智东西（公众号：zhidxcom）
编译 | 孙小雯
编辑 | 程茜
智东西10月28日消息，昨日，AMD宣布将出资认购Meta新发行的股票，投资金额为10亿美元（约合人民币67.2亿元）。其同时披露，首批Helios机架级AI计算系统已在Meta安装并开展测试，两家公司将继续扩大芯片、计算系统和软件合作，支持Meta旗下模型及AI服务。
Helios将GPU、服务器CPU、网络设备及配套软件整合在同一机架中，用于大模型训练和推理。今年2月，双方已宣布长期合作，Meta计划采用AMD GPU建设最高6吉瓦规模的AI计算基础设施。
目前，Meta正在建设四个容量分别超过1吉瓦的超级计算集群，计划于今年内将基础设施总容量较去年扩大近一倍。
AMD的投资是Meta整体融资计划的一部分。据悉，Meta正推进一轮股权融资计划，规模已达975亿美元（约合人民币6552亿元），所筹资金将支持数据中心建设、算力扩充及Muse个人智能体等业务发展。据路透社报道，本轮为Meta公司历史上规模最大的一次融资。除AMD外，英伟达也于昨日宣布，计划投资30亿美元（约合人民币201.6亿元）认购Meta股份。
一、Meta时隔13年再发新股，融资计划增至975亿美元
Meta于10月26日公布融资计划时，目标规模为950亿美元（约合人民币6384亿元）。据美国新闻网站Semafor报道，这是Meta自2013年以来首次通过公开发行新股筹资。
这笔资金分为两部分：通过此次普通股和强制可转优先股发行筹集350亿美元（约合人民币2352亿元）；其余最高600亿美元（约合人民币4032亿元）拟通过后续按市价发行股票筹集，相关计划预计于2027年第一季度启动。
10月27日，Meta将本轮普通股发行规模增加25亿美元（约合人民币168亿元）。调整后，本轮融资规模升至375亿美元（约合人民币2520亿元），后续600亿美元的发行计划保持不变，整体融资计划因此增至975亿美元。
此次融资将为Meta扩建AI基础设施提供资金支持。据Meta财报，今年第三季度，Meta资本支出达到349.1亿美元（约合人民币2345.95亿元），自由现金流为负21.2亿美元（约合人民币142.46亿元）。Meta预计，全年资本支出将达到1350亿至1450亿美元（约合人民币9072亿至9744亿元）。
目前，Meta正在建设四个容量分别超过1吉瓦的超级计算集群，计划今年将基础设施总容量较去年扩大近一倍。这些算力将用于模型训练、内容推荐、广告优化及新产品。
二、Muse上线一个半月获千万用户，但算力成本制约全球推广
Meta扩建算力的一个重要用途，是支持旗下个人智能体Muse的推广。这款于9月推出的产品，可以代用户操作网页、处理邮件、预订行程，在用户关闭应用后继续执行任务，并在发送邮件、付款等操作前请求确认。
Meta创始人兼CEO马克·扎克伯格在10月21日的财报电话会上，将Muse称为公司历来最成功的产品发布之一。他透露，上线约一个半月，美国和加拿大已有超过1000万人配置Muse，约25%的用户每天使用。
但要将Muse推向更多地区，Meta还需要解决计算成本问题。扎克伯格称，Muse运行所需算力较大，是尚未全球推出的原因之一。Meta正扩建计算基础设施，目标是在未来几年具备服务超过10亿人的能力。
AMD已经为Muse提供底层硬件支持。据此次公告披露，其EPYC服务器CPU为Muse提供专用云端计算资源，支持运行浏览器、执行代码和调用工具。此外，AMD的Instinct GPU也已用于Meta的模型训练、排序和推荐系统。
扩充算力之外，Meta还在降低运行成本。该公司计划利用闲时算力、采用更高效的模型并优化推理，预计未来一年将单位服务所需的计算成本降至目前的十分之一以下。AMD与Meta则在推进针对Meta计算任务优化的定制GPU，并共同改进软件，提高硬件运行这些任务的效率。
此次投资进一步加深了双方在采购之外的联系。今年2月，AMD已向Meta授予最多1.6亿股股票的认购权。Meta在满足GPU采购、AMD股价及其他约定条件后，可按每股0.01美元（约合人民币0.067元）认购。如今，AMD也计划出资认购Meta股份，从芯片供应商进一步成为这位大客户的投资方。
结语：芯片厂商出资支持，但投资者需区分生态内投资与独立客户需求
据美国财经媒体MarketWatch报道，圣克拉拉大学金融学教学教授Devon Coombs称，随着AI企业之间的资金联系加深，投资者需要区分，企业收入有多少来自独立客户需求，又有多少来自生态内投资所支持的支出。
此次AMD出资支持Meta，有助于这位大客户继续扩建算力，也让自身的芯片业务与投资回报更加依赖Meta的发展。目前，Muse已经吸引千万用户，接下来降低服务成本、扩大付费收入的进展，将为这轮算力投入的商业回报提供更直接的依据。
来源：AMD、Meta、路透社
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Compiled by Sun Xiaoxuan
Editor.
According to the news of October 28, yesterday, AMD announced that it would make a contribution of $1 billion (approximately RMB 6.72 billion) for the purchase of new shares issued by Meta. At the same time, it was disclosed that the first Helios shelf-level AI computing systems had been installed and tested in Meta, and that the two companies would continue to expand their cooperation on chips, computing systems and software to support the Meta-based model and AI services.
Helios integrated GPU, server CPU, network equipment and supporting software into the same rack for LLM training and reasoning. In February this year, the two sides announced long-term cooperation, and Meta plans to use AMD GPU to build an AI computing infrastructure of up to 6 GW.
At present, Meta is building four supercalculator clusters with capacity of more than 1 gigawatt each, and plans to expand the total infrastructure capacity nearly twice as much this year as last year.
AMD investments are part of Meta ' s overall financing scheme. It is known that Meta is moving forward with a stock financing package of $975.5 billion (approximately $655.2 billion yuan), which will support business developments such as the construction of data centres, the expansion of computing power and the development of the Muse personal AI agent. According to Reuters, this round is the largest single financing in Meta's history. In addition to AMD, British Wida announced yesterday that it planned to invest $3 billion (approximately RMB 20.16 billion) in the purchase of Meta shares.
I. META ' S REGULATION IN 13 YEARS AGAIN, THE FINANCING PLAN TO GO TO US$ 97.5 billion
When Meta published its financing plan on 26 October, the target was $95 billion (approximately $6.384.4 billion yuan). According to the United States news website Semafor, this was the first time since 2013 that Meta had raised funds through a new public distribution unit.
The fund is divided into two parts: $35 billion (approximately $235.2 billion RMB) is to be raised through this general and mandatory transferable priority distribution; and the remaining $60 billion (approximately $403.2 billion RMB) is to be raised through a follow-up market-priced stock, which is expected to start in the first quarter of 2027.
On 27 October, Meta increased the size of the current round of general stock distributions by $2.5 billion (approximately RMB 16.8 billion). As a result of this adjustment, the scale of the current round rose to $37.5 billion (approximately $25.2 billion RMB) and the subsequent release plan of $60 billion remained unchanged, bringing the overall financing plan to $97.5 billion.
This financing will provide financial support for Meta ' s expansion of AI infrastructure. According to the Meta financial report, in the third quarter of this year, the capital expenditure of Meta amounted to $34.91 billion (approximately $23.4595 billion RMB) and the free cash flow was negative $2.12 billion (approximately $14.24.6 billion RMB). Meta expects capital expenditure to reach between $135.0 billion and $14.5 billion (approximately 90.7 billion to 97.4 billion yuan RMB) throughout the year.
At present, Meta is building four supercalculator clusters with a capacity of more than 1 Gwa each, and plans are to double the total infrastructure capacity this year over last year. These calculations will be used for model training, content referral, advertising optimization and new products.
ii. Muse has access to millions of users in a month and a half, but computing costs constrain global outreach
An important use of Meta ' s expansion calculus is to support the promotion of AI agentMuse, the individual under the flag. The product, which was launched in September, allows the user to operate the web page, process mail, book itineraries, continue the mission after the user has closed the application and request confirmation before the mail is sent, payments etc.
The founder of Meta and CEO Mark Zuckerberg described Muse as one of the company ' s most successful products, publish, at a press conference on 21 October. He revealed that for about a month and a half, over 10 million people in the United States and Canada had been equipped with Muse, and about 25 per cent of users used it daily.
However, to move Muse to more areas, Meta also needs to address the issue of costing. According to Zuckerberg, one of the reasons for the lack of global roll-out is the fact that Muse is operating with a great deal of effort. Meta is expanding its computing infrastructure with the goal of having the capacity to serve more than 1 billion people in the coming years.
AMD has already provided ground-level hardware support to Muse. As disclosed in this bulletin, its EPYC server CPU provides a dedicated cloud computing resource for Muse to support the running of browsers, implementation codes and call tools. In addition, AMD's Instinct GPU has been used for Meta's model training, sequencing and referral systems.
In addition to the power of expansion, Meta is reducing its operating costs. The company plans to use idle computing, introduce more efficient models and optimize reasoning, and expects to reduce the computational cost of unit services to less than one tenth of the current level in the coming year. AMD and Meta are promoting customized GPUs for Meta computing mission optimization and are working together to improve software and improve the efficiency of hardware in running these tasks.
The investment further deepened the link between the parties outside procurement. In February of this year, AMD has granted Meta up to 160 million shareholdings. Meta may subscribe to the amount of US$0.01 per share (approximately RMB 0.067) after meeting GPU procurement, AMD share price and other agreed conditions. Today, AMD also plans to fund the purchase of Meta shares, further becoming an investor to this large client from the chip supplier.
Conclusion: chip manufacturers provide financial support, but investors need to distinguish between ecological investment and independent customer needs
According to the United States financial media, Marketwatch, Devon Coombs, a professor of finance at the University of Santa Clara, stated that, as the financial links between AI enterprises deepened, investors needed to distinguish between how much business income was generated by the needs of independent clients and how much was financed by ecological investment.
This AMD-funded support to Meta helped the large client to continue to expand his budget and make her chip business and return on investment more dependent on Meta's development. At present, Muse has attracted millions of users, and subsequent progress in reducing the cost of services and expanding paid revenue will provide a more direct basis for the commercial returns on this round of calculus inputs.
Source: AMD, Meta, Reuters
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.