Nvidia 的 CUDA 面临 AI 编程 Agent 带来的新威胁
英文标题：🔶 Nvidia's CUDA Faces New Threats from AI Coding Agents
长期以来，CUDA 一直是 Nvidia 在 AI 芯片市场最重要的软件护城河，但 AI 编程 Agent 正在降低开发底层芯片软件的门槛。Infinity 创始人 Jeremy Nixon 称，其团队借助 AI 编程 Agent 仅用 10 小时便为 D-Matrix 复现了类似 CUDA 的软件；云巨头及 OpenAI、Anthropic 等也在推动替代性软硬件生态。随着 AI 从训练转向推理，企业更关注成本、专用芯片以及跨芯片运行的软件能力，这可能削弱 CUDA 的锁定效应。另一方面，业内也认为代码生成后的验证、优化和生产部署仍是关键瓶颈，而 Nvidia 的验证工具与完整生态可能反而成为其下一道护城河。
20.8 KB
128.3 KB
80.6 KB
17.3 KB
16.2 KB
26.6 KB
15.2 KB
13.0 KB
13.0 KB
13.6 KB
3.0 KB
8.8 KB
5.1 KB
1.8 KB
817 B
2.2 KB
1.1 KB
11.7 KB
2.3 KB
2.1 KB
中文
Nvidia 最大的竞争优势，已不再像过去那样牢不可破。
过去二十年里，Nvidia 最珍贵的资产不仅是芯片，更是将芯片转化为 AI 基础构件的软件——CUDA。
统一计算设备架构（Compute Unified Device Architecture，CUDA）由 Nvidia 资深高管 Ian Buck 主导开发，他目前负责高性能计算。CUDA 历经多年建设，包含面向常见 AI 任务的现成代码、用于发现 Bug 的工具，以及让数千颗芯片协同训练模型的软件。
如今，一些人认为，AI 最终可能自动化行业中最困难的工作之一：构建为 AI 自身提供支持的软件。
前 Google Brain 研究员、AI 软件创业公司 Infinity 创始人 Jeremy Nixon 表示，行业正站在一个重要门槛上。他告诉 Business Insider，其创业公司使用 AI 编程 Agent，仅用 10 小时就为芯片创业公司 D-Matrix复现了类似 CUDA 的软件——他认为，这证明 Nvidia 最重要的一道护城河之一正在被跨越。
Infinity 创始人兼 CEO Jeremy Nixon。Infinity.inc
压力并不只来自创业公司。Google、Amazon 和 Microsoft 等云计算巨头多年来一直围绕自有 AI 芯片开发软件，而 OpenAI 与 Anthropic 最近也展示了能够生成系统软件的 AI 模型。
DeepSeek 创始人梁文锋最近表示，编程 Agent与其创业公司自研编程语言 TileLang，已让 AI 软件开发变得容易得多。
编程 Agent 不仅在帮助挑战者。
Nvidia 表示，开发者正越来越多地使用 CUDA 代码库构建 AI 应用；Nvidia 开发者生态系统副总裁 Ankit Patel 称，公司也“使用 AI 编程 Agent 加快 CUDA 开发速度，并以更大规模进行验证”。
推理可能改变 CUDA 的竞争方程式
如果 CUDA 的第一项优势是软件，那么第二项优势就是构建在它之上的一切。企业已开发出数百万行代码和内部工作流程，形成了强大的锁定效应，使得切换至替代方案成本高昂且过程繁琐。
Business Insider 此前报道称，Amazon 的内部文件将 CUDA列为其 Trainium 和 Inferentia AI 芯片获得采用的主要障碍。
高通旗下 AI 软件创业公司 Modular的联合创始人兼 CEO Chris Lattner表示，CUDA 的历史既是优势，也是约束。CUDA 最初是在 AI 热潮到来很久之前为游戏而打造，因此承载着多层遗留技术，“就像试图把 Microsoft Windows 塞进一部手机”。
Modular 联合创始人兼 CEO Chris Lattner。Modular
另一些人表示，AI 从训练转向推理——即模型回答请求并得出结论——带来了另一项威胁。
韩国 AI 芯片创业公司 Rebellions 首席商务官 Marshall Choy 表示，随着这一演变，企业较少关注通过最强大的芯片实现极致性能，而更关注如何以盈利方式运行 AI。
这可能不仅会带来对专用硬件的更大需求，也会推动可在不同芯片之间运行的软件。如果企业能够在不重写软件的情况下切换芯片，就能削弱 CUDA 最重要的锁定效应之一。
“这正是 Nvidia 的 CUDA 护城河被打破的地方，因为在推理侧，CUDA 不再是决定因素，”Choy 表示。“这是一场开源竞争。”
Nvidia 表示，随着 AI 模型被投入实际工作，其紧密集成的软硬件产品变得更有价值。
“随着 AI 转向推理和 Agent 型工作负载，对深度全栈优化的需求只会持续增长，”Patel 表示。
InvestorPlace 首席技术分析师 Luke Lango 表示，市场向专用芯片和软件的转移，正使华尔街愈发质疑 Nvidia 的 CUDA 优势。他称，Nvidia 股价在过去一年停滞不前，反映出市场对这些问题的部分担忧。
Nvidia 的护城河并未消失——而是在转移
并非所有人都认为编程 Agent 正在侵蚀 CUDA 的优势。一些人认为，它们最终可能反而会强化这一优势。
AI 软件创业公司 INT21 创始人 Bing Xu 表示，虽然 Agent 让生成软件变得更容易，但 AI 生成的代码仍须经过验证和优化。他认为，CUDA 拥有最深厚的验证工具生态及其他功能，可帮助编程 Agent 更高效地工作。
他表示，随着 Agent 日益普及，这一生态系统将成为 CUDA 的下一道护城河。
“Agent 能在短时间内生成大量代码，但验证是最大的瓶颈，”Xu 表示。他此前创办的 AI 芯片软件公司 HippoML 已被 Nvidia 收购；他于 4 月离开这家芯片制造商，创办 INT21。
INT21 创始人兼 CEO Bing Xu。INT21
Lattner 表示，尽管编程 Agent 的确让构建芯片软件变得更容易，但这种提升是渐进式的。
“这种炒作并非完全没有依据，但被严重夸大了，”他说，并补充称，相比为生产环境优化软件等更复杂的任务，编写代码只是构建软件的一小部分。生产环境优化至关重要，因为能够最大化芯片性能的软件可以降低大规模运行 AI 的成本。
Lattner 表示，芯片软件也是一个相对小众的领域，通常由顶尖工程师开发。因此，相较于 AI 已在海量公开代码上接受训练的应用开发领域，编程 Agent 可供学习的样本要少得多。
Xu 表示，尽管 AI 可能帮助竞争对手追赶，Nvidia 同样受益于这些技术变革。编程 Agent 是否会削弱 CUDA，取决于竞争对手追赶 Nvidia 的速度，是否快于 Nvidia 开拓新优势的速度。
Xu 表示，这家全球领先的芯片制造商“并没有沉睡，也没有停滞不前”。
English / 原文
Nvidia's biggest competitive advantage is no longer as untouchable as it once seemed.
For two decades, Nvidia's crown jewel wasn't just chips; it was the software that turned them into the building blocks of AI, known as CUDA.
Short for Compute Unified Device Architecture, CUDA is the brainchild of longtime Nvidia executive Ian Buck, who heads high-performance computing. It took years to build, with ready-made code for common AI tasks, tools to find bugs, and software that lets thousands of chips work together to train models.
Now, some believe AI could eventually automate one of the industry's hardest jobs: building the software that powers AI itself.
The industry is at an important threshold, says Jeremy Nixon, a former Google Brain researcher and the founder of AI software startup Infinity. He told Business Insider his startup used AI coding agents to recreate CUDA-like software for the chip startup D-Matrix in 10 hours — evidence, he said, that one of Nvidia's biggest moats is being crossed.
Infinity founder and CEO Jeremy Nixon. Infinity.inc
The pressure doesn't only come from startups. Cloud giants like Google, Amazon, and Microsoft have spent years building software around their own AI chips, while OpenAI and Anthropic have recently demonstrated AI models capable of generating system software.
DeepSeek founder Liang Wenfeng recently said that coding agents, along with his startup's own programming language TileLang, have made AI software substantially easier to build.
Coding agents aren't just helping challengers.
Nvidia said developers increasingly use CUDA's code libraries to build AI applications, and that it also "uses AI coding agents to develop CUDA faster and validate at greater scale," said Ankit Patel, Nvidia's vice president of developer ecosystem.
Inference could change the CUDA equation
If CUDA's first advantage was software, the second is everything built on top of it. Millions of lines of code and internal workflows have been developed by companies, creating a powerful lock-in effect that makes switching to alternatives costly and cumbersome.
Internal documents at Amazon identified CUDA as a major roadblock to adoption of its Trainium and Inferentia AI chips, Business Insider previously reported.
CUDA's age is both an advantage and a constraint, said Chris Lattner, cofounder and CEO of Qualcomm-owned AI software startup Modular. Originally built for gaming long before the AI boom, CUDA carries layers of legacy technology, "like Microsoft Windows trying to fit onto a phone," he said.
Modular cofounder and CEO Chris Lattner. Modular
Others say AI's shift from training toward inference — where models answer requests and draw conclusions — creates another threat.
With this evolution, companies care less about maximizing performance with the most powerful chips and more about running AI profitably, said Marshall Choy, chief business officer of Korean AI chip startup Rebellions.
This could result in greater demand not only for specialized hardware but for software that can run across different chips. If companies can switch between chips without rewriting software, that reduces one of CUDA's biggest lock-ins.
"That's where the CUDA moat from Nvidia gets broken because CUDA is no longer a factor in the inference side," Choy said. "It's an open source play."
Nvidia said that its tightly integrated hardware and software offerings have become more valuable as AI models get put to work.
"As AI shifts toward inference and agentic workloads, the need for deep, full-stack optimization only grows," Patel said.
A shift toward specialized chips and software has Wall Street increasingly questioning Nvidia's CUDA advantage, said Luke Lango, chief technology analyst at InvestorPlace. He said Nvidia's stagnant stock price over the past year reflects some of these concerns.
Nvidia's moat isn't disappearing — it's shifting
Not everyone agrees that coding agents are eroding CUDA's edge. Some believe they could ultimately strengthen it instead.
Though agents make it easier to generate software, AI-generated code still has to be verified and optimized, said Bing Xu, founder of AI software startup INT21. He believes CUDA has the deepest ecosystem of verification tools and other features that help coding agents work more efficiently.
As agents become more common, he said, that ecosystem will become CUDA's next moat.
"Agents can generate a lot of code in a short time, but verification is the biggest bottleneck," said Xu, whose last AI chip software startup, HippoML, was acquired by Nvidia. He left the chipmaker in April to build INT21.
INT21 founder and CEO Bing Xu. INT21
While coding agents do make it easier to build chip software, the improvement is incremental, Lattner said.
"The hype is not complete nonsense, but it is very overblown," he said, adding that writing code is only a small part of building software compared to more complex tasks like optimizing it for production — a critical task because software that maximizes a chip's performance reduces the cost of running AI at scale.
Chip software is also something of a niche field, often worked on by elite engineers, Lattner said, giving coding agents far fewer examples to learn from than, for instance, app development, where AI has been trained on vast amounts of public code.
And while AI may help rivals catch up, Nvidia benefits from the same technological shifts, Xu said. Whether coding agents weaken CUDA depends on whether competitors catch Nvidia faster than it can gain new ground.
The world's dominant chipmaker is "not sleeping or keeping still," Xu said.