在 2026 年 2 月 4 日（GMT+8，12:30）Parmy Olson 的评论中，Nvidia 执行长 Jensen Huang 被指称向业界人士表示，先前对 OpenAI 所「宣布」的 1,000 亿美元投资其实不具约束力，并私下批评其商业纪律不足。原本与基础设施扩建绑定的承诺，现在更像是在 OpenAI 目前为潜在 IPO 进行募资时，以「数百亿美元等级」的较小下注呈现；同时，OpenAI 另被称正在与 Nvidia、Microsoft、Amazon 讨论募集约 1,000 亿美元资本，且与先前拟议的基础设施交易分开。Huang 对外否认不满，并表示仍会投入「大量资金」，但其对风险的对冲被视为值得注意的讯号。
文章将 OpenAI 的执行困境与数字化风险绑在一起：Sam Altman 在 2023 年末的戏剧性被解职后复职，随后一连串复杂交易使公司背负约 1.4 兆美元的运算承诺，约为 OpenAI 预估 2025 年营收的 100 倍。产品推进被描述为仓促，例子包括 GPT Store 与自订 GPT 的开发者市集尝试，因缺乏清晰策略而未能成功。这些迹象被用来支持「外部模型能力很强、内部执行与商业化纪律不足」的判断，并解释为何即使领导层否认问题，外部资金与合作方仍倾向收缩与分散押注。
Microsoft 被描绘为同样的「执行」失速者：其股价自 ChatGPT 上线以来几近翻倍，早期对 OpenAI 的 130 亿美元投资换得 27% 持股，估值约 1,350 亿美元，超过原始投入的 10 倍；2025 年 10 月的重组协议还让 Microsoft 取得至 2032 年对 OpenAI 智财与模型的排他性存取。即便如此，Microsoft Copilot 仍被批评混乱、受限、难用；对照 Anthropic 用自家 AI 写程式工具在 10 天内做出 Claude Cowork，并可在授权下操作个人电脑、整理档案、从文件生成 PowerPoint 与 Excel、回复 LinkedIn 讯息，差距更醒目。文中引述 Sycomore Sustainable Tech 的 David Rainville 指出，若未在未来 6 个月内推出同级产品将引发高层人事震荡；同时也提到 OpenAI 对 Microsoft 抗拒分享完整技术细节、以及 Microsoft 于 2025 年 8 月将 Copilot 接入 GPT-5 后「明显变好」但仍需具备电脑任务执行能力，以跟上 Anthropic 与 OpenClaw。结论是：在模型日益商品化的趋势下，胜负转向产品化与落地执行；Jensen Huang 的迟疑被定位为对 Microsoft 的警告——必须把特权存取转化为更好的产品，否则将被甩开。
In Parmy Olson’s Feb. 4, 2026 (GMT+8, 12:30) commentary, Nvidia CEO Jensen Huang is reported to have told industry associates that his previously “announced” $100 billion investment in OpenAI was nonbinding, while privately criticizing a lack of business discipline. What was originally tied to an infrastructure build-out now appears more likely to materialize as a smaller, tens-of-billions-scale bet within OpenAI’s current fundraising ahead of a potential IPO; separately, OpenAI is said to be discussing raising roughly $100 billion with Nvidia, Microsoft, and Amazon, apart from the proposed infrastructure deal. Huang publicly denies being unhappy and says Nvidia will still invest “a great deal of money,” but the hedging itself is treated as a notable signal.
The piece links OpenAI’s execution concerns to quantified risk: after Sam Altman’s late-2023 firing and return, a stream of complex deals allegedly left the company on the hook for about $1.4 trillion in computing commitments, roughly 100× OpenAI’s projected 2025 revenue. Product rollouts are characterized as frantic, with the developer-marketplace push around the GPT Store and custom GPTs cited as fizzling for lack of a clear strategy. These details are used to support a thesis of strong model output paired with weak operational and commercial discipline, explaining why even as leadership downplays issues, funders and partners may narrow and diversify exposure.
Microsoft is portrayed as facing a similar execution gap: its stock has nearly doubled since ChatGPT’s launch, and its early $13 billion OpenAI bet now equates to a 27% stake valued around $135 billion, more than 10× the original investment; an October 2025 restructuring also gives Microsoft exclusive access to OpenAI’s IP and models through 2032. Yet Microsoft Copilot is criticized as confusing, constrained, and hard to use, made starker by Anthropic building Claude Cowork in 10 days with its own AI coding tool and, with permission, enabling it to operate a personal computer, organize files, generate PowerPoint decks and Excel spreadsheets from documents, and reply to LinkedIn messages. The article quotes Sycomore Sustainable Tech’s David Rainville warning that if Microsoft does not ship an equivalent within the next six months, leadership changes will follow; it also cites OpenAI’s resistance to sharing full technical details with Microsoft and notes Copilot became “drastically better” after being plugged into GPT-5 in August 2025 but still needs computer-task agency to keep up with Anthropic and OpenClaw. The conclusion is that as models commoditize, competition shifts to product execution; Huang’s hesitation is framed as a warning that Microsoft must convert privileged access into better products or risk falling behind.