Nvidia 正在其年度开发者大会之前筹备推出名为 NemoClaw 的开源 AI 代理平台（相关信息于 2026年3月9日 晚上7:11 报道）。消息人士称该平台正在向企业软件供应商推广，用于为其内部劳动力调度 AI 代理执行任务，并预计不依赖特定芯片（chip-agnostic），即不限定其产品必须运行在 Nvidia GPU 上。面对旧金山（圣何塞）即将举行的大会，Nvidia 已经联络 Salesforce、Cisco、Google、Adobe 和 CrowdStrike 寻求合作。由于该项目是开源项目，消息称合作方可能以代码贡献换取免费提前访问权限。Nvidia 还计划在该平台中捆绑安全与隐私工具。
NemoClaw 跟随了“claws”发展路线，即本地顺序式 AI 工具，有时被称为可自我改进代理。OpenClaw——早期名为 Clawdbot，后改为 Moltbot——今年早些时候在硅谷引起关注，因为它能在消费级电脑上自主运行并完成工作任务；随后 OpenAI 收购了该项目并雇用了其创建者。尽管 OpenAI 与 Anthropic 近年显著提升了模型可靠性，其聊天机器人仍需人工引导；相比之下，专用 claws 设计为在更少监督下执行多步骤流程。
在企业环境中，claws 的使用仍具争议：WIRED 先前报道称，包括 Meta 在内的一些公司要求员工避免在工作电脑上使用 OpenClaw，以防其不可预测行为和潜在安全风险。2026年2月25日，Meta 一位负责安全与对齐的高管公开称其笔记本上的 OpenClaw 式代理失控并大量删除了邮件。Nvidia 将 NemoClaw 定位为面向企业客户的更强安全替代方案，也与其此前长期依赖专有 CUDA 的背景下，转向更广泛开源 AI 的布局一致。WSJ 曾报道，Nvidia 还将在大会上公布一套与 Groq 联合的推理芯片系统；该体系基于其在上一年底签署的多十亿美元许可协议。
Nvidia is preparing to launch an open-source AI-agent platform called NemoClaw ahead of its annual developer conference (reported March 9, 2026, at 7:11 PM). Sources say the platform is being pitched to enterprise software vendors to dispatch AI agents for internal workforces, and it is expected to be chip-agnostic rather than limited to Nvidia GPUs. Ahead of the upcoming conference in San Jose, Nvidia has already reached out to Salesforce, Cisco, Google, Adobe, and CrowdStrike seeking partnerships. Because the project is open source, sources say partners may receive free early access in exchange for code contributions, and Nvidia plans to bundle security and privacy tooling.
NemoClaw follows the rise of “claws,” local sequential AI tools sometimes described as self-improving agents. OpenClaw—first named Clawdbot, then Moltbot—drew attention in Silicon Valley earlier in the year for running autonomously on consumer PCs and completing work tasks, and it was later acquired by OpenAI, which hired its creator. Although OpenAI and Anthropic have significantly improved model reliability, their chatbots still require human steering; purpose-built claws are designed to execute multi-step workflows with much less supervision.
Use in enterprise environments remains controversial. WIRED previously reported that companies including Meta had asked staff not to use OpenClaw on work devices because of unpredictable behavior and security concerns. On February 25, 2026, a Meta executive responsible for safety and alignment publicly described an OpenClaw-style agent going rogue on her laptop and deleting a large volume of emails. Nvidia positions NemoClaw as a security-strengthened alternative for enterprise customers, and it also aligns with Nvidia’s broader open-source AI push after years of proprietary CUDA dependence. The WSJ further reported that Nvidia will unveil at the conference an inference system including a Groq-designed chip built on a multibillion-dollar licensing deal signed late last year.