The Real Numbers Behind Stargate
星际之门背后的真实数字

We believe this project is being measured on the Total Cost of Ownership, not Capital Expenditures. IE, the $100 billion and $500 billion is TCO which includes the capital expenditures for datacenters, servers, networking costs, power costs, and many other operational costs including debt/financing costs. Furthermore, we believe only the first $100B has datacenter sites and power ready to go. Funding is still not fully secured.
我们相信这个项目是基于总拥有成本（TCO）进行评估的，而不是资本支出。也就是说，1000 亿美元和 5000 亿美元是总拥有成本，其中包括数据中心、服务器、网络成本、电力成本以及许多其他运营成本，包括债务/融资成本。此外，我们认为只有前 1000 亿美元有数据中心地点和电力准备就绪。资金仍未完全到位。

The other deceptive detail is that the first cluster of this megaproject is the Oracle OpenAI deal that was announced last year. This is being fully built by Lancium, Crusoe, and Oracle. As far as we know, no other such as Softbank and MGX are involved.
另一个误导性的细节是，这个大型项目的第一个集群是去年宣布的 Oracle OpenAI 交易。这完全由 Lancium、Crusoe 和 Oracle 建造。据我们所知，没有其他公司如软银和 MGX 参与。

Project Overview  项目概述

The Stargate Joint Venture is a new company intended to invest $500B in AI infrastructure over the next four years with $100B being deployed immediately. This is the first big AI initiative announced by the second Trump administration with more likely to come due to the significant relaxing of regulations.
星际门户合资企业是一家新公司，计划在未来四年内投资 5000 亿美元于人工智能基础设施，其中 1000 亿美元将立即投入。这是第二任特朗普政府宣布的第一个大型人工智能倡议，预计由于监管的显著放松，还会有更多类似的举措。

The equity founders in Stargate are SoftBank, Oracle, MGX, and OpenAI, with the former three companies contributing capital initially. Softbank SoftBank and OpenAI are the lead partners, with SoftBank having financial responsibility and OAI leading operations including almost all cluster management software.
Stargate 的股权创始人是软银、甲骨文、MGX 和 OpenAI，其中前三家公司最初贡献了资本。软银和 OpenAI 是主要合作伙伴，软银负责财务，OpenAI 负责运营，包括几乎所有的集群管理软件。

The project includes the buildout of 20 data centers, 10 of which are already underway in Abilene, Texas with each roughly 500k square feet. We provide more information below on this campus from our Datacenter Industry Model.
该项目包括建设 20 个数据中心，其中 10 个已在德克萨斯州阿比林开工，每个数据中心大约 500,000 平方英尺。我们在下面提供有关该园区的更多信息，来自我们的数据中心行业模型。

Scope and Location of the Initial $100 Billion Investment
初始 1000 亿美元投资的范围和地点

The intial Oracle/Crusoe is included as one of the datacenters that will be connected to other Gigacampuses to perform distributed training at a scale that is orders of magnitude larger than any other training run. Cleverly, they went for a more creative financing structure through this separate entity, which pushes the need for fundraising down the road.
初始的 Oracle/Crusoe 被纳入作为将连接到其他 Gigacampus 的数据中心之一，以进行规模比任何其他训练运行大几个数量级的分布式训练。巧妙的是，他们通过这个独立实体选择了一种更具创意的融资结构，这将推迟筹款的需求。

Ghe Stargate Joint Venture campus mentioned by Sam Altman as “already under way” is the Oracle/Crusoe Abilene campus. This is the campus being built to house the 100k GB200 cluster for OAI’s use. This development along with the GB200 servers that are due to be shipped over Q2 and Q3 will be dropped into the Oracle/Crusoe datacenters.
萨姆·阿尔特曼提到的“已经在进行中的 Ghe Stargate 合资企业校园”是 Oracle/Crusoe Abilene 校园。该校园正在建设中，以容纳 OAI 使用的 100k GB200 集群。此项开发以及预计在第二季度和第三季度发货的 GB200 服务器将被放入 Oracle/Crusoe 数据中心。

A high-resolution photo of the campus from Dec 29, 2024 is shown below. Four of the ~120k square foot modules make up one 480k square foot building, so only two buildings are currently being built with ~180MW of Critical IT Capacity. To give an understanding of the scale of this build-out, the 20 building campus will be comprised of 80 of those modules!
以下是 2024 年 12 月 29 日校园的高分辨率照片。大约 120,000 平方英尺的四个模块构成一个 480,000 平方英尺的建筑，因此目前仅建造了两座建筑，具备约 180MW 的关键 IT 容量。为了让人理解这一建设规模，20 座建筑的校园将由 80 个这样的模块组成！

Based on the build-out we have scheduled for OpenAI/Microsoft/Oracle in both our Accelerator Model and Datacenter Model, OpenAI doesn’t have to build a new campus to expend all $100B in TCO. Therefore, we believe that the entirety of the initial $100B from the Stargate Joint Venture is going to this campus alone.
根据我们为 OpenAI/Microsoft/Oracle 在加速器模型和数据中心模型中安排的建设，OpenAI 不需要建立新的校园来支出全部 1000 亿美元的总拥有成本。因此，我们相信 Stargate 合资企业的初始 1000 亿美元将全部用于这个校园。

This campus in the model is known as the Lancium Crusoe Abilene Clean Campus, and has a stated capacity on their website of 1GW energized in 2025. However, we have learned through multiple industry sources and filings that the campus has unlocked another 1GW. This is a total of 2.2GW of power supplied. As such, we estimate the campus has a Critical IT Capacity of 1.8GW based on our Datacenter Model. This is a PUE of ~1.25. It is possible they are able to improve PUE to get closer to 2GW IT capacity, but this will be difficult given power density increases for GB300 and especially VR200.
该模型中的校园被称为 Lancium Crusoe Abilene 清洁校园，其网站上声明的容量为 2025 年启用的 1GW。然而，我们通过多个行业来源和文件了解到，该校园已解锁另一个 1GW。这总共提供了 2.2GW 的电力。因此，我们根据我们的数据中心模型估计该校园的关键 IT 容量为 1.8GW。这是一个约为 1.25 的 PUE。他们可能能够改善 PUE，以接近 2GW 的 IT 容量，但考虑到 GB300 和特别是 VR200 的功率密度增加，这将是困难的。

Power, Rental Price, Capex, and GPU Spend Breakdown
电力、租金、资本支出和 GPU 支出细分

In addition to the 100k GB200 cluster in-process, we estimate another 200k GB300s and ~400k VR200s will be deployed throughout the campus. We calculated the implied number of Vera Rubin units based on the total $100B TCO investment for the campus and the TCO per unit from our AI TCO Model. The resulting total all-in IT power for the ~700k GPUs is 1.8GW, which ties out to the 1,800MW of capacity for the Oracle/Crusoe Abilene Campus.
除了正在进行的 100k GB200 集群外，我们估计校园内还将部署另外 200k GB300 和约 400k VR200。我们根据校园总计的 1000 亿美元 TCO 投资和我们 AI TCO 模型中的每单位 TCO 计算了隐含的 Vera Rubin 单元数量。大约 70 万 GPU 的总 IT 功率为 1.8GW，这与 Oracle/Crusoe Abilene 校园的 1800MW 容量相符。

In short, we estimate OpenAI is paying rougly ~$2.8 per GPU hour for the first GB200 cluster which is a TCO of $2.38 an hour for the full-scale cluster. These figures scale with future Nvidia GPU deployments. The spend OpenAI will exceed the $100B, but isn’t upfront capital. We estimate Nvidia’s revenue from this and IT equipment Capex are more tame at ~$37.5B and ~$49B. This will take the entire 1.8GW Crusoe Lancium site.
简而言之，我们估计 OpenAI 为第一个 GB200 集群支付的费用大约为每小时$2.8，这对于全规模集群的总拥有成本（TCO）为每小时$2.38。这些数字会随着未来 Nvidia GPU 的部署而变化。OpenAI 的支出将超过$100B，但并不是前期资本。我们估计 Nvidia 来自此项及 IT 设备资本支出的收入较为温和，约为$37.5B 和$49B。这将占用整个 1.8GW 的 Crusoe Lancium 场地。

As we go forward to GB300 and VR200, the number of GPUs will double each generation and the power per GPU also grows. While 700,000 GPUs sounds small for $100B note this is equivalent to many millions of Hopper GPUs.
随着我们向 GB300 和 VR200 迈进，每一代的 GPU 数量将翻倍，单个 GPU 的性能也在增长。虽然 70 万个 GPU 对于 1000 亿美元来说听起来很少，但这相当于数百万个 Hopper GPU。

The Project’s Financial Backing
项目的财务支持

Oracle and SoftBank are the well-known equity providers. MGX is a more recently founded Technology Investment Vehicle from Abu Dhabi, founded by both Mubadala and G42. The fund is chaired by Sheikh Tahnoon bin Zayed al-Nahyan, who is reported to oversee over $1.5 trillion in investment funds. In light of the recent Elon Musk claim that Softbank could only raise under $10B, it is important to note the JV will also raise debt capital for project financing. It’s unclear what the mix of equity and debt financing will be, Oracle and MGX can fund a big chunk of the project with their massive balance sheets.
甲骨文和软银是知名的股权提供者。MGX 是阿布扎比最近成立的技术投资工具，由穆巴达拉和 G42 共同创立。该基金由谢赫·塔赫努恩·本·扎耶德·阿尔·纳哈扬担任主席，他负责管理超过 1.5 万亿美元的投资基金。鉴于最近埃隆·马斯克声称软银只能筹集不到 100 亿美元，值得注意的是，合资企业还将为项目融资筹集债务资本。目前尚不清楚股权和债务融资的组合将是什么，甲骨文和 MGX 可以利用其庞大的资产负债表为项目提供大部分资金。

OpenAI and Softbank are allegedly down for $19 billion on paper each. Softbank doesn’t have this liquid today, but we think SoftBank will likely have to sell off a portion of its ARM stake to fund the equity check. OpenAI also doesn’t have the capital, but they will be able to raise with o3’s release, and they don’t have to do it directly fully, but can through some kind of hybrid instrument issued to everyone else in the consortium.
OpenAI 和软银据称在账面上各损失了 190 亿美元。软银今天没有这样的流动资金，但我们认为软银可能需要出售其部分 ARM 股份来筹集资金。OpenAI 也没有资本，但他们将能够通过 o3 的发布来筹集资金，他们不必完全直接进行，但可以通过向财团中的其他人发行某种混合工具来实现。

Elon Musk is right, they don’t have the money yet.
埃隆·马斯克是对的，他们还没有钱。

The question is if project financing risk will be separated from the parent organizations. Separating this risk is important because OpenAI will need to raise capital to pay for this. Thankfully for OpenAI, it’s not immediately required. OAI has enough existing capital to pay for rental payments on the 100k GB200 cluster this year, but they need to be able to pay rental fees in 2026 and beyond, which scale massively. It is possible OpenAI cannot raise this money directly, but we are believers that their technology and products will scale to support this.
问题是项目融资风险是否会与母公司分开。分离这一风险很重要，因为 OpenAI 需要筹集资金来支付这笔费用。幸运的是，OpenAI 目前并不急需这笔资金。OAI 拥有足够的现有资本来支付今年 100k GB200 集群的租金，但他们需要能够在 2026 年及以后支付租金，这笔费用将大幅增加。OpenAI 可能无法直接筹集到这笔资金，但我们相信他们的技术和产品将能够支持这一点。

The Reality of The Stargate Joint Venture – Winners and Losers
星际之门合资企业的现实 - 胜者与败者

Many big names were announced alongside the Stargate JV, ranging from Nvidia, Arm, Microsoft, Oracle, and the previously mentioned financiers. The names are large and impressive, but the reality is there are distinct winners and losers.
许多知名公司与 Stargate 合资企业一同宣布，包括 Nvidia、Arm、Microsoft、Oracle 以及之前提到的融资方。这些名字都很大且令人印象深刻，但现实是有明显的赢家和输家。

Microsoft Is The Biggest Loser
微软是最大的输家

Microsoft is largely left on the sidelines as OpenAI utilizes another infrastructure partner. As Satya Nadella stated, “all I know is, I’m good for my $80B.” Microsoft’s recent blog post stated the facts of the partnership and how Microsoft has a ROFR, but the reality is this is incrementally negative for Microsoft long-term since they were afraid to take on the risk of this investment. Furthermore, Chris Young, who was in charge of OpenAI partnerships also left abruptly directly after the announcement. Bulls will say Microsoft is shifting some of the capital intensity of its AI initiatives on to Oracle, but we don’t think this short-term benefit outweighs the long-run potential harm of losing the revenue and control over OpenAI.
微软在 OpenAI 利用另一家基础设施合作伙伴时基本上被搁置。正如萨提亚·纳德拉所说：“我所知道的就是，我为我的 800 亿美元感到满意。”微软最近的博客文章陈述了合作关系的事实以及微软拥有优先购买权，但现实是，这对微软的长期发展是逐步负面的，因为他们害怕承担这项投资的风险。此外，负责 OpenAI 合作关系的克里斯·杨在公告后也突然离职。看涨者会说微软正在将其人工智能计划的一部分资本密集度转移到甲骨文，但我们认为这种短期利益无法弥补失去 OpenAI 的收入和控制权所带来的长期潜在损害。

While Elon Musk and Satya Nadella say the same thing, “Where is the actual money?” Elon Musk has a plan to grow their Memphis Tennessee site as fast as if not faster than OpenAI. Satya Nadella has no plan to build leading edge AI models internally.
尽管埃隆·马斯克和萨提亚·纳德拉都在问同样的问题：“实际的钱在哪里？”埃隆·马斯克有一个计划，旨在让他们在田纳西州孟菲斯的基地发展得和 OpenAI 一样快，甚至更快。萨提亚·纳德拉没有计划在内部构建领先的人工智能模型。

Arm – Technology Partner or Fake News
臂 – 技术合作伙伴还是假新闻

Arm rallied ~16% on the news because they were named a technology partner, but only because of the Grace and Vera CPUs that accompany Blackwell and Rubin GPUs all from Nvidia. SoftBank likely pushed for Arm to get on the PR and the optics look nice. Arm isn’t doing much.
Arm 因被指定为技术合作伙伴而上涨了约 16%，但这仅仅是因为与 Nvidia 的 Blackwell 和 Rubin GPU 配套的 Grace 和 Vera CPU。软银可能推动 Arm 参与公关，效果看起来不错。Arm 并没有太多作为。

The reality for Arm shareholders, as stated above, is SoftBank will likely have to sell off a chunk of its stake in the company to fund part of the equity check for Stargate. We think investors are largely missing this point and mistakenly view this announcement as material incremental good news.
对于 Arm 的股东来说，如上所述，软银可能需要出售其在公司的部分股份，以为 Stargate 的股权支出提供资金。我们认为投资者在很大程度上忽视了这一点，错误地将此公告视为实质性的增量好消息。

Oracle is placed with a huge burden of both tackling datacenter management and managing supply chain logistics. It is quite a task to manage a 1.8GW cluster and procure all necessary equipment at the command of another company. Nvidia is obviously involved as it’s their hardware. OpenAI will remain in control of operations and run almost all its own cluster management software.
甲骨文面临着巨大的压力，既要处理数据中心管理，又要管理供应链物流。管理一个 1.8GW 的集群并在另一家公司指挥下采购所有必要设备是一项相当艰巨的任务。英伟达显然参与其中，因为这是他们的硬件。OpenAI 将继续控制运营，并几乎运行自己所有的集群管理软件。

The Future of The Stargate Joint Venture
星际之门合资企业的未来

The scale of this project is impressive, but it does feel like the next step in an exuberant progression of the AI cycle, as we recently wrote about in Fabricated Knowledge. We know the first $100B will be spent in Abilene, and that there will be more campuses developed for the remaining $400B. Even more impressive, all these datacenter campuses will continue to be built out with the aim of distributed training.
该项目的规模令人印象深刻，但确实感觉像是人工智能周期中一个充满活力的进展的下一步，正如我们最近在《虚构知识》中所写的那样。我们知道前 1000 亿美元将用于阿比林，剩余的 4000 亿美元将开发更多校园。更令人印象深刻的是，所有这些数据中心校园将继续建设，旨在实现分布式训练。

Our Datacenter Model subscribers will be some of the first to find out where the remaining sites are. Keep in mind, simply repeating the Abilene campus four more times yields 9GW of Critical IT Capacity entirely allocated for one training.
我们的数据中心模型订阅者将是首批了解剩余站点位置的人。请记住，仅仅将阿比林校园重复四次，就可以获得 9GW 的关键 IT 容量，完全分配用于一次培训。

We think Marvell and Ciena will be the biggest winners from the continued DCI build-out, while Fabrinet will benefit a decent bit and Coherent / Lumentum will benefit to a smaller degree.
我们认为 Marvell 和 Ciena 将是持续 DCI 建设中最大的赢家，而 Fabrinet 将获得相当的好处，Coherent / Lumentum 的受益程度则较小。