11-19 更新 实时
首页/ 算力/ 正文

黄仁勋风波之前,Nvidia 就已陷入困境

算力 来源:Wired 2026-11-16 19:00 0 阅读 机器翻译
黄仁勋风波之前,Nvidia 就已陷入困境

但就在他即将离任之际,公司面对的竞争对手名单越来越长,它们手握一系列新的先进技术,正步步紧逼。对 NVIDIA 而言——如今由 CFO Colette Kress 接掌帅印——这个时机并不理想。

“今天卖出的所有东西,都是两三年前设计的,”一家大型 AI 公司的 CEO 说。他因担心损害商业关系,要求匿名接受 WIRED 采访。这位 CEO 表示,Kress——或者最终正式接替她的人——必须为芯片产业规划出两年后的走向。

而竞争对手已经开始入场。

据 Silicon Analysis 的数据,NVIDIA 主导着 AI 加速器市场,占据 80% 的销售额。长期宿敌 AMD 如今正齐心协力打造有竞争力的 AI 芯片。其市场份额已从 2022 年的区区 2% 攀升至 2026 年的 7% 到 10%。

美国大型云厂商——Google、Amazon 和 Microsoft——都在开发更专用的芯片,用于支撑 AI 模型的训练与服务。每家都取得了类似的进展。这些公司设计自研芯片,是为了降低成本,并在一定程度上摆脱对 NVIDIA 的依赖。

如今有若干初创公司提供更专用的芯片,号称比 NVIDIA 的核心产品更胜一筹。例如 Cerebras 出售的超大芯片,旨在支撑超大规模 AI 模型。AMD 最近收购了 Taalas,这家公司把模型权重烧录进硅片,以实现超高的计算效率。不过,使用这些芯片需要重写大量代码,才能让它们与前沿 AI 模型配合工作。

追踪芯片产业的机构 MoorInStrat 首席分析师 Patrick Moorhead 表示,AI 智能体的兴起让推理——也就是运行已经训练好的模型——变得更加重要。而在这一领域,Nvidia 相比超大规模云服务商和初创公司优势较小。不过,Nvidia 确实在 2025 年 12 月收购了一家名为 Grok 的专用芯片初创公司。

Moorhead 表示,由于 AI 助手和智能体的兴起,市场对通用中央处理器(CPU)的需求也在增长,而这一领域 Nvidia 几乎不参与竞争。

得益于黄仁勋的前瞻性判断,Nvidia 在 AI 领域仍保持着巨大领先优势。黄仁勋于 1993 年联合创办该公司,最初为 PC 设计图形芯片。在 2010 年代中期,他看到了 AI 的前景。他资助了该领域的学术研究,开发了支持 AI 工程的软件,并向实验室和初创公司提供硬件。

2022 年,ChatGPT 让 AI 走向主流,Nvidia 芯片的销量随之飙升。公司营收从 2022 年 1 月的 269.1 亿美元增长到 2026 年 1 月的 2159.4 亿美元。它面临的挑战或许并非关乎存亡,但确实在不断累积,尤其是在全球范围内。

Nvidia 在美国的运势一路走高,但在全球另一个 AI 大市场——中国——却在走下坡路。华盛顿正试图限制中国获取其最先进的芯片,而北京则在努力发展自己的芯片产业。Nvidia 不被允许在中国销售其最强大的芯片,但可以销售前几代产品。过去一年,这些芯片以及非 AI 芯片的销售额翻了一番,但根据其提交的文件,中国目前仅占该公司总营收的 8%,而 2024 年这一比例为 20%。目前仍不清楚,有多少出口到其他国家的芯片最终流入中国,或被中国企业远程使用。

华为——这家多年来一直是美国制裁目标的中国科技巨头——已开发出新型超级计算机,其调用的芯片数量远超常规,用于训练和运行 AI 模型。一些中国 AI 公司表示,它们完全使用华为的芯片训练了自己的模型,例如打造了全球最佳开源模型之一的 Z.ai。如果中国的开放权重模型在其他国家获得认可,就可能在其他地区为中国芯片创造市场。

Huang 与中国方面的往来,是他下台的重要原因之一。他与黄乐(又称 Alice Huang)存在恋爱关系,黄乐是新加坡公司 Megaspeed 的一名高管,美国政府怀疑她为中国国家安全部效力。据报道,Megaspeed 及其子公司已向东南亚进口了价值 46 亿至 75 亿美元的 Nvidia 芯片。有报道称,该公司可能让中国 AI 公司获得了这些芯片的使用权。

专注于技术与地缘政治的咨询机构欧亚集团(Eurasia Group)高级分析师 Babak Minovi 表示,黄仁勋指出切断对中国的供应会壮大中国本土竞争对手,这一点说得没错。“黄仁勋走的是务实路线,他懂竞争。”他说。

即便在自身占据主导的市场中,NVIDIA 也可能成为自身成功的受害者。用 NVIDIA 芯片构建的 AI 模型,正把编码、数据分析和科学研究中那些有助于设计芯片的环节自动化。

多数芯片设计公司使用专门的 AI 模型来在处理器上排布元件——这在传统上需要深厚的工程专业知识和经验。而 OpenAI 和 Anthropic 最近都大力宣扬其最新模型具备从零开始设计电子器件的能力。

Moorhead 表示,向 AI 芯片设计转变可能不会对 NVIDIA 构成巨大威胁,因为它正乘着同一股浪潮。“他们也能用上 AI。”他指出。

但 Moorhead 说,AI 模型如今能迅速重写代码以在不同硬件上运行,这或许会成为一个问题。多年来,NVIDIA 通过提供便于为其芯片编写程序、构建模型的软件来吸引开发者。其统一计算设备架构(CUDA)长期被视为一道护城河,确保 AI 公司持续依赖 NVIDIA 芯片。

但如今的 AI 模型在读写代码方面已非常出色,能够迅速把 AI 模型移植到竞争对手的芯片上。

初创公司 Wafer 利用 AI 将模型移植到 AMD 平台,然后以更低成本提供服务。其 CEO Emilio Andere 自豪地表示,近几个月该公司营收增长惊人。

他说,每当有重要的新开放权重模型发布,Wafer 都能在数小时内让它在 AMD 硬件上以至少 80% 的效率运行。Andere 表示,AI 的能力已经强到他的工程师几乎不需要编辑或审查代码。

资深芯片分析师 Dan Hutchinson 表示,NVIDIA 可能正迎来它的“Intel 天鹅时刻”,指的是 2019 年 Intel 用 CFO Bob Swan 取代了专注工程的 CEO Brian Krzanich。(听起来耳熟吗?Krzanich 也是在与一名员工发生关系后辞职;黄仁勋据称有多段类似关系。)自那之后,Intel 的重心便从技术进步转向销售。官僚体系膨胀,营收暴跌。

“NVIDIA 如今就是一个等着变成 Intel 的公司,”他说。

Airbnb 房东们怒不可遏:他们再也不能让房客干家务了

大新闻:酷小孩都用全小写打字——这让你生气吗?

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Nvidia Was Already in Trouble Before the Jensen Huang Mess)

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.

“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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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

Airbnb hosts are furious they can no longer make guests do their chores

big story: cool kids write in all lowercase—does that make you mad?

Will Knight is a senior writer for WIRED, covering artificial intelligence. He writes the AI Lab newsletter, a weekly dispatch from beyond the cutting edge of AI—sign up here. He was previously a senior editor at MIT Technology Review, where he wrote about fundamental advances in AI and China’s AI ... Read More

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 Wired(Nvidia Was Already in Trouble Before the Jensen Huang Mess),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗