以下是内容的中文翻译： 英伟达（NVIDIA）首席执行官黄仁勋坚称，认为人工智能将使软件商品化的预期是“幼稚的”，他反而预测工具用户和使用实例将呈指数级增长，从而使软件公司受益。他将英伟达的根本作用描述为将“电子转化为代币”（tokens），这是一个远离商品化的复杂“制造型科学”过程。他指出，英伟达的实力在于其庞大的生态系统，涵盖从台积电（TSMC）和美光（Micron）等上游供应链合作伙伴，到下游应用开发者和AI模型制造商，所有这些都得益于其可编程的CUDA架构。公司的理念是“只做必要之事，不做不必要之事”，专注于核心创新，同时培育广泛的合作伙伴网络。 据报道，英伟达高达数千亿美元的巨额采购承诺，彰显了其前瞻性的供应链管理策略。黄仁勋认为，制造业中的暂时瓶颈（如CoWoS封装）可在2-3年内通过集中投入和架构创新得到解决，例如从Hopper到Blackwell实现了50倍的效率飞跃。他认为能源政策是一个更具挑战性的长期制约因素。 谈到竞争，黄仁勋将英伟达的“加速计算”与谷歌TPU等专用ASIC（专用集成电路）区分开来。他强调了英伟达更广泛的应用多样性、庞大的CUDA生态系统、巨大的用户基础以及在所有主流云服务商中的无处不在的部署。他认为，这种灵活性对于算法创新至关重要，因为算法创新推动AI进步的速度远超单纯依靠摩尔定律的改进。他强调英伟达卓越的总拥有成本（TCO）和每瓦特处理代币的效率，并向竞争对手发出挑战，要求他们在MLPerf等公开基准测试中展示出更优的性能。 黄仁勋承认，尽管超大规模云服务商和大型AI实验室可能会开发定制化内核，但英伟达深厚的专业知识仍能提供显著的性能提升（通常是2-3倍），这直接转化为他们的营收。他承认过去犯了一个“错误”，即未能对Anthropic等基础性AI实验室进行早期的数十亿美元投资，认识到传统风险投资未能满足他们独特的资金需求。英伟达现在正积极投资这类实验室，目的并非为了“挑选赢家”，而是为了培育整个AI生态系统。他坚定地认为，英伟达的核心使命是构建计算平台，而非成为云服务商，同时积极支持“新云”（NeoClouds）及其他合作伙伴。 讨论中有很大一部分内容涉及中国和美国的出口管制。黄仁勋极力争辩称，限制对华芯片销售是一项“政策错误”，将损害美国的长期利益。他认为，中国拥有充足的计算资源（包括能源、现有芯片以及大量擅长算法创新的优秀AI研究人员），无论如何都能够自行发展出先进的人工智能。他警告说，拒绝中国接触“美国技术栈”（即英伟达的CUDA生态系统）将加速中国独立技术栈的发展，最终侵蚀美国的科技领导地位及其在全球推广其标准的能力。尽管承认AI模型可能被用于攻击性目的（如网络攻击），黄仁勋仍倡导国际对话，并认为美国应通过驱动创新和在全球竞争中保持领先，而非采取“失败主义心态”拱手让出重要市场。他强调美国应在AI技术栈的所有五个层面保持领先，并且需要采取审慎细致的方法，避免“幼稚的”极端做法。 最后，黄仁勋指出，即使没有深度学习革命，英伟达仍然会是一家实力雄厚的公司，专注于在涵盖科学、工程和图形等多个领域的应用中进行加速计算。通用计算的局限性确保了对专业化加速的持续需求，这是英伟达几十年来一直追求的使命，使其成为超越人工智能的众多领域的基础性推动者。

Jensen Huang, CEO of NVIDIA, asserts that the expectation of AI commoditizing software is "naive," instead predicting an exponential growth in tool users and instances, benefiting software companies. He describes NVIDIA's fundamental role as transforming "electrons to tokens," a complex process of "manufacturing science" far from commoditization. NVIDIA's strength, he argues, lies in its vast ecosystem, stretching from upstream supply chain partners like TSMC and Micron to downstream application developers and AI model makers, all enabled by its programmable CUDA architecture. The company's philosophy is to do "as much as necessary, as little as possible," focusing on core innovations while fostering a broad partner network. NVIDIA's massive purchase commitments, reportedly reaching hundreds of billions, are a testament to its proactive supply chain management. Huang believes temporary bottlenecks in manufacturing (like CoWoS packaging) are resolvable within 2-3 years through focused investment and architectural innovation, such as the 50x efficiency leap from Hopper to Blackwell. He views energy policy as a more challenging, long-term constraint. Regarding competition, Huang differentiates NVIDIA's "accelerated computing" from specialized ASICs like Google's TPUs. He highlights NVIDIA's greater diversity in applications, its extensive CUDA ecosystem, large install base, and ubiquitous presence across all major cloud providers. This flexibility, he argues, is crucial for algorithmic innovation, which drives AI advancements faster than mere Moore's Law improvements. He stresses NVIDIA's superior Total Cost of Ownership (TCO) and tokens-per-watt efficiency, challenging competitors to demonstrate better performance in public benchmarks like MLPerf. Huang acknowledges that while hyperscalers and large AI labs may develop custom kernels, NVIDIA's deep expertise still yields significant performance gains (often 2-3x), which translates directly to their revenue. He admits a past "mistake" in not making early multi-billion dollar investments in foundational AI labs like Anthropic, recognizing that their unique funding needs weren't met by traditional VCs. NVIDIA is now actively investing in such labs, not to pick winners, but to foster the overall AI ecosystem. He firmly maintains that NVIDIA's core mission is building the computing platform, not becoming a cloud provider, while actively supporting "NeoClouds" and other partners. A significant portion of the discussion addresses China and US export controls. Huang vehemently argues that restricting chip sales to China is a "policy mistake" detrimental to US long-term interests. He contends that China has abundant computing resources (energy, existing chips, and a massive pool of talented AI researchers who excel at algorithmic innovation) that would allow them to develop advanced AI regardless. He warns that denying access to the "American tech stack" (NVIDIA's CUDA ecosystem) would accelerate the development of an independent Chinese tech stack, ultimately eroding US technological leadership and its ability to diffuse its standards globally. While acknowledging the potential for AI models to be used for offensive purposes (like cyberattacks), Huang advocates for international dialogue and for the US to maintain its lead by driving innovation and competing globally, rather than adopting a "losing mindset" that cedes a significant market. He stresses that the US should lead in all five layers of the AI stack and that a nuanced approach is required, avoiding "childish" extremes. Finally, Huang notes that even without the deep learning revolution, NVIDIA would still be a formidable company, dedicated to accelerated computing across diverse scientific, engineering, and graphics applications. The limits of general-purpose computing ensure a continued need for specialized acceleration, a mission NVIDIA has pursued for decades, making it a foundational enabler across numerous fields beyond just AI.