内容来源：https://www.wired.com/story/openai-oracle-softbank-data-center-stargate-us/

内容总结：

OpenAI启动“星门”计划 在美国新建五大AI数据中心

人工智能公司OpenAI于本周二宣布，将启动名为“星门”的重大基础设施计划，在美国新建五个大型数据中心。此举旨在大幅提升其人工智能算力，以巩固美国在AI领域的全球领先地位。

巨量投资与合作伙伴

该计划由OpenAI与甲骨文、软银集团合作推进。据披露，这些新设施将使“星门”计划的总规划容量提升至近7吉瓦，其耗电量约相当于七座大型核电站的发电量。OpenAI首席执行官萨姆·奥尔特曼在德克萨斯州阿比林市的发布会上强调，AI发展离不开庞大的基础设施支撑，并称赞德克萨斯州的“创新精神”为如何实现“更大、更快、更便宜、更好”的扩展提供了范本。

具体布局与就业预期

五个新数据中心的地点分别位于：

• 德克萨斯州沙克尔福德县、新墨西哥州多纳安娜县和中西部一未公开地点：这三个站点由OpenAI与甲骨文合作开发。
• 俄亥俄州洛兹敦和德克萨斯州米拉姆县：这两个站点由OpenAI与软银旗下SB Energy公司共同主导。

OpenAI声称，这些新数据中心连同阿比林现有设施的600兆瓦扩容计划，将创造超过2.5万个现场工作岗位。

战略背景与中美竞争视角

“星门”计划被视为美国确保其AI领导地位的关键举措。OpenAI在一份政策白皮书中直言，将AI基础设施视为美国超越中国、重振制造业的重要工具，并指出中国在电力资源开发和AI基础设施部署上的速度令人瞩目。

这一观点也得到了部分美国政界的呼应。参议员特德·克鲁兹在发布会上表示，美国应采取“轻触式监管”，以赢得与中国的“AI竞赛”。

挑战与不确定性

尽管野心勃勃，但该计划也面临诸多挑战。行业专家指出，巨大的投资规模并不必然保证成功，并质疑科技行业是否在追求算力的过程中放弃了环保承诺。此外，自计划公布以来，贸易不确定性和硬件成本飙升已导致项目进展出现延迟，复杂的合作结构也可能带来潜在风险。

未来展望

与此同时，OpenAI的扩张步伐并未停歇。本周初，该公司还宣布与英伟达达成一项规模更大的新合作，旨在部署高达10吉瓦的数据中心容量。分析认为，OpenAI正押注于“规模越大，模型性能越好”的发展路径，但其最终能否找到足以支撑如此巨大投入的市场应用，仍有待时间检验。

中文翻译：

OpenAI周二宣布，作为"星门计划"的组成部分，公司将在美国新建五座数据中心。这些与甲骨文、软银合作开发的站点将使星门计划当前规划容量接近7吉瓦——相当于七座大型核电站的发电量。

"人工智能在许多方面与互联网不同，其中之一就是其所需的基础设施规模，"OpenAI首席执行官萨姆·奥尔特曼在德州阿比林市的发布会上表示。他强调美国"在这方面不能落后"，并称德州的"创新精神"为如何实现"更大、更快、更便宜、更好"的扩展提供了范本。

其中三处位于德州沙克尔福德县、新墨西哥州多纳安娜县以及中西部未公开地点的站点正与甲骨文合作开发。此举延续了双方七月达成的协议，即在阿比林市首座星门设施基础上新增4.5吉瓦数据中心容量。

OpenAI宣称这些新数据中心及阿比林站点计划扩建的600兆瓦容量将创造超过2.5万个现场岗位，不过数据中心建设所需劳动力通常远超后期运维需求。

其余两处站点由OpenAI与软银旗下太阳能和电池项目开发商SB Energy共同主导，分别位于俄亥俄州洛兹敦和德州米拉姆县。

星门计划是自特朗普总统年初就职以来宣布的多个美国重大科技基础设施项目之一。OpenAI在一月声明中称，这家ChatGPT制造商与软银、甲骨文、MGX达成的5000亿美元、10吉瓦投资承诺将"巩固美国在AI领域的领导地位"并"创造数十万就业岗位"。

特朗普重返白宫仅两天后便大力宣传这项庞大计划，承诺将加速美国人工智能发展，助力与中国等国的竞争。七月他宣布的AI行动计划要求在美国争夺先进AI主导权时加速基建并减少官僚程序。"我们认为正处于AI竞赛中，"白宫AI事务负责人戴维·萨克斯当时表示，"我们要让美国赢得竞赛。"

OpenAI最初将星门定位为由软银CEO孙正义担任主席的"新公司"。但如今项目相关高管透露，这实则是涵盖OpenAI所有数据中心项目的统一品牌（微软合作项目除外）。

据知情高管透露，阿比林的旗舰站点主要由甲骨文持有运营，OpenAI作为主要租户。由数据中心初创企业Crusoe负责的建设项目预计2026年中完工，目前已依托甲骨云基础设施支持OpenAI训练与推理任务。

甲骨文正在阿比林建设八个数据中心大厅，每个支持约100兆瓦功耗。知情人士称，当前一栋建筑已竣工，另一栋接近完工，全部落成后将容纳超40万块GPU，支持超1.4吉瓦功耗。

星门计划初始公告曾邀请拥有土地设备等资源的企业参与。OpenAI高管表示公告发布后收到大量合作咨询。

有高管估算OpenAI已评估全美约700处潜在地点。周二宣布的五个项目据称是与合作伙伴从30个州收到的300多份方案中筛选得出。

在周二发布的政策白皮书中，OpenAI将AI基础设施称为美国战胜中国、重塑制造业优势的关键工具。文件指出："如今中共领导下的中国正以空前速度发展电力资源，而美国正在落后。中国巨大的电力消耗不仅彰显其工业统治力，更体现其快速部署AI基础设施（包括数据中心、半导体晶圆厂和制造中心）的能力。"

本周初OpenAI还宣布与英伟达达成新战略合作，部署10吉瓦数据中心容量——规模甚至远超甲骨文合作项目。英伟达将向OpenAI投资高达1000亿美元，但双方未透露具体资金结构。英伟达表示首期建设将包含400-500万块GPU，公司将随容量部署"逐步推进"投资。

双方称首套英伟达Vera Rubin GPU系统将于2026年下半年投产。奥尔特曼向CNBC表示该项目将"拓展星门计划的雄心"。

研究计算效率数十年的加州大学伯克利分校客座教授乔纳森·库米指出，OpenAI巨额投资未必能保证成功。今年中国初创企业深度求索以较少算力打造出惊艳AI模型的案例表明，资源较少的竞争者仍可能超越OpenAI。

"OpenAI押注规模扩张仍是提升大模型性能的最佳途径。如果该判断正确——且能找到用户真正愿意付费的应用场景——那么一切都会顺利，"库米表示。但他补充道这一假设尚未得到验证。

库米还担忧科技行业可能为追求新数据中心容量而放弃环保承诺，质疑这种牺牲是否合理："我会质疑这种紧迫感的必要性。"

尽管Meta在路易斯安那州的大型数据中心等项目遭遇当地反对，但阿比林站点似乎获得政界强力支持。市长韦尔登·赫特与参议员特德·克鲁兹共同出席周二发布会力挺该项目。

"我认为应采用轻度监管模式，因为我们正在与中国赛跑，"克鲁兹对记者团表示。他强调若要让"自由、言论自由、自由企业和个人自由主导AI"，美国必须"赢得AI竞赛。这一切正在发生，这正是星门计划如此重要的原因。"

五月以来，OpenAI还宣布了旨在海外建设大型数据中心的"星门英国"和"星门阿联酋"计划。特朗普政府官员已协助与盟国安排部分基建协议，以期助力美国企业抢占全球市场份额。

但星门计划自启动以来始终伴随质疑。上月彭博社和华尔街日报均报道称贸易不确定性与硬件成本飙升导致项目进展延迟。

库米指出部分星门协议的复杂性可能成为另一隐患："从会计视角看极其复杂，若出现问题可能酿成危机。"

英文来源：

OpenAI is planning to build five new data centers in the United States as part of the Stargate initiative, the company announced on Tuesday. The sites, which are being developed in partnership with Oracle and SoftBank, bring Stargate’s current planned capacity to nearly 7 gigawatts—roughly the same amount of power as seven large-scale nuclear reactors.
“AI is different from the internet in a lot of ways, but one of them is just how much infrastructure it takes,” OpenAI CEO Sam Altman said during a press briefing in Abilene, Texas, on Tuesday. He argued that the US “cannot fall behind on this” and the “innovative spirit” of Texas provides a model for how to scale “bigger, faster, cheaper, better.”
Three of the new sites, in Shackelford County, Texas; Doña Ana County, New Mexico; and a yet-to-be-disclosed location in the Midwest, are being developed in partnership with Oracle. The move follows an agreement Oracle and OpenAI announced in July to develop up to 4.5 gigawatts of US data center capacity on top of what the two companies are already building at the first Stargate facility in Abilene.
OpenAI claims the new data centers, along with a planned 600 megawatt expansion of the Abilene site, will create more than 25,000 onsite jobs, though the number of workers required to build data centers typically dwarfs the amount needed to maintain them afterwards.
The two remaining sites are being helmed by OpenAI and SB Energy, a SoftBank subsidiary that develops solar and battery projects. These are located in Lordstown, Ohio, and Milam County, Texas.
Stargate is one of several major US technology infrastructure projects that have been announced since President Donald Trump took office at the start of the year. OpenAI said in January that the $500 billion, 10 gigawatt commitment between the ChatGPT maker, SoftBank, Oracle, and MGX would “secure American leadership in AI” and “create hundreds of thousands of American jobs.”
Trump touted the mammoth initiative just two days after he returned to the White House, promising that it would accelerate American progress in artificial intelligence and help the US compete against China and other nations. In July, Trump announced an AI action plan that called for speedy infrastructure development and limited red tape as the US tries to beat other countries in the quest for advanced AI. “We believe we’re in an AI race,” White House AI czar David Sacks said at the time. “We want the United States to win that race.”
OpenAI initially framed Stargate as a “new company” that would be chaired by Softbank CEO Masayoshi Son. Now, however, executives close to the project say it’s an umbrella brand name used to refer to all of OpenAI’s data center projects—except those developed in partnership with Microsoft.
The flagship site in Abilene is primarily owned and operated by Oracle, with OpenAI acting as the primary tenant, according to executives close to the project. The buildout, which is being managed by the data center startup Crusoe, is on track to be completed by mid-2026, sources close to the project say. It is already running on Oracle Cloud Infrastructure and supporting OpenAI training and inference workloads, those sources add.
Oracle is in the process of standing up eight data center halls in Abilene, which will each support roughly 100 megawatts of power. One of those buildings is currently complete, and another one is close to being finalized. When it is done, the facility will house more than 400,000 GPUs and support more than 1.4 gigawatts of power, sources close to the project say.
The initial Stargate announcement called for companies with land, equipment, and other relevant resources to get involved. After it was published, OpenAI executives say they were inundated with messages from firms that wanted to participate.
One executive estimates that OpenAI has surveyed around 700 sites in the US where data centers could potentially be built. For the five projects announced Tuesday, OpenAI claims it reviewed more than 300 proposals together with its partners from roughly 30 states.
In a policy white paper released Tuesday, OpenAI framed AI infrastructure as a crucial tool the United States will need to beat China and become a manufacturing powerhouse. “Today, Communist Party–led China is developing electricity resources at unmatched velocity while the US falls behind,” the document argues. “China’s immense electricity consumption underscores not only its industrial dominance but its ability to rapidly deploy AI infrastructure including new data centers, semiconductor fabs, and manufacturing hubs.”
Earlier this week, OpenAI also announced a new strategic partnership with Nvidia to deploy 10 gigawatts of data center capacity—dwarfing even the size of OpenAI’s deal with Oracle. As part of the arrangement, Nvidia is investing up to $100 billion in OpenAI, though neither company shared specifics about the exact structure of the funding. The initial build-out will include 4 to 5 million GPUs, Nvidia said, and the chipmaker will make investments “progressively” as each gigawatt is deployed.
OpenAI’s first Nvidia-powered system, which will be built with the hardware maker’s Vera Rubin GPUs, will be operational in the second half of 2026, the two companies said. In an interview with CNBC, Altman said that the undertaking would “expand on the Stargate ambitions.”
Jonathan Koomey, a visiting professor at UC Berkeley who has studied computing and data center efficiency for decades, says the enormous size of OpenAI’s data center investments may not be enough to guarantee the company’s success. Earlier this year, the Chinese startup DeepSeek debuted an impressive AI model built using relatively little computing power, a feat that suggested OpenAI could still be beat out by lesser-resourced competitors.
“OpenAI are making a bet that scale will continue to be the best way to get performance from large learning models, and if that’s true—and if they’re able to find applications that people really want and are willing to pay for—then it will all work out fine,” Koomey says. But that thesis has still yet to be proven, he adds.
Koomey is also concerned that the tech industry may be abandoning its environmental commitments in pursuit of new data center capacity, and he questions whether such sacrifices are really warranted. “I would question the sense of urgency,” he says.
While some data center projects, like a massive complex being developed by Meta in Louisiana, have received pushback from the local community, the site in Abilene appears to have strong support from local politicians. Mayor Weldon Hurt attended the press briefing alongside Senator Ted Cruz on Tuesday to champion the initiative.
“In my view, we should have a very light-touch regulatory approach because we are in a race with China,” Cruz told a group of reporters. He said that if “we want freedom and free speech and free enterprise and individual liberty to dictate AI” the United States needs to “win the race for AI. That's happening right now. And that's why Stargate is so important.”
Since May, OpenAI has also announced Stargate UK and Stargate UAE, which aim to build out large-scale data center projects abroad. Trump administration officials have helped arrange some of the infrastructure deals with allied nations, betting the deals would help American companies capture market share globally.
But questions have swirled around Stargate ever since its inception. Last month, Bloomberg and The Wall Street Journal both reported that progress had been delayed by trade uncertainty and soaring hardware costs.
Koomey says the convoluted nature of some of the Stargate deals could prove to be another issue. “From an accounting perspective it’s very complicated and it has the potential to be a problem if things go south,” he says.