, according to Silicon Analysis. Longtime nemesis AMD is now making a concerted effort to build competitive AI chips. Its share of the market has crept up from just two percent in 2022 to between 7 and 10 percent in 2026.

美国大型云服务商——Google、Amazon 和 Microsoft——都在开发更专用的芯片,为 AI 模型的训练和推理提供动力。它们都取得了类似的增长。这些公司设计自有芯片,旨在降低成本并减少对 Nvidia 的依赖。

Big US cloud players—Google, Amazon, and Microsoft—are all developing more specialized chips to power the training and serving of AI models. Each has seen similar gains. These companies are designing their own silicon to reduce costs and to gain some independence from Nvidia.

多家初创公司现在提供更专用的芯片,承诺优于 Nvidia 的核心产品。例如,Cerebras 销售巨型芯片,旨在为巨型 AI 模型提供动力。AMD 最近收购了 Taalas,该公司将模型权重烧录进硅片,使计算效率极高。不过,使用这些芯片需要重写大量代码,才能使其与前沿 AI 模型配合工作。

Several startups now offer more specialized chips that promise improvements over Nvidia’s core products. Cerebras, for instance, sells giant ones meant to power giant AI models. AMD recently bought Taalas, a company that burns models’ weights into silicon to make computations ultra efficient. Using these chips requires rewriting a lot of code, however, to make them work with cutting-edge AI models.

跟踪芯片行业的公司 MoorInStrat 首席分析师 Patrick Moorhead 表示,AI 智能体的兴起使得推理——即运行已经训练好的模型——变得更加重要。这是一个英伟达相比云超大规模厂商和初创公司优势较小的领域。不过,英伟达确实在 2025 年 12 月收购了一家名为 Grok 的专用芯片初创公司。

Patrick Moorhead, chief analyst at MoorInStrat, a company that tracks the chip industry, says the rise of AI agents makes inference, or running models that have already been trained, more important. That’s an area where Nvidia has less of an edge compared to cloud hyperscalers and startups. Nvidia did, however, acquire a startup making specialized chips, called Grok, in December of 2025.

Moorhead 说,由于 AI 助手和智能体的兴起,对通用中央处理器(CPU)的需求也在增长,这是英伟达几乎不涉足的领域。

And because of the rise in AI assistants and agents, there’s also growing demand for general purpose central processing units (CPUs), Moorehead said, an area where Nvidia barely competes.

得益于黄仁勋的远见,英伟达在 AI 领域仍保持巨大领先优势。黄仁勋于 1993 年联合创立公司,旨在为 PC 设计图形芯片。2010 年代中期,他看到了 AI 的前景。他资助相关学术研究、开发支持 AI 工程的软件,并向实验室和初创公司提供硬件。

Nvidia still maintains a huge lead in AI thanks to Huang’s foresight. Huang cofounded the company in 1993 to design graphics chips for PCs. In the mid-2010s, he saw the promise of AI. He funded academic research on the topic, developed software to support AI engineering, and gave hardware to labs and startups.

2022 年,ChatGPT 让 AI 走向大众,英伟达芯片销量飙升。公司营收从 2022 年 1 月的 269.1 亿美元增至 2026 年 1 月的 2159.4 亿美元。它面临的挑战可能不构成生存威胁——但它们累积起来影响不小,尤其是在全球层面。

In 2022, ChatGPT took AI mainstream and sales of Nvidia chips skyrocketed. The company’s revenue went from $26.91 billion in January 2022 to $215.94 billion in January 2026. The challenges it faces may not be existential—but they do add up, especially on a global scale.

中国挑战 当英伟达在美国如日中天时,它在全球另一个 AI 大市场中国的命运却在下滑。华盛顿试图限制获取其最先进芯片,而北京试图发展自己的芯片产业。英伟达不得向中国出售其最强大的芯片,但可以出售上一代产品。根据其文件披露,这些芯片以及非 AI 芯片的销量在过去一年翻倍,但中国现在仅占公司总营收的 8%,而 2024 年为 20%。目前仍不清楚出口到其他国家的芯片中有多少最终流入中国,或被中国公司远程使用。

China Challenge While Nvidia’s fortunes have risen in the US, they have declined in the world’s other big market for AI, China. Washington is trying to restrict access to its most advanced chips, while Beijing is trying to develop its own chip industry. Nvidia is not allowed to sell its most powerful chips in China, but it can sell previous generations. Sales of these chips as well as non AI chips doubled over the past year, but China now accounts for only eight percent of the company’s overall revenue, compared with 20 percent in 2024, according to its fillings. It remains unclear how many chips exported to other countries may either end up in China or end up being used remotely by Chinese firms.

中国科技巨头华为因遭受美国多年制裁,已开发出新型超级计算机,能够调用比常规更多的芯片来训练和运行 AI 模型。一些中国 AI 公司,如制造出全球最优秀开源模型之一的 Z.ai,表示它们完全使用华为芯片训练模型。如果中国开放权重模型在其他国家获得普及,可能会为中国芯片在海外创造市场。

Huawei, a Chinese tech giant targeted with US sanctions for several years, has developed new kinds of supercomputers that harness many more chips than usual to train and run AI models. Some Chinese AI companies, such as Z.ai, which makes one of the world’s best open source models, say they’ve trained their models entirely using Huawei’s chips. If Chinese open-weight models gain traction in other countries, it could create a market for Chinese chips elsewhere.

黄仁勋与中国的交往在其倒台中扮演了重要角色。他与黄乐(Alice Huang)保持恋爱关系,后者是新加坡公司 Megaspeed 的中国籍高管,美国政府怀疑她是中国国家安全部的资产。据报道,Megaspeed 及其子公司已向东南亚进口了 46 亿至 75 亿美元的英伟达芯片。报道表明,该公司可能已向中国 AI 公司提供这些芯片的使用权。

Huang’s dealings with China played a major role in his downfall. He had a romantic relationship with Huang Le, also known as Alice Huang, a Chinese executive at a Singaporean company called Megaspeed who the US government suspects of being an asset of China’s Ministry of State Security. Megaspeed and its subsidiaries have reportedly imported between $4.6 and $7.5 billion Nvidia chips into southeast Asia. Reports suggest that the company may have provided access to the chips to Chinese AI firms.

专注于技术与地缘政治的咨询公司欧亚集团高级分析师巴巴克·米诺维表示,黄仁勋指出切断中国获取渠道将增强国内竞争对手实力,这一观点是正确的。“黄仁勋采取的是务实路线,他了解竞争,”他说。

Babak Minovi, a senior analyst at the Eurasia Group, a consultancy that focuses on technology and geopolitics, says Huang was right to point out that cutting off access to China would empower domestic rivals. “Huang was taking a pragmatic line, and he knows competition,” he says.

趋势观察 即使在其占据主导地位的市场,英伟达也可能成为自身成功的受害者。使用英伟达芯片构建的 AI 模型正在自动化编程、数据分析和科学研究中有助于芯片设计的那些环节。

Trend Spotting Even in markets that it dominates, Nvidia may turn out to be a victim of its own success. The AI models built using Nvidia’s chip are automating the very elements of coding, data analysis, and scientific research that help design chips.

大多数芯片设计公司使用专门的 AI 模型在处理器上排布组件——这传统上需要深厚的工程专长和经验。OpenAI 和 Anthropic 最近都宣称其最新模型具备从零设计电子产品的能力。

Most chip design firms use specialized AI models to arrange components on processors—something that traditionally required deep engineering expertise and experience. And both OpenAI and Anthropic have recently touted the ability of their newest models to design electronics from scratch.

穆尔黑德表示,向 AI 芯片设计的转变可能不会对英伟达构成巨大威胁,因为它也在乘同一浪潮。“它们也能使用 AI,”他指出。

Moorhead says the shift towards AI chip design may not pose a huge threat to Nvidia because it’s riding the same wave.“They have access to AI too,” he notes.

但 Moorhead 表示,AI 模型如今能快速重写代码以在不同硬件上运行,这可能会成为一个问题。多年来,Nvidia 通过提供软件吸引开发者,使其能轻松为其芯片编写程序和构建模型。其统一计算设备架构(CUDA)长期以来被视为一道护城河,确保 AI 公司持续依赖 Nvidia 芯片。

But Moorhead says the way AI models can now quickly rewrite code to run on different hardware may prove a problem. For years, Nvidia has attracted developers by providing software that makes it easy to write programs and build models for its silicon. Its Compute Unified Device Architecture, or CUDA, has long been seen as a moat to ensure AI companies continue to rely on Nvidia chips.

但 AI 模型现在读写代码的能力已如此之强,以至于能以极快的速度将 AI 模型移植到竞争对手的芯片上。

But AI models are now so good at reading and writing code that they can port an AI model over to a competitor's silicon at speed.

初创公司 Wafer 利用 AI 将模型移植到 AMD 硬件上,并以更低成本提供服务,其 CEO Emilio Andere 自称近几个月收入实现了惊人增长。

Wafer, a startup that uses AI to port models to AMD and then serves them more cheaply, has seen revenue grow spectacularly in recent months, boasts CEO Emilio Andere.

他表示,每当发布一个主要的新开放权重模型,Wafer 都能在数小时内让其在 AMD 硬件上运行,效率至少达到 80%。Andere 称 AI 能力已变得如此强大,以至于他的工程师几乎不需要编辑或审查代码。

Every time a major new open weight model is released Wafer can have it running on AMD hardware at least 80 percent as efficiently within hours, he says. Andere says AI is getting so capable that his engineers barely have to edit or review the code.

资深芯片分析师 Dan Hutchinson 表示,Nvidia 可能正步入其“Intel Swan 时刻”,指的是 2019 年 Intel 以工程为导向的 CEO Brian Krzanich 被其 CFO Bob Swan 取代。(听起来熟悉吗?Krzanich 也因与员工的关系辞职;Huang 据称也有过多次此类关系。)此后,公司重心从技术进步转向销售。官僚主义膨胀,营收暴跌。

Dan Hutchinson, a veteran chip analyst, says Nvidia may be entering its “Intel swan moment,” referring to the 2019 replacement of engineering-focused CEO Brian Krzanich with its CFO Bob Swan. (Sound familiar? Krzanich also resigned after a relationship with an employee; Huang is alleged to have had multiple.) From there, the company’s focus shifted to sales more than technical progress. Bureaucracy ballooned and revenue tanked.

“Nvidia 现在就是一个即将发生的 Intel,”他说。

“Nvidia is now an Intel waiting to happen,” he says.

阅读媒体原文 ↗