经济学人｜Day1004-AI淘金热：英伟达不是唯一获利的公司

文章来源：《经济学人》May 27th 2023 期 Business 栏目

Nvidia is not the only firm cashing in on the AI gold rush

英伟达并不是唯一一家在人工智能淘金热中获利的公司

There is AI in them thar hills

Selling specialist chips and infrastructure is becoming a trillion-dollar industry

销售专业芯片和基础设施正在成为一个价值数万亿美元的产业

May 29th 2023

A GREY RECTANGULAR building on the outskirts of San Jose houses rows upon rows of blinking machines. Tangles of colourful wires connect high-end servers, networking gear and data-storage systems. Bulky air-conditioning units whirr overhead. The noise forces visitors to shout.

在圣何塞郊区，一座灰色的长方形建筑里摆放着一排排闪烁的机器。五颜六色的电线连接着高端服务器、网络设备和数据存储系统。笨重的空调机组在头顶上嗡嗡作响。噪音迫使游客大声喊叫。

The building belongs to Equinix, a company which leases data-centre space. The equipment inside belongs to companies from corporate giants to startups, which are increasingly using it to run their artificial-intelligence (AI) systems. The AI gold rush, spurred by the astounding sophistication of “generative” systems such as ChatGPT, a hit virtual conversationalist, promises to generate rich profits for those who harness the technology’s potential. As in the early days of any gold rush, though, it is already minting fortunes for the sellers of the requisite picks and shovels.

该建筑属于Equinix公司，这是一家租赁数据中心空间的公司。里面的设备属于从企业巨头到初创公司，这些公司越来越多地使用它来运行他们的人工智能系统。“生成式”系统(如热门的虚拟对话助手ChatGPT)令人震惊的复杂程度，刺激了这波AI淘金热，这有望为那些利用这项技术潜力的人带来丰厚的利润。然而，就像任何淘金热的早期一样，它已经为那些必备的镐头和铲子的卖家创造了财富。

On May 24th Nvidia, which designs the semiconductors of choice for many AI servers, beat analysts’ revenue and profit forecasts for the three months to April. It expects sales of $11bn in its current quarter, half as much again as what Wall Street was predicting. As its share price leapt by 30% the next day, the company’s market value flirted with $1trn. Nvidia’s chief executive, Jensen Huang, declared on May 29th that the world is at “the tipping point of a new computing era”.

5月24日，为许多AI服务器设计半导体的英伟达在截至4月的三个月里收入和利润超过了分析师的预测。该公司预计，当前季度的销售额将达到110亿美元，是华尔街预测的一半。随着其股价在第二天暴涨30%，该公司的市值接近1万亿美元。英伟达首席执行官黄仁勋在5月29日宣布，世界正处于“新计算时代的转折点”。

Other chip firms, from fellow designers like AMD to manufacturers such as TSMC of Taiwan, have been swept up in the AI excitement. So have providers of other computing infrastructure—which includes everything from those colourful cables, noisy air-conditioning units and data-centre floor space to the software that helps run the AI models and marshal the data. An equally weighted index of 30-odd such companies has risen by 40% since ChatGPT’s launch in November, compared with 13% for the tech-heavy NASDAQ index (see chart). “A new tech stack is emerging,” sums up Daniel Jeffries of the AI Infrastructure Alliance, a lobby group.

其他芯片公司，从AMD这样的设计同行到台积电这样的制造商，都被卷入了人工智能的热潮中。其他计算基础设施的供应商也是如此，这些基础设施包括从彩色电缆、嘈杂的空调设备、数据中心的占地面积到帮助运行人工智能模型和整理数据的软件。自ChatGPT于11月推出以来，由30多家这样的公司组成的等权指数已经上涨了40%，而以科技为主的纳斯达克指数仅上涨了13%(见图表)。游说团体人工智能基础设施联盟的丹尼尔•杰弗里斯总结道:“一种新的技术堆栈正在出现。”

On the face of it, the AI gubbins seems far less exciting than the clever “large language models” behind ChatGPT and its fast-expanding array of rivals. But as the model-builders and makers of applications that piggyback on those models vie for a slice of the future AI pie, they all need computing power in the here and now—and lots of it.

从表面上看，AI设备似乎远不如ChatGPT及其快速扩张的竞争对手背后聪明的“大语言模型”令人兴奋。但是，随着基于这些模型的模型构建者和应用程序制造商争夺未来的人工智能蛋糕，他们现在都需要算力，而且需要大量的算力。

The latest AI systems, including the generative sort, are much more computing-intensive than older ones, let alone non-AI applications. Amin Vahdat, head of AI infrastructure at Google Cloud Platform, the internet giant’s cloud-computing arm, observes that model sizes have grown ten-fold each year for the past six years. GPT-4, the latest version of the one which powers ChatGPT, analyses data using perhaps 1trn parameters, more than five times as many as its predecessor. As the models grow in complexity, the computational needs for training them increase correspondingly.

最新的AI系统，包括生成式系统，比旧的计算密集得多，更不用说非人工智能应用了。互联网巨头谷歌云平台的人工智能基础设施主管阿明·瓦赫达特观察到，在过去的六年里，模型的大小每年都增长十倍。为ChatGPT提供动力的最新版本GPT-4使用了大约1万亿参数分析数据，这一数字是其前身（1750亿）的五倍多。随着模型复杂度的增加，训练模型所需的计算量也相应增加。

Once trained, AIs require less number-crunching capacity to be used in a process called inference. But given the range of applications on offer, inference will, cumulatively, also demand plenty of processing oomph. Microsoft has more than 2,500 customers for a service that uses technology from OpenAI, ChatGPT’s creator, of which the software giant owns nearly half. That is up ten-fold since the previous quarter. Google’s parent company, Alphabet, has six products with 2bn or more users globally—and plans to turbocharge them with generative AI.

【1】oomph 力量

一旦训练好，人工智能在推理的过程中所需的数字处理能力就会减少。但考虑到所提供的应用范围，推理也将逐渐需要大量的处理能力。微软有2500多名客户使用了ChatGPT的创造者OpenAI的技术提供的服务，这家软件巨头拥有OpenAI近一半的股份。这比上一季度增长了10倍。谷歌的母公司Alphabet拥有6款产品，在全球拥有20亿或更多的用户，并计划用生成式人工智能为这些产品赋能。

The most obvious winners from surging demand for computing power are the chipmakers. The products of companies like Nvidia and AMD, which design chips and have them made at foundries like TSMC, are in hot demand, notably from the big providers of cloud computing that powers most AI applications. AI is thus a boon to the chip designers, since it benefits from more powerful chips (which tend to generate higher margins), and more of them. UBS, a bank, reckons that in the next one or two years AI will increase demand for specialist chips known as graphics-processing units (GPUs) by $10bn-15bn.

对算力的需求激增，最明显的赢家是芯片制造商。英伟达和AMD等公司设计芯片，并在台积电等代工厂制造芯片，它们的产品需求旺盛，尤其是来自大型云计算提供商的需求。大多数人工智能应用都是云计算提供的。因此，人工智能对芯片设计师来说是一个福音，因为它受益于更强大的芯片(往往会产生更高的利润率)和更多的芯片。瑞银估计，在未来一两年里，人工智能将使对图形处理单元(GPU)等专业芯片的需求增加100亿至150亿美元。

As a result, Nvidia’s annual data-centre revenue, which accounts for 56% of its sales, could double. AMD is bringing out a new GPU later this year. Although it is a much smaller player in the GPU-design game than Nvidia, the scale of the AI boom means that the firm is poised to benefit “even if it just gets the dregs” of the market, says Stacy Rasgon of Bernstein, a broker. Chip-design startups focused on AI, such as Cerebras and Graphcore, are trying to make a name for themselves. PitchBook, a data provider, counts about 300 such firms.

因此，英伟达占其销售额56%的数据中心年度收入可能翻一番。AMD将在今年晚些时候推出一款新的GPU。伯恩斯坦的经纪人斯泰西•拉斯贡表示，尽管AMD在GPU设计领域的规模远远小于英伟达，但人工智能热潮的规模意味着，该公司“即使只分得市场的渣渣”，也能从中受益。专注于人工智能的芯片设计初创公司，如Cerebras和Graphcore，正试图为自己赢得声誉。数据提供商PitchBook统计了大约300家这样的公司。

Naturally, some of the windfall will also accrue to the manufacturers. In April TSMC’s boss, C.C. Wei, talked cautiously of “incremental upside in AI-related demand”. Investors have been rather more enthusiastic. The company’s share price rose by 10% after Nvidia’s latest earnings, adding around $20bn to its market capitalisation. Less obvious beneficiaries also include companies that allow more chips to be packaged into a single processing unit. Besi, a Dutch firm, makes the tools that help bond chips together. According to Pierre Ferragu of New Street Research, another firm of analysts, the Dutch company controls three-quarters of the market for high-precision bonding. Its share price has jumped by more than half this year.

当然，部分意外之财也会落入制造商的腰包。4月，台积电的老板魏哲家谨慎地谈到了“人工智能相关需求的增长”。投资者的热情要高得多。在英伟达公布最新收益后，该公司股价上涨了10%，市值增加了约200亿美元。不太明显的受益者还包括那些允许在单个处理单元中封装更多芯片的公司。Besi是一家荷兰公司，他们生产帮助将芯片粘合在一起的工具。据另一家分析公司New Street Research的皮埃尔•费拉居称，这家荷兰公司控制着高精度粘合市场的四分之三。它的股价今年已经上涨了一半以上。

UBS estimates that gpus make up about half the cost of specialised AI servers, compared with a tenth for standard servers. But they are not the only necessary gear. To work as a single computer, a data centre’s GPUs also need to talk to each other.

瑞银估计，GPU约占专用人工智能服务器成本的一半，而标准服务器的成本仅为十分之一。但它们并不是唯一的必备组件。为了像一台计算机一样工作，数据中心的GPU也需要彼此通信。

That, in turn, requires increasingly advanced networking equipment, such as switches, routers and specialist chips. The market for such kit is expected to grow by 40% annually in the next few years, to nearly $9bn by 2027, according to 650 Group, a research firm. Nvidia, which also licenses such kit, accounts for 78% of global sales. But competitors like Arista Networks, a Californian firm, are getting a look-in from investors, too: its share price is up by nearly 70% in the past year. Broadcom, which sells specialist chips that help networks operate, said that its annual sales of such semiconductors would quadruple in 2023, to $800m.

【1】get a look in 得到机会

这反过来又需要越来越先进的网络设备，如交换机、路由器和专业芯片。据研究公司650 Group称，未来几年，此类配套设备的市场预计将以每年40%的速度增长，到2027年将达到近90亿美元。英伟达也获得了此类设备的授权，其占全球销售额的78%。但是像Arista Networks(一家加州公司)这样的竞争对手也受到了投资者的关注: 它的股价在过去一年中上涨了近70%。销售帮助网络通信的专业芯片的博通表示，到2023年，其此类半导体的年销售额将翻两番，达到8亿美元。

The AI boom is also good news for companies that assemble the servers that go into data centres, notes Peter Rutten of IDC, another research firm. Dell’Oro Group, one more firm of analysts, predicts that data centres across the world will increase the share of servers dedicated to AI from less than 10% today to about 20% within five years, and that kit’s share of data centres’ capital spending on servers will rise from about 20% today to 45%.

另一家研究公司IDC的彼得•鲁滕指出，人工智能的繁荣对那些为数据中心组装服务器的公司来说也是个好消息。另一家分析公司Dell ' oro Group预测，全球数据中心将在五年内将专用于人工智能的服务器份额从目前的不到10%增加到20%左右，而配套设备在数据中心服务器资本支出中的份额将从目前的约20%上升到45%。

This will benefit server manufacturers like Wistron and Inventec, both from Taiwan, which produce custom-built servers chiefly for giant cloud providers such as Amazon Web Services (AWS) and Microsoft’s Azure. Smaller manufacturers should do well, too. The bosses of Wiwynn, another Taiwanese server-maker, recently said that AI-related projects account for more than half of their current order book. Super Micro, an American firm, said that in the three months to April AI products accounted for 29% of its sales, up from an average of 20% in the previous 12 months.

这将使纬创和英业达等服务器制造商受益，这两家公司都来自台湾，主要为亚马逊网络服务和微软Azure等大型云提供商生产定制服务器。规模较小的制造商也会有不错的表现。另一家台湾服务器制造商纬颖的老板最近表示，与人工智能相关的项目占他们当前订单的一半以上。美国公司超微表示，在截至今年4月的三个月里，人工智能产品占其销售额的29%，高于此前12个月的平均20%。

All this AI hardware requires specialist software to operate. Some of these programs come from the hardware firms; Nvidia’s software platform, called CUDA, allows customers to make the most of its GPUs, for example. Other firms create applications that let AI firms manage data (Datagen, Pinecone, Scale AI) or host large language models (HuggingFace, Replicate). PitchBook counts around 80 such startups. More than 20 have raised new capital so far this year; Pinecone counts Andreessen Horowitz and Tiger Global, two giants of venture capital, as investors.

所有这些人工智能硬件都需要专业软件来操作。其中一些程序来自硬件公司；例如，英伟达的软件平台CUDA允许客户充分利用其GPU。其他公司创建应用程序，让人工智能公司管理数据(Datagen, Pinecone, Scale AI)或托管大型语言模型(HuggingFace, Replicate)。PitchBook统计了大约80家这样的初创公司。今年到目前为止，已有20多家银行筹集了新资金；Pinecone的投资者包括Andreessen Horowitz和Tiger Global这两家风险投资巨头。

As with the hardware, the main customers for a lot of this software are the cloud giants. Together Amazon, Alphabet and Microsoft plan to undertake capital spending of around $120bn this year, up from $78bn in 2022. Much of that will go to expanding their cloud capacity. Even so, demand for AI computing is so high that even they are struggling to keep up.

与硬件一样，许多软件的主要客户是云巨头。亚马逊、Alphabet和微软计划今年的资本支出总额约为1200亿美元，高于2022年的780亿美元。其中大部分将用于扩大他们的云计算能力。即便如此，对人工智能计算的需求很高，即使是他们也很难跟上。

That has created an opening for challengers. In the past few years IBM, Nvidia and Equinix have started to offer access to GPUs “as a service”. AI-focused cloud startups are proliferating, too. In March one of them, Lambda, raised $44m from investors such as Gradient Ventures, one of Google’s venture arms, and Greg Brockman, co-founder of OpenAI. The deal valued the firm at around $200m. A similar outfit, CoreWeave, raised $221m in April, including from Nvidia, at a valuation of $2bn. Brannin McBee, CoreWeave’s co-founder, argues that a focus on customer service and infrastructure designed around AI help it compete with the cloud giants.

这为挑战者创造了机会。在过去的几年里，IBM、英伟达和Equinix已经开始提供GPU“即服务”。专注于人工智能的云计算初创公司也在激增。今年3月，其中一家公司Lambda从谷歌旗下风险投资公司Gradient Ventures和OpenAI联合创始人格雷格•布罗克曼等投资者那里筹集了4400万美元。这笔交易对该公司的估值约为2亿美元。今年4月，一家类似的公司CoreWeave融资2.21亿美元，估值为20亿美元，其中包括英伟达。CoreWeave的联合创始人布兰宁•麦克比认为，专注于客户服务和围绕人工智能设计的基础设施，有助于该公司与云计算巨头竞争。

The last group of AI-infrastructure winners are closest to providing actual shovels: the data-centre landlords. As demand for cloud computing surges, their properties are filling up. In the second half of 2022 data-centre vacancy rates stood at 3%, a record low. Specialists such as Equinix or its rival, Digital Realty, increasingly compete with large asset managers, who are keen to add data centres to their property portfolios. In 2021 Blackstone, a private-markets giant, paid $10bn for QTS Realty Trust, one of America’s biggest data-centre operators. In April Brookfield, Blackstone’s Canadian rival which has been investing heavily in data centres, bought Data4, a French data-centre firm.

最后一批人工智能基础设施的赢家最接近于提供真正的铲子: 数据中心的房东。随着对云计算需求的激增，他们的房产也被装满了。2022年下半年，数据中心空置率为3%，创历史新低。Equinix或其竞争对手Digital Realty等专业公司与大型资产管理公司的竞争日益激烈，后者热衷于将数据中心纳入其房地产投资组合。2021年，私人市场巨头黑石集团斥资100亿美元收购了美国最大的数据中心运营商之一QTS Realty Trust。今年4月，黑石的加拿大竞争对手Brookfield收购了法国数据中心公司Data4, Brookfield一直在大力投资数据中心。

Continued growth of the AI-infrastructure stack could yet run up against constraints . One is energy . A big investor in data centres notes that access to electricity, of which data centres are prodigious users, is expected to slow development of new data centres in hubs like northern Virginia and Silicon Valley. Another potential block is a shift away from vast AI models and cloud-based inference to smaller systems that require less computing power to train and can run inference on a smartphone , as is the case for Google’s recently unveiled scaled-down version of its PaLM model.

人工智能基础设施堆栈的持续增长可能会遇到限制。一个是能源。一位数据中心的大投资者指出，数据中心是电力的巨大用户，电力的使用预计会减缓像北弗吉尼亚和硅谷这样的枢纽地区新数据中心的发展。另一个潜在的障碍是从庞大的人工智能模型和基于云的推理转向更小的系统，这些系统需要更少的计算能力来训练，并可以在智能手机上运行推理，就像谷歌最近发布的缩小版PaLM模型一样。

The biggest question-mark hangs over the permanence of the AI boom itself. Despite the popularity of ChatGPT and its ilk, profitable use cases for the technology remain unclear. In Silicon Valley, hype can turn to disappointment on a dime. Nvidia’s market value surged in 2021, as its GPUs turned out to be perfect for mining bitcoin and other cryptocurrencies, then collapsed as the crypto boom turned to bust.

最大的问题在于人工智能繁荣本身的持久性。尽管ChatGPT及其同类产品很受欢迎，但该技术的盈利用例仍不清楚。在硅谷，炒作可能会瞬间变成失望。英伟达的市值在2021年飙升，因为其GPU被证明非常适合挖掘比特币和其他加密货币，然后随着加密货币热潮转向萧条而崩溃。

And if the technology does live up to its transformative billing, regulators could clamp down. Policymakers around the world, worried about generative AI’s potential to eliminate jobs or spread misinformation, are already mulling guardrails. Indeed, on May 11th lawmakers in the EU proposed a set of rules that would restrict chatbots.

如果这项技术真的达到了它的变革性收费标准，监管机构可能会取缔它。世界各地的政策制定者担心生成式人工智能可能会消除就业机会或传播错误信息，他们已经在考虑防范措施。事实上，5月11日，欧盟的立法者提出了一套限制聊天机器人的规则。

All these limiting factors may slow AI’s deployment, and in doing so dampen the prospects for AI-infrastructure firms. But probably only a bit. Even if generative AI does not turn out to be as transformative as its boosters claim, it will almost certainly be more useful than crypto. And there are plenty of other, non-generative AIs that also need lots of computing power. Nothing short of a global ban on generative AI, which is not on the horizon, is likely to stop the gold rush. And so long as everybody is rushing, the pedlars of picks and shovels will be cashing in. ■

所有这些限制因素可能会减缓人工智能的部署，从而削弱人工智能基础设施公司的前景。但可能只会削弱一点点。即使生成式人工智能并不像其支持者声称的那样具有变革性，它也几乎肯定会比加密技术更有用。还有很多其他的非生成型人工智能也需要大量的计算能力。只有在全球范围内禁止生成式人工智能(目前还没有出台)，才有可能阻止这场淘金热。只要大家都在“冲动行事“，卖镐头和铁锹的小贩就会大赚一笔。■