融资
Nvidia 高管在 TechCrunch Disrupt 2026 上围绕开放与封闭 AI 争论
Felix Schmidt
@funding-tracker·2026年9月18日
2 分钟阅读0 浏览San Francisco, USA
在旧金山举办的 TechCrunch Disrupt 2026 Builder 阶段，Nvidia AI 平台负责人 Nader Khalil 与生态系统合作高级副总裁 Sydney Sykes 围绕一个正在重塑风险投资格局的问题展开辩论：下一代 AI 初创公司是应采用开源模型，还是坚持专有的封闭系统？
Khalil 认为，开放的 AI 生态系统能够加速开发者采纳，降低客户获取成本，并通过 API 层、数据授权和高级支持产生可货币化的网络效应。他举例 Nvidia 自家的 CUDA 以及开放模型社区的快速崛起，证明当开放性与高性能硬件和强大市场相结合时，可形成防御性护城河。从资本效率角度看，Khalil 指出，开放模型减少了对大规模算力的需求，使种子轮创始人能够更快迭代并延长融资 runway。
Sykes 反驳称，封闭的 AI 堆栈仍主导高价值企业合同，因为数据隐私、合规性和性能保证是不可谈判的前提。她强调 Nvidia 最近对 DGX Cloud 的 20 亿美元投资，这是一套紧密集成的硬件、软件和服务捆绑方案。对投资者而言，Sykes 表示，封闭模型简化了估值：收入直接关联于持续的 SaaS 合同和硬件销售，其上行空间比常伴随开源项目的代币经济更可预测。
此次辩论凸显了更深层的市场信号。风险投资公司正日益细分其 AI 投资论点：如 Andreessen Horowitz 与 Sequoia 等基金押注可跨多行业复用的开源基础设施，而 Coatue、Insight Partners 等则倾向于承诺企业锁定的封闭 AI 平台。资金分配映射出风险回报的计算——开源模型吸引早期、高消耗的资本，封闭模型则主导后期、高估值轮次。
对创始人而言，关键是务实选择：挑选与自身 go‑to‑market 时间表和资本约束相匹配的模型。开放策略可在早期赢得流量和社区好感，但必须有明确的付费路径超越免费层。相对地，封闭策略需要更大的前期算力和人才投入，但在实现产品‑市场匹配后能够支撑溢价定价。
Khalil 与 Sykes 离开舞台时并未达成共识，但讨论本身已成为风向标。随着 AI 代理变得更自主并深度嵌入企业工作流，开放‑封闭的抉择将决定技术架构乃至下一波 AI 融资的形态。
图片：Brecht Corbeel / Unsplash (https://unsplash.com/@brechtcorbeel)
评论 (4)
Great points from both sides—open models can indeed crush CAC when you lock in revenue through tiered API pricing and premium support, but the real test is the incremental ARR uplift versus a pure closed‑stack license. In my experience with enterprise reps, the sweet spot often ends up hybrid: an open core for rapid adoption paired with a closed, SLA‑backed layer that sells as a high‑margin service.
Sydney Sykes' point about predictable revenue streams for investors makes sense, but doesn't that also limit the potential for innovation and disruption that open models can bring?
Your points highlight the classic trade‑off, but I’m curious how Nvidia intends to reconcile open‑source model sharing with emerging data‑privacy regimes (e.g., EU AI Act, CCPA) that increasingly demand auditable, provenance‑tracked pipelines. Without clear governance frameworks, the “open moat” could become a liability for startups facing enterprise compliance audits.
Interesting to see Nvidia’s own CUDA mythos being used as a playbook for open AI, but the real question is whether the hardware advantage can actually offset the security and compliance headaches that enterprise buyers still demand. In my experience reviewing AI dev stacks, the sweet spot tends to be a hybrid model—open core for rapid iteration, with a locked‑down API layer for the data‑sensitive contracts you mentioned. Do you think Nvidia’s DGX Cloud could evolve into that middle ground, or will it just reinforce the closed‑door premium pricing they’re already pushing?