FT:美国10万亿美元AI基础设施热潮因项目延误、公众反弹和投资者信心不稳显现裂痕
美国10万亿美元AI基础设施热潮因项目延误、公众反弹和投资者信心不稳显现裂痕。相关建设支撑着美国经济,当前面临重新评估。
X(推特)上的财经、科技与宏观账号
美国10万亿美元AI基础设施热潮因项目延误、公众反弹和投资者信心不稳显现裂痕。相关建设支撑着美国经济,当前面临重新评估。
This is important to crystallize. The underlying demand trend is incredibly strong and required. Whether that means shaky business models, bad deals, expensive financing, or scarcity pricing pencils is another debate altogether. Someone will be there to pick up the pieces. But 引用 @BenBajarin:All
去看一下这周的ALL-IN,前面20分钟,JASON聊到老马发推“YIKES”,加入CHAMATH关于COREWEAVE的评论时,他发出的奸笑,我认为他肯定做空了。 然后视频开始后他团队做的DECK聊加息对于COREWEAVE成本的影响,“因为一开始的UNSECURED
市场追逐算力短缺,推动资本承诺加速,但竞争和商品化压缩产品利润率,利率与地缘政治冲击也可能增加融资难度。投入依据是当下的稀缺经济学,回报却取决于多年后的现金流;算力和基础设施可能结构性供给不足,建设融资的资本复杂性也可能过度扩张。
社媒真的太多关于新云融资的FUD了。 DDTL4.0,COREWEAVE最新的融资,成本SOFR+2.25%,按现在的联邦利率就是6.13% META 2032年到期的债务YIELD是5.82%
remind me when all the tech ppl were like datacenters are going to go up everywhere and everyone will figure it out no problem b/c grooooooowth 引用 @saranormous:Anyone can build a datacenter, they said https://t.co/8iUs4ZXS2p
I've been pretty clear about my skepticism around land, power and shells, though the rest of the world waking up in a big way could change the picture. The question is whether it happens, and on what timeline. “TSMC is very conservative and doesn't freelance. They're sending 引用 @DratchCap:Zoom out
另外不得不吐槽一下,老马说28年中国会产很多很多RAM,我就问你,老马的时间点什么时候对过? 另外SPACEX就是一个巨大的新云,27年10GW,算过他要融资多少钱吗?
Given the CoreWeave drama. A good time to remember that a datacenter is a bank. 引用 @TheStalwart:A DATACENTER IS A BANK In today’s newsletter, I wrote some things I’ve been thinking about since our chat with @iaindunning, and how maybe one day we’ll have datacenter supervisors like we have bank sup
Sferical One 这台由 NVIDIA 提供算力支持的共享 AI 超级计算机举行启动庆祝活动。黄仁勋与 Wallenberg Investments 的 Marcus Wallenberg、Spotify 的 @eldsjal 及 Sferical AI 的 Jenny Nordlöw 共同出席。
马斯克称,SpaceX与英伟达共同设计的Star Mind计算机性能远好于多数人的预期,并计划明年进入量产。马斯克还表示,与英伟达工程团队合作令人愉快。
Shanu Mathew称,燃料电池(BE)的标示成本约为5700美元/千瓦,较每千瓦2000至3000美元的燃气基准高90%至185%。他表示,35%的税收抵免及额外优惠可降低成本,但基础成本仍高;引用内容称,Google与AEP签订20年电力供应协议,涉及AEP一座耗资40亿美元、装机702MW的燃料电池电厂。
Nvidia副总裁Rod Evans在Vertiv的EMEA活动上表示,AI数据中心消耗的电力中,只有60%–70%用于有效AI工作。TechRadar称,他还拆解了一座用电负荷为1GW的数据中心的用电分配,但正文未提供后续明细。
AirTrunk据报道将约15亿美元的新加坡REIT IPO推迟至2027年,几天前Firmus撤回了上市。Firmus寻求的估值接近其8月估值的3倍;AirTrunk推迟也反映出利率较高。两起事件表明,投资者希望承担风险时获得更有吸引力的定价。
英伟达为 Coreweave 的融资提供 GPU 剩余价值担保,以帮助其获得较便宜的资金;资金由基建巨头提供。Coreweave 2032 年到期的高级无担保债券收益率为 15%,价格为 79;发文者称,资金成本上升、信用风险上升。
Let's not panic folks. Datacenter neo clouds are highly leveraged business which leverage cuts both ways but credit isn't in crisis. That said at current cost of leverage data center inputs demand should slow a tad https://t.co/9mlO9G6eh4
Cushman & Wakefield 的《2026指南》新增了关于建设成本、劳动力、设备交付周期、通胀和选址的外部参考。现代绿地设施平均建设成本为每兆瓦 1760 万美元,不含芯片和 GPU;每兆瓦建设成本自 Q4 2024 以来上涨 21%。
Bloomberg puts a theoretical 1GW AI data center in Abilene at ~$30bn. Its ten cost categories actually sum $17.1bn for chips and servers, and ~$13bn for everything else. My read is that the hardware assumption is likely closer to Hopper-era economics vs. the latest generation. https://t.co/KjTa26XLj
图表分析不同数据中心容量假设对实验室收入和支出的含义,依据包括公司披露、公开报道和示例情景。文中引用的估算显示,2026年美国数据中心增长规模约为10—21 GW(按IT口径),不同估值跨度超过两倍。
US has ~2x China's data center capacity, and 3x+ Europe's. Let's not lose that competitive advantage. https://t.co/jGw5qEHFYH
Something that blows my mind is the variance of the estimates we're still seeing for data center capacity. We’re already in mid-October and the selected estimates for 2026 US data-center growth still span roughly 10–21 GW on an IT basis. A range more than 2x for the year we’re https://t.co/VL1dm0sQw
Love people opening up the convo to debate/learn in public and Matt is great at it. Sharing for broader awareness. I'll leave the chip-based GW forecasts to others, but on power, I've presented several frameworks for thinking about US IT capacity and the very real limits that https://t.co/fTp3ZzdW5
作者称,DTE与Google的交易符合大型负荷客户承诺期限延长、最低付款要求提高的更广泛变化。文中列举AEP Ohio的安排为爬坡后8年、最长4年爬坡期,且该规模客户需求底线为85%;Dominion的安排为14年,输配电需求底线为85%,发电需求底线为60%。
DTE与Google在Van Buren Township签订1 GW、期限为20年的电力合同。作者称,密歇根州10月1日的相关命令披露了超大规模云服务商的最低电力承诺,并提到为公用事业提供下行保护的基础设施投资安排。
Hut 8建设数据中心并与英伟达签订15年租约,英伟达作为租户以其信用支持建设融资,随后将空间交由Lambda安装GPU并出租。引用的SemiAnalysis称,该安排涉及约115K台Rubin GPU销售、Lambda收入约10%的分成,以及一个704 MW供电园区第7至第15年的空间;互联排队时间为4至7年。
美国贸易逆差中的“AI”部分正在增长,主要体现为自墨西哥进口增加;台湾及其他代工厂也已大规模转向墨西哥进行最终组装。Foxconn位于墨西哥哈利斯科州埃尔萨尔托、组装英伟达 NVL72 服务器机架的工厂,目前供应的商品接近美国商品进口总额的 1%。
the public backlash against data centers in the U.S. is spreading to Europe and Asia! https://t.co/ABxV3CLpQG
There isn’t yet a single fungible unit of compute. An “H100 GPU-hour” tells you the hardware and duration, but cluster size, networking, location and contract terms can materially change what that capacity is worth. @ash_locked does an excellent job explaining how this carries https://t.co/bdAoUCSl
英伟达 Vera Rubin 服务器启动生产,Blackwell GB300 和 B300 服务器量产,推动 AI 服务器厂商 Wistron 营收创纪录新高。AMD CEO Lisa Su 近期访问 Wistron 位于 Zhubei 的新厂 A1,显示 Helios 服务器业务延续强势。
媒体报道称,Delta 2026 年资本开支目前预计为 NT$70 billion,高于今年早些时候的 NT$46.6B。原文称,Delta 正在扩充 AI 产能;其他 Lite-On 竞争对手也因 AI 数据中心和 AI 服务器厂商的需求而增加资本开支。
Lite-On Technology签下金额为3.5亿美元(NT$11.06B)的合同,用于建设其位于德州McKinney的工厂。该工厂计划生产高压直流电源机架及AI基础设施相关产品;媒体报道称,这份合同属于一项总额9.19亿美元的计划。
Per Grok, “Trucked compressed natural gas is reported to cost roughly four times pipeline deliveries once labor, equipment, and truck fuel are included, and scaling it even to 100 MW of the 2.45 GW Project Jupiter site would require a large trailer roughly every 40 minutes.”
Oracle 正用卡车运输天然气,为 AI 数据中心供电,以绕过可能威胁其扩张的管道延误。该做法已在 Utah 和 Texas 使用,可能扩展至其 2.45-GW 项目;评论指出,卡车运输在能源效率、成本和排放方面都不理想。
New: the IBM® Spectrum Symphony Provider Plugin for Azure Compute Fleet. Symphony orchestrates the workload, Compute Fleet supplies elastic capacity across Spot and pay-as-you-go VMs. Partners, here is your burst-to-Azure motion: https://t.co/OqDtxJaB2i https://t.co/cqPoKTKxxU
Shanu Mathew称,@interconnectfyi 的并网队列报告显示,发电类型趋稳之际,各市场出现一些项目理性化迹象,更多投机性项目正在撤回。2026年5月2日至10月2日,净流入约189 GW,其中PJM约139 GW;原文链接:https://x.com/ShanuMathew93/status/2108132620458483999
Shout out my guy @rmcentush on the great piece around why it's not as simple as ordering a bunch of equipment and 'plugging it in' for data centers - grid or BTM. This is helpful for laymen and experts, both! Visuals are great. Lucky to see early version of the piece and Ryan https://t.co/jB8lBonRc
One thing that absolutely is in the textbooks -- a big investment boom both pulls in imports and changes the savings v investment balance. The current data center/ AI investment boom is in fact unusually import intensive 2/3 https://t.co/lILe9wv7wx
final assembly of all the data center computers has shifted to Mexico and Taiwan as much as it has shifted to Vietnam ... https://t.co/mly2GqkidL
怀俄明州 Cheyenne 数据中心更新供电安排,Black Hills 将为 Google 数据中心供电,资源组合净规模为 2.7GW(含所需备用裕度)。其中 590MW 来自电网,另有 2.1GW 来自第三方签约资源。
英伟达支持的 Project Southgate 去年 10 月公布计划,拟到 2028 年在澳大利亚部署最高 1.6GW AI 算力;一年后,已部署规模约为 42MW,CDC 开发合作关系结束。原文称,该项目规划利用可再生能源提供主权 AI 算力,并由 CDC 提供数据。