How Elon Musk’s Supercomputer Freaked Out AI Rivals
埃隆·马斯克的超级计算机如何让人工智能竞争对手感到恐慌

The spy plane in Memphis is one sign of the high stakes involved in one of the most expensive races in the history of technology. Microsoft, Meta Platforms, Google and Amazon are each pouring tens of billions of dollars into new data centers to power the advanced new forms of AI that undergird ChatGPT and other applications.
孟菲斯的间谍飞机是技术史上最昂贵竞赛之一所涉及高风险的一个标志。微软、Meta Platforms、谷歌和亚马逊各自投入数百亿美元建设新的数据中心，以支持支撑 ChatGPT 和其他应用程序的先进新型人工智能。

It’s a risky bet based on a simple belief that the bigger a cluster of servers is, the better the AI it can produce. The scramble to supersize those clusters began in late 2022 with OpenAI’s launch of ChatGPT, the chatbot whose popularity sent shock waves across the tech industry.
这是一项基于简单信念的风险投资，即服务器集群越大，能够产生的人工智能就越好。为了扩大这些集群的规模的争夺战始于 2022 年底，OpenAI 推出了 ChatGPT，这款聊天机器人因其受欢迎程度在科技行业引发了震动。

Musk—who co-founded and initially funded OpenAI but later split from it—joined the data center race after it had already started. But through a combination of ambition, indefatigability and disregard for some of the conventional methods of building data centers, he has still managed to make a major splash.
马斯克——他是 OpenAI 的联合创始人并最初提供资金，但后来与其分道扬镳——在数据中心竞赛已经开始后才加入其中。但通过雄心、精力充沛以及对一些传统数据中心建设方法的无视，他仍然成功地引起了重大关注。

Two things about Musk’s supercomputer have jolted competitors: its size and the speed with which xAI built it. The supercomputer, fittingly known as Colossus, consists of 100,000 graphics processing units, the chips best suited to training and running AI software. That is several times bigger than similar supercomputers built in the past by Meta and other tech giants.
关于马斯克的超级计算机，有两件事让竞争对手感到震惊：其规模和 xAI 构建它的速度。这台名为“巨人”的超级计算机由 100,000 个图形处理单元组成，这些芯片最适合用于训练和运行人工智能软件。这比 Meta 和其他科技巨头过去建造的类似超级计算机大几倍。

Stringing together so many GPUs into a single supercomputer isn’t as simple as it sounds because of how much power the servers consume and bottlenecks in the networking equipment used to connect the chips to each other. And completing the project as quickly as xAI did is unheard of.
将如此多的 GPU 串联成一台超级计算机并不像听起来那么简单，因为服务器消耗的电力以及用于将芯片相互连接的网络设备中的瓶颈。而像 xAI 那样迅速完成项目是前所未闻的。

Musk and Nvidia, the AI chip powerhouse that supplied the GPUs for Colossus, said the data center and supercomputer were built in just 122 days. On a recent podcast, Nvidia CEO Jensen Huang said a GPU cluster of that size would normally take three years to plan and design and an additional year to get working.
马斯克和为巨型计算机提供 GPU 的人工智能芯片巨头英伟达表示，该数据中心和超级计算机仅在 122 天内建成。在最近的一期播客中，英伟达首席执行官黄仁勋表示，通常这样规模的 GPU 集群需要三年的时间进行规划和设计，以及额外一年的时间才能投入使用。

“No question that nobody slept,” Huang said of the Colossus project on a recent podcast.
“毫无疑问，没有人睡觉，”黄在最近的一期播客中谈到巨像项目时说道。

“As far as I know, there’s only one person in the world who could do that,” Huang added. “Elon is singular in this understanding of engineering and construction and large systems and marshaling resources.”
“据我所知，世界上只有一个人能做到这一点，”黄补充道。“埃隆在工程、建筑、大型系统以及调动资源方面的理解是独一无二的。”

Musk appears to have built the Memphis data center so quickly in part by cutting a few key corners—for example, by moving forward without having secured enough power from the electric grid to run Colossus. But defying such norms is part of the playbook Musk has used again and again at his other companies.
马斯克似乎通过在一些关键环节上削减成本，迅速建成了孟菲斯数据中心——例如，在没有从电网获得足够电力以运行巨兽的情况下就开始了建设。但无视这些规范是马斯克在其他公司一再使用的策略之一。

At Tesla, for example, he once sidestepped the need for permits to expand a car factory in California by setting up an assembly line in a parking lot for its Model 3 vehicle. And at SpaceX, he constantly pushes engineers to get rid of parts on its rockets that he views as unnecessary or to use cheaper components that were not designed for space applications.
在特斯拉，例如，他曾通过在加利福尼亚的停车场为 Model 3 车辆建立一个装配线，绕过扩建汽车工厂所需的许可证。而在 SpaceX，他不断推动工程师去除他认为不必要的火箭部件，或者使用未设计用于太空应用的更便宜的组件。

Even though xAI’s AI tools are still far behind those of OpenAI, the speed with which he built his supercomputer raised alarms with Sam Altman, OpenAI’s CEO. After Musk posted on X about it, Altman got into an argument with infrastructure executives at Microsoft, telling them he was concerned xAI was moving faster than Microsoft, according to a person who heard his remarks.
尽管 xAI 的人工智能工具仍远远落后于 OpenAI，但他构建超级计算机的速度引起了 OpenAI 首席执行官萨姆·阿尔特曼的警觉。根据一位听到他发言的人士的说法，在马斯克在 X 上发布相关内容后，阿尔特曼与微软的基础设施高管发生了争执，告诉他们他担心 xAI 的进展速度超过了微软。

He worried that xAI could soon have a more powerful supercomputer than OpenAI did. That concern has prompted OpenAI to seek alternatives to Microsoft for the first time.
他担心 xAI 很快会拥有比 OpenAI 更强大的超级计算机。这一担忧促使 OpenAI 首次寻求微软以外的替代方案。

Right now, one of those alternatives is under construction along a dusty stretch of flat land in Abilene, Texas, roughly three hours west of Dallas, where a group of companies is preparing the site for a data center that will eventually house a 100,000-chip cluster for OpenAI next year.
目前，其中一个替代方案正在德克萨斯州阿比林的一片尘土飞扬的平坦土地上建设，距离达拉斯大约三个小时的车程，一组公司正在为一个数据中心准备场地，该数据中心将于明年最终容纳一个 10 万个芯片的集群，供 OpenAI 使用。

Construction at the site is moving quickly. On a recent tour, a guide who worked for one of the project’s contractors pointed out how most of the buildings don’t yet have walls on all of their sides. The contractors are building most of the facility’s components off-site so that they can install parts quickly when they arrive.
工地的建设进展迅速。在最近的一次参观中，一位为项目承包商工作的导游指出，大多数建筑物的四面尚未完全封闭。承包商正在场外建造大部分设施的组件，以便在到达时能够快速安装部件。

It may not be long before even the supercomputers in Abilene and Memphis look relatively small. Some big tech companies, including Microsoft, have discussed data center projects that would contain millions of GPUs and cost more than $100 billion each.
可能不久之后，阿比林和孟菲斯的超级计算机看起来也会相对较小。一些大型科技公司，包括微软，已经讨论了将包含数百万个 GPU 且每个项目成本超过 1000 亿美元的数据中心项目。

The one-upmanship is likely to continue, as nearly everyone in the data center business is keeping close tabs on what their competitors are up to.
这种攀比可能会继续，因为几乎所有数据中心行业的人都在密切关注竞争对手的动向。

“The data center market is very small, and everybody is paying attention to what is going on,” said John Arcello, who leads the advanced data center team at DPR Construction, which builds data centers for large companies including Meta and is working on the Abilene project.
“数据中心市场非常小，大家都在关注发生了什么，”DPR Construction 的高级数据中心团队负责人约翰·阿切洛（John Arcello）表示，该公司为包括 Meta 在内的大型公司建设数据中心，并正在进行阿比林项目。

‘Gigafactory of Compute’ “计算的超级工厂”

Early this year, Musk began putting together the computing horsepower he needed to build xAI, a company he had founded in 2023 as a serious contender in AI.  At the time, he was already renting GPUs from Oracle to train the initial version of Grok, xAI’s large language model.
今年年初，马斯克开始组建他所需的计算能力，以建立 xAI，这是一家他在 2023 年创立的人工智能领域的有力竞争者。当时，他已经在从甲骨文租用 GPU，以训练 Grok 的初始版本，Grok 是 xAI 的大型语言模型。

To improve Grok’s quality, he needed access to much more computing capacity. In May, he held a video call with prospective xAI investors as part of an effort to raise billions of dollars for the startup. He laid out a vision to them of building the world’s largest supercomputer, which he called the “gigafactory of compute”—a reference to Tesla’s gargantuan factories around the world—according to an investor who attended the meeting.
为了提高 Grok 的质量，他需要获得更多的计算能力。5 月，他与潜在的 xAI 投资者进行了视频通话，作为筹集数十亿美元资金以支持初创公司的努力的一部分。他向他们阐述了建立世界上最大的超级计算机的愿景，他称之为“计算的千兆工厂”——这是对特斯拉在全球的巨大工厂的引用——根据一位参加会议的投资者所说。

Huddled around a table with fewer than a dozen other xAI employees, Musk revealed his plan to connect 100,000 of Nvidia’s H100s—the most advanced GPUs from the company then on the market—into a single cluster. A chart on the screen said xAI would build its supercomputer in about one-fifth the time it would take most companies to do so.
在与不到十名其他 xAI 员工围坐在一张桌子旁时，马斯克透露了他的计划，即将 100,000 个 Nvidia 的 H100——当时市场上最先进的 GPU——连接成一个单一的集群。屏幕上的一张图表显示，xAI 将以大约五分之一的时间建造其超级计算机，而大多数公司则需要更长的时间。

One of the xAI slides said that the company was operating at “ludicrous speed” and promised that “Elon is personally responsible for delivering the data center on time.”
xAI 的一张幻灯片表示，该公司正在以“荒谬的速度”运营，并承诺“埃隆亲自负责按时交付数据中心。”

Musk told investors he hadn’t yet decided whether xAI would partner with a cloud provider on the project or proceed on its own.
马斯克告诉投资者，他尚未决定 xAI 是否会与云服务提供商合作进行该项目，还是独立进行。

A few weeks later, a handful of Oracle executives logged onto a video meeting with Musk to discuss the first option. Musk proposed that Oracle—whose founder, Larry Ellison, is a close friend of Musk—build xAI’s supercomputer, which would have made the AI startup one of Oracle’s largest customers, The Information previously reported.
几周后，一小部分甲骨文高管与马斯克进行视频会议，讨论第一个选项。马斯克提议甲骨文——其创始人拉里·埃里森是马斯克的密友——为 xAI 构建超级计算机，这将使这家人工智能初创公司成为甲骨文最大的客户之一，《信息》之前报道。

Musk wanted the xAI data center to be located in a former Electrolux appliance manufacturing facility in Memphis, and he wanted it done by fall 2024. But on the call, Oracle executives told Musk they didn’t think they could build it as quickly as he was demanding, according to someone who attended the meeting.
马斯克希望 xAI 数据中心位于孟菲斯的一家前 Electrolux 家电制造厂，并希望在 2024 年秋季之前完成。但根据一位参加会议的人士的说法，甲骨文高管在电话中告诉马斯克，他们认为无法像他要求的那样快速建造。

Oracle staff noted that the building Musk wanted to use didn’t have access to enough power to support the number of chips he wanted Oracle to put in it, the person said. Musk quickly grew frustrated with the pushback from the Oracle executives.
甲骨文的员工指出，马斯克想要使用的建筑没有足够的电力来支持他希望甲骨文放入的芯片数量，这位人士表示。马斯克对甲骨文高管的反对迅速感到沮丧。

Ultimately Musk decided xAI would work on the Memphis data center without Oracle.
最终，马斯克决定 xAI 将在没有 Oracle 的情况下在孟菲斯数据中心工作。

“Oracle is a great company…but, when our fate depends on being the fastest by far, we must have our own hands on the steering wheel, rather than be a backseat driver,” he said on X, following publication of a story in The Information about how the talks broke down.
“甲骨文是一家伟大的公司……但是，当我们的命运完全依赖于速度时，我们必须亲自掌握方向盘，而不是当个后座司机，”他在 X 上说道，此前《信息》发布了一篇关于谈判破裂的报道。

Power Plans 电力计划

To meet his aggressive timeline, Musk pushed local officials in Memphis to approve the data center in record time. Luckily for xAI, Memphis was eager to give him what he wanted so the city could attract his business.
为了满足他激进的时间表，马斯克推动孟菲斯的地方官员以创纪录的速度批准数据中心。幸运的是，对于 xAI 来说，孟菲斯渴望满足他的需求，以便吸引他的业务。

“We worked even longer hours, took every text and phone call at any hour of the day so that we were reflecting this sense of drive that matches this company and its expectations,” Ted Townsend, chair of the Greater Memphis Chamber, told the Daily Memphian.
“我们工作了更长的时间，随时接听每一条短信和电话，以反映出与这家公司及其期望相匹配的动力感，”大孟菲斯商会主席泰德·汤森对《每日孟菲斯人》说道。

In early June, Townsend announced publicly that Musk had picked Memphis as the location for xAI’s supercomputer.
在六月初，汤森德公开宣布马斯克选择孟菲斯作为 xAI 超级计算机的地点。

Over the next few weeks, Musk and his team at xAI gutted the Memphis manufacturing facility to make room for rows and rows of racks that would house the Nvidia GPUs. They set up electrical, mechanical and plumbing equipment and installed a water cooling system for the servers.
在接下来的几周里，马斯克和他在 xAI 的团队对孟菲斯制造设施进行了彻底改造，以腾出空间放置一排排将容纳 Nvidia GPU 的货架。他们安装了电气、机械和管道设备，并为服务器安装了水冷系统。

One major hitch for Musk’s breakneck construction schedule could have been electrical power. Initially, there wasn’t enough of it at the Memphis site for all of xAI’s power-hungry GPUs. Normally, that’s the kind of problem that could derail or delay a data center project.
马斯克快速建设进度的一个主要障碍可能是电力。最初，孟菲斯现场的电力不足以满足 xAI 所有耗电量大的 GPU。通常，这种问题可能会使数据中心项目脱轨或延迟。

But Musk came up with a stopgap solution: He brought in mobile, natural gas–powered turbines to provide supplemental power while he waited for local authorities to approve a request for an additional 100 megawatts of power at the site. The Tennessee Valley Authority agreed to that request last week.
但是马斯克想出了一个权宜之计：他引入了移动的天然气发电机组，以提供补充电力，同时等待当地政府批准在该地点增加 100 兆瓦电力的请求。田纳西河谷管理局上周同意了该请求。

Musk’s move prompted immediate pushback from a coalition of local environmental groups, who wrote to the local health department that xAI was polluting the air by operating gas combustion turbines without permits. An executive who works on data centers at Microsoft said there’s no way the company would have been able to do something similar, given its climate goals and initiatives.
马斯克的举动引发了当地环保组织联盟的立即反对，他们向当地卫生部门写信称，xAI 在没有许可证的情况下运营燃气燃烧涡轮机，正在污染空气。一位在微软数据中心工作的高管表示，考虑到公司的气候目标和举措，微软绝对不可能做类似的事情。

“To have basically an un-permitted power plant come in and set up shop is alarming and really disrespectful to the community,” said Amanda Garcia, a senior attorney at the Southern Environmental Law Center, which opposed the Tennesse Valley Authority’s decision. “Air pollution is a huge challenge in southwest Memphis.”
“让一个基本上没有许可证的发电厂进驻并设立营运是令人震惊的，并且对社区非常不尊重，”南方环境法律中心的高级律师阿曼达·加西亚说，该中心反对田纳西河谷管理局的决定。“空气污染是西南孟菲斯面临的巨大挑战。”

Other factors likely helped Musk finish the project quickly. For example, executives in the data center business said Colossus almost certainly didn’t have to undergo and pass any compliance tests before xAI could start using the cluster. That’s mainly because xAI was planning to use the supercomputer for its own needs rather than renting it out to other clients.
其他因素可能帮助马斯克迅速完成该项目。例如，数据中心业务的高管表示，Colossus 几乎肯定不需要在 xAI 开始使用该集群之前进行并通过任何合规测试。这主要是因为 xAI 计划将超级计算机用于自身需求，而不是将其租赁给其他客户。

In contrast, Microsoft has to undergo several data security tests before handing over servers to OpenAI or other Azure cloud customers, who expect a certain level of uptime or privacy standards, according to someone with direct knowledge of the process.
相比之下，微软在将服务器交给 OpenAI 或其他 Azure 云客户之前，必须经过几项数据安全测试，这些客户期望达到一定的正常运行时间或隐私标准，依据一位对该过程有直接了解的人士。

“We have all these different industry certifications that we have to pass,” said Raul Martynek, CEO of data center operator DataBank. “I’ll guarantee you the [xAI] data center can’t pass those types of things.”
“我们有很多不同的行业认证必须通过，”数据中心运营商 DataBank 的首席执行官 Raul Martynek 说。“我可以保证[xAI]数据中心无法通过这些类型的认证。”

Musk’s effort to stand up Colossus has encountered a fair amount of skepticism. Several data center executives said it is extremely difficult to retrofit buildings, such as manufacturing plants, for GPU servers and liquid cooling systems. Over the past few months, the data center has experienced outages, according to two people who have spoken with xAI employees.
马斯克建立巨像的努力遭遇了相当多的怀疑。几位数据中心高管表示，将制造工厂等建筑改造为 GPU 服务器和液冷系统是极其困难的。根据与 xAI 员工交谈的两个人的说法，过去几个月该数据中心经历了停机。

The problems haven't appeared to slow xAI down. Musk and Nvidia have said they started the first training run for the next Grok model only 19 days after they brought the first server rack into the data center.
这些问题似乎并没有减缓 xAI 的进展。马斯克和英伟达表示，他们在将第一台服务器机架引入数据中心仅 19 天后，就开始了下一个 Grok 模型的第一次训练。

In a recent interview, Antonio Gracias—a close friend of Musk’s and a longtime investor in his companies, including xAI—said xAI is rethinking the entire process of building data centers “from first principles and trying to make it cheaper, better, faster.
在最近的一次采访中，安东尼奥·格拉西亚斯——马斯克的密友以及他公司（包括 xAI）的长期投资者——表示，xAI 正在从“第一原则”重新思考构建数据中心的整个过程，试图使其更便宜、更好、更快。

“I’ve seen this movie at Tesla, SpaceX, other companies, where it’s Elon, but it is also dozens of engineers being led in a mission to create the very best, most effective system possible,” he said.
“我在特斯拉、SpaceX 和其他公司见过这样的情况，虽然是埃隆在领导，但也是数十名工程师在共同努力，致力于创造出最好的、最有效的系统，”他说。

Peer Pressure 同伴压力

As word of Musk’s supercomputer progress spread this summer, senior data center executives at Amazon, Microsoft and Google began calling staffers at Nvidia with versions of the same question: How was Musk able to move so quickly on his supercomputer?
随着关于马斯克超级计算机进展的消息在这个夏天传播，亚马逊、微软和谷歌的高级数据中心高管开始联系英伟达的员工，问同样的问题：马斯克是如何如此迅速推进他的超级计算机的？

Officials from some of these and other companies, including Meta, also called a small competing cloud provider to see whether the company could provide them with data center capacity faster than what they could build themselves, according to someone who spoke to the companies.
一些这些公司及其他公司的官员，包括 Meta，也联系了一家小型竞争云服务提供商，询问该公司是否能够比他们自己建设更快地提供数据中心容量，依据一位与这些公司交谈过的人士。

Their eagerness to unpack the mysteries of the project has only increased as more information about the Memphis data center has trickled out. Data center and cloud executives have closely studied images from the facility to see what they can glean about its design.
他们对揭开该项目神秘面纱的渴望随着关于孟菲斯数据中心的信息逐渐披露而愈发增强。数据中心和云计算高管们仔细研究了该设施的图像，以了解其设计的相关信息。

Musk has published several photos from inside the data center on X. And last month, an industry analyst published a video on YouTube after taking a tour of the Colossus data center (the video, an unusual peek behind the curtain of the facility, was sponsored by Super Micro Computer, which supplied some of the data center servers to Musk).
马斯克在 X 上发布了几张数据中心内部的照片。上个月，一位行业分析师在参观了巨人数据中心后在 YouTube 上发布了一段视频（这段视频是对该设施的非同寻常的幕后探秘，由超级微型计算机赞助，该公司向马斯克提供了一些数据中心服务器）。

Meanwhile, Oracle—xAI’s would-be partner on the Memphis project—signed its deal to provide computing capacity to OpenAI not long after the talks with xAI broke down over the summer. The Information was first to report that the new OpenAI data center would be located in Abilene, where Oracle had signed an agreement to work with startups Crusoe and Lancium to develop the site.
与此同时，Oracle——xAI 在孟菲斯项目上的潜在合作伙伴——在与 xAI 的谈判于夏季破裂后不久签署了向 OpenAI 提供计算能力的协议。《信息》首次报道新的 OpenAI 数据中心将位于阿比林，Oracle 已与初创公司 Crusoe 和 Lancium 签署协议以开发该地点。

Last month, Crusoe raised more than $3 billion to develop the initial phase of the data center, which will contain 100,000 of Nvidia’s forthcoming GPUs, known as GB200s.
上个月，克鲁索筹集了超过 30 亿美元用于开发数据中心的初始阶段，该中心将包含 10 万个英伟达即将推出的 GPU，称为 GB200s。

As Musk did in Memphis, Crusoe is pushing to get the project done quickly. Arcello from DPR, whose firm is contracting with Crusoe on the project, said it’s one of the fastest builds he’s ever worked on. The firms began discussing the data center design in March and broke ground on the project in June.
正如马斯克在孟菲斯所做的，Crusoe 正在推动项目快速完成。与 Crusoe 合作承包该项目的 DPR 的 Arcello 表示，这是他参与过的最快的建设之一。两家公司于 3 月开始讨论数据中心设计，并于 6 月动工。

A few weeks ago, construction crews at the Abilene site were busy cutting down trees to make room for a new electrical substation and pouring as much concrete as they could truck in on a daily basis. OpenAI has asked its partners on the project to consider using gas turbines in case there are any issues getting power to the site on time, according to three people with direct knowledge of the request.
几周前，阿比林工地的施工队忙于砍伐树木，以为新的电力变电站腾出空间，并每天尽可能多地浇筑混凝土。根据三位知情人士的说法，OpenAI 已要求其项目合作伙伴考虑使用燃气涡轮，以防在按时为工地供电时出现任何问题。

On a recent tour at the site, a guide was asked why there was so much fuss about building the data center so quickly.
在最近的现场参观中，一位导游被问到为什么对如此快速建设数据中心如此关注。

“Whoever can get their [supercomputer] faster…could pretty much rule the world,” the guide said.
“谁能更快地获得他们的[超级计算机]…几乎可以统治世界，”向导说道。