据报道，亚马逊云计算部门AWS正在以前所未有的速度采购英伟达最新AI服务器系统。根据内部规划，AWS一度预计接收价值约50亿美元的英伟达GB300服务器机架，未来还计划持续以每月约30亿美元的规模采购相关设备。然而，部分设备在交付后并未立即投入使用，而是被暂时存放在仓库或中转设施中。
造成这一现象的原因并非需求不足，而是部分数据中心虽然已经具备安装空间，却缺乏足够的供电能力来驱动这些高功耗AI系统。业内将这类已经采购但暂时无法运行的设备称为“暗GPU（Dark GPU）”。这一概念既可以指因电力或基础设施限制而闲置的硬件，也可用于描述已部署但未被充分利用的AI计算资源。
过去几年里，人工智能产业面临的主要挑战是先进芯片供应紧张。各大科技企业争相争夺英伟达GPU配额，芯片往往供不应求。但随着企业提前锁定供应链资源，新的限制因素开始转向电网和能源基础设施。对于大型数据中心运营商而言，获得稳定且充足的电力供应正变得和获取芯片同样重要。
为了确保不会在未来芯片竞争中落于下风，亚马逊调整了采购模式。按照以往流程，公司通常会在服务器机架送达可投入运营的数据中心后才正式接收并付款。而在新的“提前接收机架”机制下，亚马逊会更早完成接收和付款流程，即使设备尚未被安装到最终运行地点。
这一做法有助于提前锁定稀缺硬件资源，但也意味着企业需要提前投入大量资金，而这些设备在正式上线并向客户提供云计算服务之前无法产生收入。因此，巨额资本支出与延迟回报之间的矛盾正在加剧。
亚马逊今年计划投入约2200亿美元资本支出用于AI和云计算基础设施建设，规模超过许多科技同行。与此同时，公司的自由现金流已经承受明显压力，并出现近年来首次转负的情况。对于AWS而言，提前占有关键硬件资源被视为一项战略选择，即便部分设备暂时无法运行。
分析人士指出，当前AI产业的发展重心正在发生变化。两年前行业关注的是谁能够获得更多英伟达GPU，如今越来越多企业发现，真正稀缺的资源或许已不是芯片本身，而是支撑这些芯片运行所需的能源和电力网络。
随着AI模型规模不断扩大、算力需求持续增长，电力供应、输电能力以及数据中心建设周期正逐渐成为影响产业发展的关键因素。大量暂时处于闲置状态的GPU也从侧面反映出一个现实：人工智能竞赛已不仅是芯片之争，更是能源与基础设施之争。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
It is reported that AWS, Amazon Cloud Calculator, is purchasing the latest AI server system in Weida at an unprecedented rate. Based on internal planning, AWS had at one point expected to receive British Wida GB300 server racks valued at approximately $5 billion, and it was planned to purchase related equipment on a continuing basis at a scale of approximately $3 billion per month. However, some of the equipment was not put into service immediately after delivery but was temporarily stored in a warehouse or transit facility.
This is not due to inadequate demand, but rather to the lack of sufficient power to drive these high-capacity AI systems in some data centres, which already have installed space. This type of equipment, which has been procured but is temporarily not operational, is referred to by industry as “Dark GPU”. This concept can be used to describe both unoccupied hardware due to power or infrastructure constraints and AI computing resources that have been deployed but not fully utilized.
The main challenge facing the AI industry over the past few years has been the supply of advanced chips. Major technology firms compete for a GPU quota, and chips are often scarce. However, as enterprises pre-empt supply chain resources, new constraints begin to shift to grid and energy infrastructure. For large data centre operators, access to a stable and adequate power supply is becoming as important as access to chips.
In order to ensure that future chip competition would not fall short, the Amazon adjusted its procurement model. In accordance with past processes, companies usually formally receive and pay for server racks after they reach the operational data centre. Under the new “early-receiving rack” mechanism, the Amazon would have completed the receipt and payment process earlier, even if the equipment had not yet been installed at the final operational location.
This approach helps to target scarce hardware resources in advance, but it also means that enterprises need to invest large amounts of money in advance, and that equipment cannot generate income until it is formally online and provides cloud computing services to clients. As a result, the tension between large capital expenditures and delayed returns is increasing.
This year, the Amazon plans to invest some $22 billion in capital spending on AI and cloud computing infrastructure, which is larger than many of its science and technology counterparts. At the same time, the company ' s free cash flow has come under clear pressure and has been reversed for the first time in recent years. For AWS, pre-positioning of critical hardware resources is considered a strategic option, even if some of the equipment is temporarily not operational.
According to analysts, the current focus of development of the AI industry is changing. Two years ago, the industry focused on who could get more GPUs, and more and more companies now find that the real scarce resources may no longer be the chips themselves, but rather the energy and power networks that support their operation.
As the size of the AI model expands and computing demand continues to grow, power supply, transmission capacity and the data centre construction cycle are becoming key factors influencing industrial development. The large number of temporarily idle GPUs also reflects the reality that artificially intelligent competitions are not just a chip contest, but energy versus infrastructure.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.