Just how rich are businesses getting in the AI gold rush?
企业在人工智能热潮中到底赚了多少钱？

Barely a day goes by without excitement about artificial intelligence (AI) sending another company’s market value through the roof. Earlier this month the share price of Dell, a hardware manufacturer, jumped by over 30% in a day because of hopes that the technology will boost sales. Days later Together AI, a cloud-computing startup, raised new funding at a valuation of $1.3bn, up from $500m in November. One of its investors is Nvidia, a maker of AI chips that is itself on an extended bull run. Before the launch of ChatGPT, a “generative” AI that responds to queries in uncannily humanlike ways, in November 2022 its market capitalisation was about $300bn, similar to that of Home Depot, a home-improvement chain. Today it sits at $2.3trn, $300bn or so short of Apple’s.
人工智能（AI）的兴奋情绪每天都在推动另一家公司的市值飙升。本月早些时候，硬件制造商戴尔的股价一天内飙升了30%，因为人们希望这项技术能够提升销售额。几天后，云计算初创公司Together AI以13亿美元的估值融资，比去年11月的5亿美元估值高出了一倍。其中一家投资者是Nvidia，一家制造AI芯片的公司，其市值也一直在持续上涨。ChatGPT是一种“生成式”AI，于2022年11月发布，以令人难以置信的人类方式回答问题。在那之前，它的市值约为3000亿美元，与家居装修连锁店Home Depot相似。如今，它的市值达到了2.3万亿美元，仅比苹果公司少了约3000亿美元。

The relentless stream of AI headlines makes it hard to get a sense of which businesses are real winners in the AI boom—and which will win in the longer run. To help answer this question The Economist has looked where value has accrued so far and how this tallies with the expected sales of products and services in the AI “stack”, as technologists call the various layers of hardware and software on which AI relies to work its magic. On March 18th many companies up and down the stack will descend on San Jose for a four-day jamboree hosted by Nvidia. With talks on everything from robotics to drug discovery, the shindig will show off the latest AI innovations. It will also highlight furious competition between firms within layers of the stack and, increasingly, between them.
人工智能的不断涌现让人很难判断哪些企业是人工智能繁荣中的真正赢家，以及哪些将在更长远的未来获胜。为了帮助回答这个问题，《经济学人》杂志已经研究了迄今为止价值积累的地方，以及这与人工智能“堆栈”中产品和服务的预期销售情况是否相符。在3月18日，许多涉及堆栈的公司将齐聚圣何塞，参加由英伟达主办的为期四天的盛会。从机器人技术到药物发现，这次盛会将展示最新的人工智能创新。它还将突显堆栈内部和之间公司之间的激烈竞争。

Our analysis examined four of these layers and the companies that inhabit them: AI-powered applications sold to businesses outside the stack; the AI models themselves, such as GPT-4, the brain behind ChatGPT, and repositories of them (for example, Hugging Face); the cloud-computing platforms which host many of these models and some of the applications (Amazon Web Services, Google Cloud Platform, Microsoft Azure); and the hardware, such as semiconductors (made by firms such as AMD, Intel and Nvidia), servers (Dell) and networking gear (Arista), responsible for the clouds’ computing oomph (see chart 1).
我们的分析考察了其中四个层面以及驻留在其中的公司：面向企业的人工智能应用程序（不在堆栈内部销售）；人工智能模型本身，比如Chatgpt背后的GPT-4，以及它们的存储库（例如Hugging Face）；托管许多这些模型和一些应用程序的云计算平台（亚马逊云服务、谷歌云平台、微软Azure）；以及半导体（由AMD、英特尔和Nvidia等公司制造）、服务器（戴尔）和网络设备（Arista）等硬件，负责云计算的计算能力（见图表1）。

Technological breakthroughs tend to elevate new tech giants. The PC boom in the 1980s and 1990s propelled Microsoft, which made the Windows operating system, and Intel, which manufactured the chips needed to run it, to the top of the corporate pecking order. By the 2000s “Wintel” was capturing four-fifths of the operating profits from the PC industry, according to Jefferies, an investment bank. The smartphone era did the same to Apple. Only a few years after it launched the iPhone in 2007, it was capturing more than half of handset-makers’ global operating profits.
技术突破往往会推动新科技巨头的崛起。上世纪80年代和90年代的个人电脑热潮推动了微软和英特尔成为企业界的佼佼者，微软生产Windows操作系统，而英特尔则生产运行该系统所需的芯片。根据投资银行杰富瑞的数据，到了21世纪初，“Wintel”占据了个人电脑行业四分之四的营业利润。智能手机时代也同样对苹果产生了影响。仅仅在2007年推出iPhone几年后，苹果就占据了全球手机制造商超过一半的营业利润。

The world is still in the early days of the generative-AI epoch. Even so, it has already been immensely lucrative. All told, the 100 or so companies that we examined have together created $8trn in value for their owners since its start—which, for the purposes of this article, we define as October 2022, just before the launch of ChatGPT (see chart 2). Not all of these gains are the result of the AI frenzy—stockmarkets have been on a broader tear of late—but many are.
世界仍处于生成式人工智能时代的早期阶段。尽管如此，它已经带来了巨大的利润。我们调查的大约100家公司总共为它们的所有者创造了8万亿美元的价值，自从它开始——对于本文的目的，我们定义为2022年10月，就在Chatgpt推出之前（见图表2）。并非所有这些收益都是人工智能狂热的结果——股市最近一直处于更广泛的繁荣中，但其中许多是。

At every layer of the stack, value is becoming more concentrated in a handful of leading firms. In hardware, model-making and applications, the biggest three companies have increased their share of overall value created by a median of 14 percentage points in the past year and a half. In the cloud layer Microsoft, which has a partnership with ChatGPT’s maker, OpenAI, has pulled ahead of Amazon and Alphabet (Google’s parent company). Its market capitalisation now accounts for 46% of the cloud trio’s total, up from 41% before the release of ChatGPT.
在堆栈的每一层，价值都在一小部分领先企业中更加集中。在硬件、模型制作和应用方面，过去一年半时间里，排名前三的公司在整体价值中所占份额平均增加了14个百分点。在云层中，与Chatgpt的制造商Openai合作的微软已经超过了亚马逊和Alphabet（谷歌的母公司）。其市值现在占云三巨头总市值的46%，高于Chatgpt发布前的41%。

Skimming the cream 撇取奶油

The spread of value is uneven between layers, too. In absolute terms the most riches have accrued to the hardware-makers. This bucket includes chip firms (such as Nvidia), companies that build servers (Dell) and those that make networking equipment (Arista). In October 2022 the 27 public hardware companies in our sample were worth around $1.5trn. Today that figure is $5trn. This is what you would expect in a technology boom: the underlying physical infrastructure needs to be built first in order for software to be offered. In the late 1990s, as the internet boom was getting going, providers of things like modems and other telecoms gubbins, such as Cisco and WorldCom, were the early winners.
价值的传播在各层之间也是不均匀的。就绝对值而言，最富有的是硬件制造商。这个范畴包括芯片公司（如英伟达）、建造服务器的公司（戴尔）以及制造网络设备的公司（阿里斯塔）。在我们的样本中，2022年10月，27家公开硬件公司的市值约为1.5万亿美元。而今天这个数字是5万亿美元。这正是你在技术繁荣中所期望的：底层的物理基础设施需要首先建立，以便提供软件。在上世纪90年代末，随着互联网繁荣的开始，像调制解调器和其他电信设备供应商，比如思科和世界通信，成为了早期的赢家。

So far the host of the San Jose gabfest is by far the biggest victor. Nvidia accounts for some 57% of the increase in the market capitalisation of our hardware firms. The company makes more than 80% of all AI chips, according to IDC, a research firm. It also enjoys a near-monopoly in the networking equipment used to yoke the chips together inside the AI servers in data centres. Revenues from Nvidia’s data-centre business more than tripled in the 12 months to the end of January, compared with the year before. Its gross margins grew from 59% to 74%.
到目前为止，圣何塞峰会的主办方迄今为止是最大的赢家。英伟达占我们硬件公司市值增长的约57%。根据研究公司IDC的数据，该公司生产了超过80%的人工智能芯片。此外，它在数据中心内用于连接芯片的网络设备上几乎垄断了市场。与去年同期相比，英伟达数据中心业务的收入在截至1月底的12个月内增长了两倍多。其毛利率从59%增长至74%。

Nvidia’s chipmaking rivals want a piece of these riches. Established ones, such as AMD and Intel, are launching rival products. So are startups like Groq, which makes super-fast AI chips, and Cerebras, which makes super-sized ones. Nvidia’s biggest customers, the three cloud giants, are all designing their own chips, too—as a way both to reduce reliance on one provider and to steal some of Nvidia’s juicy margins for themselves. Lisa Su, chief executive of AMD, has forecast that revenues from the sale of AI chips could balloon to $400bn by 2027, from $45bn in 2023. That would be far too much for Nvidia alone to digest.
Nvidia的芯片制造竞争对手想要分一杯羹。像amd和Intel这样的老牌公司正在推出竞争产品。像Groq这样制造超快AI芯片的初创公司，以及制造超大规模芯片的Cerebras也在推出竞争产品。Nvidia最大的客户，三大云计算巨头，也在设计自己的芯片，以减少对单一供应商的依赖，并且窃取Nvidia丰厚的利润。amd首席执行官苏姿丰预测，到2027年，来自AI芯片销售的收入可能会从2023年的450亿美元激增至4,000亿美元。这对于Nvidia来说是难以独吞的。

As AI applications become more widespread, a growing share of that demand will also shift from chips required for training models, which consists in analysing mountains of data in order to teach algorithms to predict the next word or pixel in a sequence, to those needed actually to use them to respond to queries, (“inference”, in tech-speak). In the past year about two-fifths of Nvidia’s AI revenues came from customers using its chips for inference. Experts expect some inference to start moving from specialist graphics-processing units (GPUs), which are Nvidia’s forte, to general-purpose central processing units (CPUs) like those used in laptops and smartphones, which are dominated by AMD and Intel. Before long even some training may be done on CPUs rather than GPUs.
随着人工智能应用的普及，对用于训练模型的芯片的需求也将逐渐增加，这些芯片用于分析大量数据，以教授算法预测序列中的下一个单词或像素。过去一年，约有五分之二的英伟达人工智能收入来自于客户使用其芯片进行推理。专家预计，一些推理工作将从专用的图形处理单元（GPU，英伟达的专长）转移到笔记本电脑和智能手机中使用的通用中央处理单元（CPU，由AMD和英特尔主导）。不久之后，甚至一些训练工作也可能会从GPU转移到CPU上进行。

Still, Nvidia’s grip on the hardware market seems secure for the next few years. Startups with no track record will struggle to convince big clients to reconfigure corporate hardware systems for their novel technology. The cloud giants’ deployment of their own chips is still limited. And Nvidia has CUDA, a software platform which allows customers to tailor chips to their needs. It is popular with programmers and makes it hard for customers to switch to rival semiconductors, which CUDA does not support.
尽管如此，Nvidia对硬件市场的控制似乎在未来几年内是稳固的。没有任何记录的初创公司将很难说服大客户重新配置公司硬件系统以适应其新技术。云巨头部署自己的芯片仍然有限。而Nvidia拥有cuda，这是一个软件平台，允许客户根据自己的需求定制芯片。它受程序员欢迎，并且使客户很难转向不支持cuda的竞争半导体产品。

Whereas hardware wins the value-accrual race hands down in absolute terms, it is the independent model-makers that have enjoyed the biggest proportional gains. The collective value of 11 such firms we have looked at has jumped from $29bn to about $138bn in the past 16 months. OpenAI is thought to be worth some $100bn, up from $20bn in October 2022. Anthropic’s valuation surged from $3.4bn in April 2022 to $18bn. Mistral, a French startup founded less than a year ago, is now worth around $2bn.
硬件在绝对值方面赢得了增值竞赛，但独立的模型制造商获得了最大比例的收益。我们研究的11家这样的公司的总价值在过去16个月内从290亿美元增长到约1380亿美元。据说Openai的价值从2022年10月的200亿美元增加到了1000亿美元。Anthropic的估值从2022年4月的34亿美元激增至180亿美元。法国初创公司Mistral成立不到一年，现在价值约20亿美元。

Some of that value is tied up in hardware. The startups buy piles of chips, mostly from Nvidia, in order to train their models. Imbue, which like OpenAI and Anthropic is based in San Francisco, has 10,000 such chips. Cohere, a Canadian rival, has 16,000. These semiconductors can sell for tens of thousands of dollars apiece. As the models get ever more sophisticated, they need ever more. GPT-4 reportedly cost about $100m to train. Some suspect that training its successor could cost OpenAI ten times as much.
部分价值与硬件相关。初创公司购买大量芯片，主要来自英伟达，用于训练他们的模型。Imbue总部位于旧金山，像Openai和Anthropic一样，拥有1万个这样的芯片。加拿大的竞争对手Cohere拥有1.6万个。这些半导体每个售价可能达到数万美元。随着模型变得越来越复杂，它们需要越来越多的芯片。据报道，gpt-4的训练成本约1亿美元。有人怀疑，训练其继任者可能会花费Openai十倍的成本。

Yet the model-makers’ true worth lies in their intellectual property, and the profits that it may generate. The true extent of those profits will depend on just how fierce competition among model providers will get—and how long it will last. Right now the rivalry is white hot, which may explain why the layer has not gained as much value in absolute terms.
然而，模型制造商的真正价值在于他们的知识产权，以及可能带来的利润。这些利润的真正程度将取决于模型提供商之间竞争的激烈程度以及持续时间。目前竞争异常激烈，这可能解释了为什么该层在绝对价值上没有获得太多增值。

Although OpenAI seized an early lead, challengers have been catching up fast. They have been able to tap the same data as the maker of ChatGPT (which is to say text and images on the internet) and, also like it, free of charge. Anthropic’s Claude 3 is snapping at GPT-4’s heels. Four months after the release of GPT-4, Meta, Facebook’s parent company, released Llama 2, a powerful rival that, in contrast to OpenAI’s and Anthropic’s proprietary models, is open and can be tinkered with at will by others. In February Mistral, which has fewer than 40 employees, wowed the industry by releasing an open model whose performance almost rivals that of GPT-4, despite requiring much less computational power to train and run.
虽然Openai占据了早期领先地位，但挑战者们迅速在追赶。他们能够利用与Chatgpt制造商相同的数据（也就是互联网上的文本和图像），而且像它一样，是免费的。Anthropic的Claude 3正在追赶gpt-4。在gpt-4发布四个月后，Facebook的母公司Meta发布了强大的竞争对手Llama 2，与Openai和Anthropic的专有模型相比，它是开放的，可以被其他人随意调整。在二月，拥有不到40名员工的Mistral通过发布一个开放模型，其性能几乎与gpt-4相媲美，尽管训练和运行所需的计算能力要少得多。

Even smaller models increasingly offer good performance at a low price, points out Stephanie Zhan of Sequoia, a venture-capital firm. Some are designed for specific tasks. A startup called Nixtla developed TimeGPT, a model for financial forecasting. Another, Hippocratic AI, has trained its model on data from exams to enter medical school, to give accurate medical advice.
甚至更小的模型也越来越能以低价提供良好的性能，Sequoia风险投资公司的Stephanie Zhan指出。有些是为特定任务而设计的。一家名为Nixtla的初创公司开发了Timegpt，这是一个用于财务预测的模型。另一家名为Hippocratic ai的初创公司已经用医学院入学考试的数据来训练其模型，以提供准确的医疗建议。

The abundance of models has also enabled the growth of the application layer. The value of the 19 publicly traded software companies in our application group has jumped by $1.1trn, or 35%, since October 2022. This includes big software providers that are adding generative AI to their services. Zoom uses the technology to let users summarise video calls. ServiceNow, which provides tech, human-resources and other support to companies, has introduced chatbots to help resolve customers’ IT queries. Adobe, maker of Photoshop, has an app called Firefly, which uses AI to edit pictures.
模型的丰富多样也促进了应用层的增长。自2022年10月以来，我们应用组中的19家上市软件公司的价值已经增加了1.1万亿美元，增长了35%。这包括一些大型软件提供商正在将生成式人工智能应用到他们的服务中。Zoom利用这项技术让用户总结视频通话。ServiceNow为公司提供技术、人力资源和其他支持，已经推出了聊天机器人来帮助解决客户的IT问题。Adobe，Photoshop的制造商，推出了一个名为Firefly的应用，利用人工智能来编辑图片。

Newcomers are adding more variety. “There’s An AI For That”, a website, counts over 12,000 applications, up from fewer than 1,000 in 2022. DeepScribe helps transcribe doctors’ notes. Harvey AI assists lawyers. More idiosyncratically, 32 chatbots promise “sarcastic conversation” and 20 generate tattoo designs. Fierce competition and low barriers to entry mean that some, if not many, applications could struggle to capture value.
新手正在增加更多的多样性。“有一个ai适用于那个”，一个网站，统计超过12,000个应用程序，比2022年的不到1,000个多。DeepScribe帮助转录医生的笔记。Harvey ai协助律师。更具个性化的是，32个聊天机器人承诺“讽刺对话”，20个生成纹身设计。激烈的竞争和低进入壁垒意味着一些，如果不是很多，应用程序可能会难以捕捉价值。

Then there is the cloud layer. The combined market capitialsation of Alphabet, Amazon and Microsoft has jumped by $2.5trn since the start of the AI boom. Counted in dollars that is less than three-quarters of the growth of the hardware layer, and barely a quarter in percentage terms. Yet compared with actual revenues that AI is expected to generate for the big-tech trio in the near term, this value creation far exceeds that in all the other layers. It is 120 times the $20bn in revenue that generative AI is forecast to add to the cloud giants’ sales in 2024. The comparable ratio is about 40 for the hardware firms and around 30 for the model-makers.
然后是云层。自人工智能繁荣开始以来，谷歌母公司Alphabet、亚马逊和微软的市值总额已经增加了2.5万亿美元。以美元计算，这不到硬件层增长的四分之三，而在百分比方面几乎不到四分之一。然而，与大型科技公司在短期内预计从人工智能中获得的实际收入相比，这种价值创造远远超过其他层面。这相当于2024年生成式人工智能预计为云巨头销售额增加的200亿美元的120倍。硬件公司的可比比率约为40，而模型制造商的比率约为30。

This implies that investors believe that the cloud giants will be the biggest winners in the long run. The companies’ ratio of share price to earnings, another gauge of expected future profits, tells a similar story. The big three cloud firms average 29. That is above 50% higher than for the typical non-tech firm in the S&P 500 index of large American companies—and up from 21 in early 2023 (see chart 3).
这意味着投资者相信云巨头将是长期内最大的赢家。公司的股价与收益比，另一个衡量预期未来利润的标准，也讲述了类似的故事。三大云公司的平均值为29。这高于标准普尔500指数中典型的非科技公司的50%以上，而在2023年初为21（见图3）。

Investors’ cloud bullishness can be explained by three factors. First, the tech titans possess all the ingredients to develop world-beating AI systems: troves of data, armies of researchers, huge data centres and plenty of spare cash. Second, the buyers of AI services, such as big corporations, prefer to do business with established commercial parters than with untested upstarts (see chart 4). Third, and most important, big tech has the greatest potential to control every layer of the stack, from chips to applications. Besides designing some of their own chips, Amazon, Google and Microsoft are investing in both models and applications. Of the 11 model-makers in our sample, nine have the support of at least one of the three giants. That includes the Microsoft-backed OpenAI, Anthropic (Google and Amazon) and Mistral (Microsoft again).
投资者对云计算的乐观情绪可以通过三个因素来解释。首先，科技巨头拥有开发世界领先人工智能系统的所有要素：大量数据、大批研究人员、庞大的数据中心和充裕的现金。其次，购买人工智能服务的买家，如大型企业，更倾向于与成熟的商业合作伙伴做生意，而不是与未经测试的新兴公司合作（见图表4）。第三，最重要的是，大型科技公司有最大的潜力控制从芯片到应用的每一层技术堆栈。除了设计一些自己的芯片外，亚马逊、谷歌和微软还在投资模型和应用。在我们的样本中，有11家模型制造商，其中有9家得到了这三家巨头中至少一家的支持。这包括得到微软支持的Openai、得到谷歌和亚马逊支持的Anthropic以及得到微软支持的Mistral。

Have the layer cake and eat it
想要又吃又有蛋糕

The potential profits that come from controlling more of the layers are also leading hitherto layer-specific firms to branch out. OpenAI’s in-house venture-capital arm has invested in 14 companies since its launch in January 2021, including Harvey AI and Ambience Healthcare, another medical startup. Sam Altman, boss of OpenAI, is reportedly seeking investors to bankroll a pharaonic $7trn chipmaking venture.
控制更多层面可能带来的潜在利润也导致迄今专注于某一层面的公司开始拓展。Openai的内部风险投资部门自2021年1月成立以来已投资了14家公司，包括Harvey ai和另一家医疗初创公司Ambience Healthcare。据报道，Openai的老板Sam Altman正在寻求投资者支持一项规模达7万亿美元的芯片制造业务。

Nvidia is becoming more ambitious, too. It has taken stakes in seven of the model-makers, and now offers its own AI models. It has also invested in startups such as Together AI and CoreWeave, which compete with its big cloud customers. At its San Jose event it is expected to unveil a snazzy new GPU and, just maybe, AI tools from other layers of the stack. The AI boom’s biggest single value-creator is in no mood to cede its crown. ■
英伟达也变得更加雄心勃勃。它已经持有七家模型制造商的股份，并现在提供自己的人工智能模型。它还投资了一些初创公司，如Together ai和CoreWeave，它们与其大型云客户竞争。在其圣何塞活动上，预计将推出一款时髦的新GPU，也许还会推出来自堆栈其他层的人工智能工具。人工智能繁荣的最大价值创造者并不打算放弃其王位。■