在9月的一次内部员工会议上，亚马逊云科技（AWS）首席财务官约翰·费尔顿对普拉萨德·卡利亚纳拉曼开了一个玩笑，后者是负责亚马逊大部分人工智能基础设施扩张的高管。“你让我们公司变成了自由现金流为负，所以恭喜你，”根据《商业内幕》网站获得的一份会议录音，费尔顿说道。这个评论本意轻松。但它捕捉到了一种惊人的逆转。
At an internal staff meeting in September, AWS CFO John Felton had a joke for Prasad Kalyanaraman, the executive overseeing much of Amazon's AI infrastructure expansion. "You've taken us free cash flow negative as a company, so congratulations," Felton said, according to a recording of the meeting obtained by Business Insider. The remark was lighthearted. But it captured a stunning reversal.
就在两年前，华尔街还在讨论亚马逊将如何处理其不断增加的现金储备。如今，随着亚马逊开启其有史以来最大规模的押注——仅今年一年就花费2200亿美元来建设人工智能的物理骨干网，现金流已经蒸发。在会议上，卡利亚纳拉曼没有反驳费尔顿的前提。亚马逊的机器学习和人工智能业务“自由现金流严重为负”，他说道，并补充预计公司的自由现金流在“未来两年左右”将保持负值。
Just two years ago, Wall Street was debating what Amazon would do with its growing cash hoard. Now, cash flow has evaporated as Amazon embarks on its biggest bet ever, spending $220 billion this year alone to build the physical backbone for AI. At the meeting, Kalyanaraman did not dispute Felton's premise. Amazon's machine-learning and AI business is "hugely free cash flow negative," he said, adding that he expects the company's free cash flow to remain negative for "the next couple of years or so."
亚马逊以前也做过重大押注。它曾在云计算、物流中心和配送网络上投入巨资，远在这些业务成为公司最有价值的资产之前。当时投资者对这些投资提出质疑，但亚马逊最终利用这些业务建立了庞大且盈利的庞然大物。人工智能正在以大得多的规模检验这一游戏规则。在投资者了解这些投资将以多快速度产生收入之前，亚马逊已向数据中心、服务器和电力投入了数千亿美元。
Amazon has made big bets before. It spent heavily on cloud computing, fulfillment centers, and its delivery network long before they became some of the company's most valuable assets. Investors questioned those investments at the time, but Amazon ultimately built enormous, profitable businesses from them. AI is testing this playbook at a far larger scale. Amazon is committing hundreds of billions of dollars to data centers, servers, and power before investors know how quickly those investments will generate revenue.
凭借几十年的经验，该公司是或一直是建设数据中心最优秀的公司之一。但这些项目的规模和复杂性，以及阻碍进展的瓶颈，即使对这样一位备受尊崇的云计算先驱来说也极具挑战性。《商业内幕》审阅的内部文件和知情人士透露，该公司正试图协调一个庞大的工业系统，而该系统的各个组成部分以不同的时间表运转。
The company is the best, or among the best, at building data centers, with decades of experience. But the size and complexity of these projects, and the bottlenecks holding progress back, are challenging — even for such a respected cloud pioneer. Internal documents reviewed by Business Insider and people familiar with the matter show the company trying to coordinate a sprawling industrial system whose parts move on different schedules.
电力、设备或建设方面的延误可能会使价值数十亿美元的硬件闲置——并导致亚马逊的资金被占用——同时预测和成本也在不断变化。亚马逊表示，有强有力的证据表明市场将需要这些产能。亚马逊网络服务（AWS）第二季度营收增长37%，创下18个季度以来的最快增速，其订单积压量则达到了4960亿美元。首席执行官安迪·贾西（Andy Jassy）认为，公司对从这一轮创纪录的投资周期中获得丰厚财务回报有着“清晰的前景”。
A delay in power, equipment, or construction can leave billions of dollars in hardware waiting — and Amazon's cash tied up — while forecasts and costs keep shifting. Amazon says it has strong evidence that the capacity will be needed. AWS revenue grew 37% in the second quarter, its fastest pace in 18 quarters, while its backlog reached $496 billion. CEO Andy Jassy thinks the company has a "clear line of sight" to strong financial returns from this record investment cycle.
他在最新的年度致股东信中将人工智能称为“一生难得的机遇”，并表示亚马逊在投资方面不会保守。贾西写道：“我们经历过AWS第一波大增长周期，并且很满意当时的结果。我们预计对这下一波浪潮也会有同样的感受。”投资者对行业扩张的规模感到担忧，尤其是随着人工智能数据中心的建设成本不断增加、完工时间不断延长。
In his latest annual shareholder letter, he called AI a "once in a lifetime opportunity" and said Amazon would not be conservative in its investments. "We've been through this cycle with the first big AWS growth wave, and liked the results," Jassy wrote. "We expect to feel similarly about this next wave." Investors are concerned by the scale of the industry's expansion, especially as AI data centers are costing more to build and taking longer to complete.
Yorkville Ives分析师丹·艾夫斯（Dan Ives）在最近的一份报告中写道，市场正在“重新对任何针对随后到来的收入进行多年期建设融资的行为进行定价”。他补充说，围绕数据中心审批日益增加的不确定性以及关于电价的政治争端，给投资者提供了又一个质疑这些投资回报速度的理由。行业面临的一个重大挑战是确保获得充足的电力，以为数据中心内部的设备供电。
The market is "repricing anything that funds multi-year builds against revenue that arrives later," Dan Ives, an analyst at Yorkville Ives, wrote in a recent report. Growing uncertainty around data center permitting and the political fight over electricity prices have given investors another reason to question how quickly these investments will pay off, he added. One major industry challenge is securing enough power to run the equipment that goes inside data centers.
这包括装有GPU和其他芯片的昂贵服务器机架，这些芯片用于训练和运行ChatGPT等人工智能模型及服务。8月份的内部邮件显示，亚马逊正在为美国设施中预期到来的一波搭载英伟达GB300芯片的服务器机架筹备一项新的“早期机架收货”流程，以在最终部署前做好准备。
This includes expensive racks of servers that contain GPUs and other chips that train and run AI models and services such as ChatGPT. Internal emails from August show Amazon preparing a new "Early Rack Receipt" process for an expected wave of server racks carrying Nvidia's GB300 chips at US facilities before final deployment.
一封亚马逊的电子邮件预计，该公司将在8月份接收价值约50亿美元的英伟达GB300机架，随后在“可预见的未来”每月再接收价值约30亿美元的机架。
One Amazon email projected the company would take delivery of about $5 billion of Nvidia GB300 racks in August, followed by another roughly $3 billion each month "for the foreseeable future."
据知情人士透露，一些亚马逊网络服务（AWS）数据中心站点有空间容纳更多机架，但没有足够的可用电力来启动它们，导致这些设备只能暂存于中转仓库，直到设施准备就绪。由于讨论的是敏感事项，这些知情人士要求匿名。
Some AWS data center sites had room for more racks but not enough available power to turn them on, people familiar with the effort said, leaving the equipment to be staged at intermediary warehouses until the facilities were ready. These people asked not to be identified discussing sensitive matters.
根据《商业内幕》获得的内部文件，通常情况下，亚马逊在设备被放入数据中心之前不会为机架付款，从而让供应商承担运输中设备的成本。
Usually, Amazon doesn't pay for racks until the gear is put into a data center, leaving suppliers to carry the cost of the equipment in transit, according to the internal documents obtained by Business Insider.
这有助于该公司保存现金。然而，人工智能数据中心建设热潮的强度，使得这种协调变得更加困难。
This helps the company conserve cash. However, the intensity of the AI data center building boom has made this type of coordination harder.
随着亚马逊与竞争对手争夺英伟达图形处理器（GPU）和其他人工智能组件的获取权限，它必须为稀缺资源展开更激烈的竞争。
Amazon has to compete more for scarce resources as it jockeys with rivals for access to Nvidia GPUs and other AI components.
根据新的“提前接收机架”流程，亚马逊现在在供应链的更早期阶段就接管了这些机架，而不是等到设备运达数据中心。
Under the new "Early Rack Receipt" process, Amazon now takes ownership of these racks earlier in the supply chain, rather than waiting until the equipment reaches a data center.
这意味着更快地支出现金。任何导致机架安装并接入可靠电源的延迟，都可能会推迟这些资产开始产生收益的时间。
The means spending cash sooner. And any delays getting the racks installed and plugged into a reliable power source potentially pushes back the moment when revenue starts flowing from these assets.
根据《商业内幕》审阅的一封电子邮件，至少有一个亚马逊网络服务供应链团队没有关于将机架送入数据中心的“清晰预测路径”。
At least one AWS supply chain team did not have "a clear forecasted path" for getting racks into data centers, according to one of the emails reviewed by Business Insider.
亚马逊的一名发言人表示，该公司例行在部署站点附近暂存成品的硬件，以便在电力容量可用时将其安装，并称这种方法是“最佳商业做法”。
An Amazon spokesperson said the company routinely stages finished hardware near deployment sites so it can be installed once power capacity is available, calling that approach "best business practice."
该发言人补充说，亚马逊网络服务拥有“长期确立的系统”来追踪基础设施资产，并对更广泛的公司缺乏将机架送入数据中心的预测路径这一说法提出异议。这个问题并非亚马逊所独有。
The spokesperson added that AWS has "long-established systems" to track infrastructure assets and disputed that the broader company lacks a forecasted path for getting racks into data centers. This problem is not unique to Amazon.
伯恩斯坦分析师在最近的一份报告中写道，电力许可、冷却系统、电气设备、建筑和硬件的进展时间表各不相同。“时间安排并不完美，”分析师们说。
Bernstein analysts wrote in a recent note that power permits, cooling systems, electrical equipment, construction, and hardware all move on separate schedules. "Timing is imperfect," the analysts said.
在9月份的内部员工会议上，AWS首席财务官费尔顿对这一更广泛的问题作出了生动的描述：“当未来本身充满不确定性时，你该如何在未来工作中？”人工智能基础设施负责人卡利亚拉曼表示，亚马逊对其人工智能投资回报的内部预期有所改善。但这些回报背后的假设不断在变。“预测每天都在变，”卡利亚拉曼说，“我们生活的世界极其动态。”
During the internal staff meeting in September, Felton, the AWS CFO, described the broader problem eloquently: "How do you work in the future when the future is inherently uncertain?" Kalyanaraman, the AI infrastructure leader, said Amazon's internal outlook for the return on its AI investments had improved. But the assumptions behind those returns keep shifting. "The forecast is changing on a daily basis," Kalyanaraman said. "The world that we live in is extremely dynamic."
他解释说，这些不断变化的假设包括数据中心何时准备就绪，以及AWS需要备有多少设备。八周的供应缓冲对于一家小公司来说可能是例行公事。但卡利亚拉曼表示，在亚马逊的规模下，这可能意味着持有价值“数十亿美元”的额外设备库存——从而占用宝贵的现金。
Those shifting assumptions include when data centers will be ready and how much equipment AWS needs to keep on hand, he explained. An eight-week supply buffer might be routine for a smaller company. But at Amazon's scale, Kalyanaraman said, this could mean holding "billions and billions of dollars" worth of additional equipment inventory — tying up precious cash.
这种紧张关系对卡利亚拉曼来说尤为引人注目。这位在亚马逊工作了21年的老员工在今年早些时候晋升为直接向贾西汇报的核心管理层S团队。亚马逊的创业故事以节俭著称。20世纪90年代中期创始人杰夫·贝索斯在西雅图创业时，曾用廉价的门板当办公桌，这广为人知。卡利亚拉曼已经将这种传说融会贯通。2023年底，在AWS面临增长放缓和裁员之际，他敦促员工以更少的资源做更多的事。“节俭孕育创新，”他当时说。
Such tensions are especially striking for Kalyanaraman, a 21-year Amazon veteran who earlier this year was promoted to the elite S-team that reports directly to Jassy. Amazon's founding story is defined by frugality. It famously used cheap doors as desks when founder Jeff Bezos started in Seattle in the mid-90s. Kalyanaraman has absorbed this lore into bones. In late 2023, as AWS faced slower growth and layoffs, he urged employees to do more with less. "Frugality breeds innovation," he said back then.
这种本能依然塑造着他的方法。在今年9月的会议上，卡利亚拉曼告诉费尔顿，AWS试图在“最后负责任的时刻”进行重大投资，也就是在对需求有最清晰展望的时候。“如果我们能尽可能晚地把资本花出去，对我们来说就更好，”他说。但人工智能热潮的狂热步伐使得这一原则很难遵循。
That instinct still shapes his approach. In the September meeting this year, Kalyanaraman told Felton that AWS tries to make major investments at the "last responsible moment," when it has the clearest possible view of demand. "If we can wait to spend the capital as late as possible, it's better for us," he said. But the frenetic pace of the AI boom makes this principle difficult to follow.
等到需求完全显现时，AWS所需的服务器、电力和建设能力可能已经无法获得。这给该公司带来了压力，迫使其更快地投入更多资金，以确保供应链安全并实现其数据中心容量目标。卡利亚拉曼回忆说，他在2023年曾告诉杰西，AWS到2027年“肯定”能摆脱产能危机。然而这位人工智能基础设施负责人在今年9月的会议上表示，亚马逊目前预计，到2027年需求仍将超过其可用产能。
By the time demand is fully visible, the servers, power, and construction capabilities that AWS needs may no longer be available. That puts pressure on the company to commit more money sooner to secure supply and meet its data center capacity goals. Kalyanaraman recalled telling Jassy in 2023 that AWS would "surely" be out of its capacity crunch by 2027. Instead, Amazon still expects demand to exceed its available capacity in 2027, the AI infrastructure chief said during the meeting this September.
对亚马逊这样的行业领头羊来说，获得充足的电力或许是最大的挑战。根据Cleanview Research的估计，该公司在16个州开发的数据中心容量达到44吉瓦，是谷歌、Meta和微软总和的两倍多。一个名为“受保障电力风险”的AWS内部仪表盘，用于追踪影响AWS为其数据中心筹措的电力容量的公用事业和监管限制。这些风险包括输电瓶颈、在电网负荷过重时减少用电的要求，以及如果数据中心项目错过使用或建设里程碑便可能危及容量的“不用即作废”条款。
Getting enough electricity may be the biggest challenge for an industry leader like Amazon. The company has 44 gigawatts of data center capacity under development across 16 states, more than double Google, Meta, and Microsoft, according to Cleanview Research estimates. An internal AWS dashboard called "Secured Power at Risk" tracks utility and regulatory constraints affecting electricity capacity AWS has lined up for its data centers. The risks include transmission bottlenecks, requirements to reduce use during grid stress, and "use it or lose it" provisions that can jeopardize capacity if a data center project misses usage or construction milestones.
换句话说，AWS并非只是坐等电力到位。它正在管理一组承诺，如果数据中心项目延期，这些承诺的成本可能会更高。通过公共记录申请获得的一份去年向马里兰州公用事业监管机构提交的文件，说明了大规模数据中心建设可能伴随的风险。亚马逊向马里兰州公共服务委员会表示，大型数据中心客户可以帮助支付电网升级费用，或将新的电力容量投放市场，而不会给其他电价纳税人带来负担。该公司还提议预留抵押品，以帮助公用事业公司放心地进行为数据中心产生更多电力所需的大额投资。
In other words, AWS is not simply waiting for electricity to arrive. It is managing a portfolio of commitments that can become more costly if a data center project is delayed. A submission to Maryland's utility regulator last year, obtained through a public records request, illustrates the kind of exposure that can accompany massive data center buildouts. Amazon told the Maryland Public Service Commission that large data center customers could help pay for grid upgrades or bring new power capacity to market without burdening other electricity ratepayers. The company also proposed setting aside collateral to help utilities get more comfortable with the large investments needed to generate more electricity for data centers.
亚马逊提议提供三年的抵押担保，以便在大型电力客户放弃未使用的容量时，公用事业公司有时间转售电力。该提议并不证明AWS已经进行了此类付款或提供了抵押担保。不过，这确实表明，如果项目延迟或取消，大型数据中心除了自身的电费之外，还可能面临额外的财务义务。亚马逊发言人表示，此类公用事业规则并非AWS所独有。
Amazon proposed three years of collateral so that if a large electricity customer abandoned unused capacity, the utility would have time to re-sell the power. The proposal is not evidence that AWS has made such a payment or posted collateral. It does, however, show that large data centers can face extra financial obligations, beyond their own electricity bills, if projects are delayed or cancelled. The Amazon spokesperson said such utility rules are not unique to AWS.
该公司尚未披露为确保电力供应已预先支付了多少现金。人工智能的建设还带来了比传统资本支出计划更难建模的费用。卡利亚拉曼在今年9月的一次会议上表示，在审视西得克萨斯州吉瓦级容量的规划时，AWS遇到了劳动力问题：附近没有足够的管工、电工和其他熟练工人来建设这些设施。他说，公司的应对措施是出资在数据中心附近建造住宅综合体，以便工人们可以在那里居住两年或三年，直到项目完成。
The company has not disclosed how much cash it has paid upfront to secure power. The AI buildout also incurs expenses that are harder to model than a conventional capex plan. In the September meeting this year, Kalyanaraman said that while reviewing plans for gigawatts of capacity in West Texas, AWS confronted a labor problem: There weren't enough nearby pipefitters, electricians, and other skilled workers to build the facilities. The company's response, he said, was to pay to construct housing complexes near its data centers so workers could live there for the two or three years needed to complete the projects.
甚至用于运输服务器的板条箱也成为了一个规划问题。卡利亚拉曼说，当服务器价值数十万美元时，担心集装箱的价格似乎很荒谬。但在AWS的规模下，由于公司需要如此多的板条箱，即使是微不足道的物品也会成为一项实质性的开支。“我们不断遇到新的限制，”卡利亚拉曼说。
Even the crates used to ship servers have become a planning issue. It may seem absurd to worry about the price of a container when servers cost hundreds of thousands of dollars, Kalyanaraman said. But at AWS's scale, where the company needs so many crates, even trivial items can become a material expense. "We have new constraints that keep popping up," Kalyanaraman said.