Why Top Founders Are Racing Into AI Infrastructure

以下是将内容翻译为中文： 新成立的“机器时代基金”是为了应对被誉为我们这个时代最重要的技术革命而设立的，其重要性可能超越互联网、微处理器，甚至车轮。这个人工智能新时代需要一套全新的基础设施，它远远超出了传统的服务器、存储和网络，甚至“深入到矿山，特别是铜矿”。 基金合伙人（Ben、Martin和Raghu）阐述的核心问题是前所未有的资源限制，而不仅仅是工程挑战。对人工智能的需求被描述为“无限”，正迅速超越供应链各个环节的供应能力。超大规模公司正将资本支出提高到创纪录水平（预计明年将达1万亿美元），这表明了巨大的真实需求，而非仅仅是炒作周期。GPU价格 historically 趋于下降，但现在却急剧上涨，关键组件已被预订到2028年，内存供应商表示，仅当前需求就需要三年才能满足。 这种情况与互联网繁荣时期投机性的“暗光纤”建设形成鲜明对比；今天的人工智能基础设施，每一部分都是预售的。行业面临电力、散热、内存和GPU的关键短缺，芯片周期（3-4年）和数据中心建设（4-5年，包括电力来源）的漫长交付周期更是雪上加霜。根本问题在于现有的硬件架构和系统从未为人工智能工作负载设计，导致了物理限制和效率低下。 基金的成立反映了创业格局的重大转变，顶尖创始人解决复杂硬件问题的比例大幅增加（从5%上升到20-30%的交易）。这表明整个行业普遍认识到“瓶颈在于我所说的模型以下的一切”，即物理基础设施而非人工智能模型本身。人工智能倾向于通过“更多人工智能”（例如通过推理、链式思维、长期运行的代理）来解决问题，这不断增加token消耗，从而产生持续升级的需求。 基础设施需求令人震惊： * **电力：** 机架功率从5-10千瓦增至100-150千瓦，需要从交流电转向更危险的直流电，认证电工严重短缺。数据中心需要环保、回馈社区并提供自己的电力解决方案，以应对到2028年所需的额外44吉瓦电力（而电网预计仅增加25吉瓦）。 * **散热：** 从风冷到液冷的转变已在进行中，并推动环保解决方案。 * **物理厂房：** 设备重量和密度的增加要求更坚固的地板、更厚的墙壁（由于噪音），以及可能用于施工和维护的机器人。 * **供应链：** 变压器和涡轮机短缺，加上监管障碍和许可证审批流程，阻碍了快速扩张。 机器时代基金将专门关注“计算机科学基础设施”——芯片、网络、互连、存储和基础电力解决方案。尽管英伟达等现有巨头占据主导地位，但他们对“金砖”（指其核心、高利润产品）的关注，为新公司在边缘领域创新并实现效率和能力上所需的十倍提升留下了巨大机会。这些挑战的复杂性意味着成功的创始人往往是经验更丰富的“系统创始人”，他们能够整合设计、制造和整个生态系统。 “机器时代”这个名称强调的是“机器智能”，而非“人工智能”，承认了物理机器在这个新时代的关键作用。长期愿景（5-10年）是让美国在这场基础设施竞赛中获胜，培育大量环保、高效的数据中心、芯片和电力，从而保持国家的科技领先地位及其独特的深刻创新能力。

The newly launched Machine Age Fund is introduced as a response to what is being hailed as the most significant technological revolution of our time, potentially surpassing the internet, the microprocessor, or even the wheel. This new era of AI necessitates an entirely new infrastructure, one that extends far beyond traditional servers, storage, and networks, reaching "all the way down to the mines, copper mines." The core problem articulated by the fund's partners (Ben, Martin, and Raghu) is an unprecedented resource limitation, not merely an engineering challenge. Demand for AI, described as "infinite," is rapidly outpacing supply across every part of the supply chain. Hyperscale companies are increasing their capital expenditure to record levels (projected $1 trillion next year), indicating massive real demand, not just a hype cycle. GPU prices, which historically decline, have instead risen sharply, and essential components are booked out to 2028, with memory suppliers stating current demand alone would take three years to fulfill. This situation is a stark contrast to the speculative "dark fiber" build-out during the internet boom; every piece of AI infrastructure today is pre-sold. The industry faces critical shortages in power, cooling, memory, and GPUs, compounded by long lead times for chip cycles (3-4 years) and data center construction (4-5 years), including securing power sources. The fundamental issue is that existing hardware architectures and systems were never designed for AI workloads, leading to physical limits and inefficiencies. The fund's inception reflects a significant shift in the entrepreneurial landscape, with a dramatic increase in top founders tackling complex hardware problems (from 5% to 20-30% of deals). This indicates a community-wide recognition that the "bottleneck is all what I call south of the model," meaning the physical infrastructure rather than the AI models themselves. AI's tendency to solve problems by using "more AI" (e.g., through inference, chained thought, long-running agents) continuously multiplies token consumption, creating an ever-escalating demand. The infrastructure requirements are staggering: * **Power:** Rack power moving from 5-10 kilowatts to 100-150 kilowatts, requiring a shift from AC to more dangerous DC power, with a severe shortage of certified electricians. Data centers will need to be eco-friendly, contribute back to the community, and provide their own power solutions to address the 44 gigawatts of additional power needed by 2028 (against an expected 25 gigawatt grid addition). * **Cooling:** Transition from air to liquid cooling is already underway, with a push for eco-friendly solutions. * **Physical Plants:** The increased weight and density of equipment demand stronger floors, thicker walls (due to noise), and potentially robots for construction and maintenance. * **Supply Chain:** Shortages of transformers and turbines, coupled with regulatory hurdles and permit processes, hinder rapid expansion. The Machine Age Fund will specifically target "computer science infrastructure" – chips, networks, interconnects, storage, and foundational power solutions. While incumbents like NVIDIA are dominant, their focus on "gold bricks" leaves significant opportunity for new companies to innovate at the margins and drive the necessary 10x improvements in efficiency and capability. The complexity of these challenges means successful founders are often more experienced "systems founders" who can integrate design, manufacturing, and the entire ecosystem. The name "Machine Age" emphasizes "machine intelligence" over "artificial intelligence," acknowledging the critical role of physical machines in this new era. The long-term vision (5-10 years) is for America to win in this infrastructure game, fostering an abundance of eco-friendly, efficient data centers, chips, and power, thereby maintaining the country's technological leadership and its unique capacity for profound innovation.