信仰推动了多次技术革命。没有人有确凿证据表明个人电脑、互联网或社交网络会创造出价值数十亿美元的数字经济，但史蒂夫·乔布斯、杰夫·贝索斯和马克·扎克伯格等远见者的坚持使之成为现实。如今，科技行业面临着日益增长的怀疑，即其能否用人工智能重复这一成功。
Faith has propelled several technology revolutions. No one had hard evidence that personal computers, the Internet or social networks would create digital economies worth billions of dollars, but the perseverance of visionaries like Steve Jobs, Jeff Bezos and Mark Zuckerberg made it happen. Now the tech industry faces mounting skepticism about its ability to repeat the trick with artificial intelligence.
人们对Anthropic PBC、OpenAI以及Alphabet Inc.旗下谷歌未来盈利潜力的疑虑正在增加：当企业迄今为止从使用这些模型中几乎未见生产力提升时，庞大计算机仓库的账单又将由谁来支付？
Doubts about the future earnings potential of Anthropic PBC, OpenAI and Alphabet Inc.’s Google are growing: How will the bills for vast computer warehouses get paid when businesses are so far seeing few productivity gains from using the models?
最近我前往旧金山，探访硅谷的脉搏，发现那里热闹非凡——但对人工智能未来的信念，更多是基于轶事、理论和直觉，而非数据和事实。一位初创公司创始人引用他们“强烈的直觉”，认为人工智能工具将在编程之外得到更广泛的应用。来自可能是历史上最昂贵的基础设施建设背后的人们的这一结论令人不安，如果预计的收入不能尽快到来，这可能会向金融市场发送冲击波。
On a recent trip to San Francisco to take the pulse of Silicon Valley, I found it racing — but the belief in AI’s future was also underpinned more by anecdotes, theories and hunches than data and facts. One startup founder cited their “strong intuition” that AI tools will be used more broadly beyond coding. It’s a disconcerting takeaway from the people behind what may be the most expensive infrastructure buildout in history, which risks sending shockwaves through financial markets if the projected revenue doesn’t arrive quickly enough.
我发现，即使那些预期短期市场修正的人，仍然相信人工智能将产生有意义的长期影响，并且人类会将更多任务和决策让渡给软件。一位连续创始人夸耀说，他们仅用三周时间和精简团队就构建了最新产品，雇佣了一名营销人员来完成之前公司25个人的工作。在自己的工作中发现即将到来的革命的证据，让科技高管们产生近视的视角，并带有傲慢的色彩。
I found even those expecting a near-term market correction still believed in the meaningful, long-term impact of AI, and that humans would cede more tasks and decision-making to software. A serial startup founder boasted that they’d built their latest product in three weeks with a skeleton crew, hiring one marketing person to do the job of 25 in their previous firm. Finding evidence of the coming revolution in their own work gives tech executives a myopic perspective that’s tainted with hubris.
我听到的支撑这场数万亿美元建设的最薄弱理由，来自一些抱怨称其初创公司难以找到提供新型AI服务所需算力的风险投资人。有些创始人甚至急于以股权换取英伟达公司芯片的使用权。当我告诉一位风投，这未必能转化为大型银行或零售商的长期采用时，他反驳说，人们常常忽略“复合式技术增长向量”。这种所谓缺乏想象力的现象，在我在硅谷的讨论中反复出现：世界不理解指数增长的概念，也不理解轮子、电力以及如今AI等发明的二阶效应。但即便如此，支付所有这些数据中心的收入又将从何而来？
Perhaps the weakest justification I heard for the multi-trillion-dollar buildout came from venture capitalists complaining that their startups were struggling to find the computing capacity they needed to provide new AI services. Some founders were desperate enough to offer equity in exchange for access to Nvidia Inc. chips. When I told one VC that this might not translate to long-term adoption by large banks or retailers, they retorted that people often miss “compounding technical growth vectors.”This alleged lack of imagination was a recurring theme in my Silicon Valley discussions; that the world doesn’t grasp the concept of exponential growth or the second-order effects of inventions such as the wheel or electricity, and now AI. But even if that’s true, where will the revenue come from to pay for all those data centers?
答案几乎总是智能体。与人工智能处理为人们预订餐厅或为企业跟踪客户支持等现实工作所需的众多步骤和查询相比，你在ChatGPT上关于烹饪的查询使用的计算资源微不足道。据Tokenomics基金会主任J. R. Storment称，全球代币消耗量——衡量人工智能使用量的指标——在过去四个月里翻了三倍，很大程度上是因为模型陷入循环、生成其他智能体或在夜间工作。
The answer was almost always agents. Your ChatGPT queries about cooking use tiny computing resources compared with the many steps and queries AI will need to handle real-world work like making restaurant reservations for people or tracking customer support for businesses. Global token consumption — a measure of AI usage — tripled over the last four months in large part because of models getting caught in loops, spawning other agents or working overnight, according to J. R. Storment, director of the Tokenomics Foundation, set up this summer to help create industry standards.
该基金会今年夏天成立，旨在帮助制定行业标准。“我现在手机上运行着九个智能体，在管理事情，我还不是重度用户，”他在圣何塞Linux基金会主办的AGNTCon会议侧边告诉我，随后指向与会者。“这里有些人有数百个智能体。”专注于物理空间来执行诸如处理视觉数据等任务需要更多的计算能力。例如，一家大型物流公司使用软件处理来自其仓库的摄像头画面，以提出诸如移动托盘等建议。
“I now have nine agents running on my phone, managing things, and I'm not a powerful user,” he told me on the sidelines of AGNTCon, a San Jose conference hosted by the Linux Foundation, before gesturing at the attendees. “Some people here have hundreds of agents.”Focusing on physical spaces to perform tasks such as processing visual data needs even computing power. One large logistics company, for instance, uses software to process camera feeds coming from its warehouses to make suggestions like moving pallets around.
办公室里的人工审查所有这些建议，但该系统最终将被信任直接指导工人，据旧金山初创公司Claryo的技术提供者Mohamed Amer称。
A human in an office reviews all those recommendations, but the system will eventually be trusted to guide workers directly, according to Mohamed Amer, whose San Francisco startup Claryo provides the technology.
当人工智能模型变得越来越便宜时，这种额外的使用量会产生更多的收入和利润吗？技术人员说是的，并且不断增长的需求将超过价格下降。他们特别指出净收入留存率（NRR），即来自人工智能服务现有客户的销售增长。“如果你今天在我们这里花费一百万美元，12个月后你将花费400万美元，”估值90亿美元的编程公司Replit的创始人兼首席执行官Amjad Masad说。
Will that additional usage generate more revenue and profit when AI models are getting cheaper? Technologists say yes, and that growing demand will outweigh falling prices. They point specifically to net revenue retention (NRR), the sales growth coming from existing customers of AI services. “If you spend a million dollars with us today, in 12 months you'll be spending $4 million,” says Amjad Masad, founder and chief executive officer of Replit, a coding company valued at $9 billion.
根据该公司最近的首次公开募股文件，Anthropic的净收入留存率（NRR）甚至更高；其用户在Claude上的支出是12个月前的九倍。该公司的收入也没有像许多人曾担心的那样高度集中。Anthropic的前五大客户约占其收入的25%，其余大部分分散在数千家付费企业客户之间。
Anthropic’s NRR is even higher; its users are spending nine times more on Claude now than they were 12 months ago, according to the company’s recent initial public offering documentation. And its income isn’t as concentrated as many once feared. The company’s top five customers account for roughly 25% of revenue, with much of the rest spread out across thousands of paying enterprise customers.
当Anthropic上市的话题被提及时，科技企业家们的自信开始动摇。如此多的个人财富都与这笔可能将公司估值达到惊人的1.5万亿美元的股票出售挂钩。有人认为该估值过高。另一些人则采取了更为谨慎的模式，初创公司削减了差旅预算，风险投资公司在投资前要求手头有两年的现金储备，而一年前它们还不需要。行业风向标的首次公开募股成败，关系重大。
When the topic of Anthropic’s listing came up, the self-belief of the tech entrepreneurs started to wobble. So much personal wealth is tied to the share sale that could value the company at an incredible $1.5 trillion. Some mused that valuation was too high. Others had hunkered into a more cautious mode, with startups reducing their travel budgets and VCs demanding two years worth of cash on hand before investing, having needed none a year ago. A lot is riding on the success or failure of the industry bellwether’s IPO.
只有少数人公开表示：Anthropic和人工智能行业可能会在经济蛋糕扩大的同时亏损。科技界最根深蒂固的信念是，人工智能的涟漪效应将在医疗保健和金融等领域创造前所未有的发展，从而提升各地的生产力。但构建软件模型的公司是否会受益，则是另一回事。
What only a few said out loud: Anthropic and the AI industry could lose money even as the economic pie grows. The most deeply held belief in the tech world is that AI’s ripple effects will create hitherto unimaginable developments in areas like healthcare and finance, boosting productivity everywhere. Whether the companies building the software models will benefit is another matter.
无论发生什么，硅谷大多数富有的真正信徒都会没事。如果他们的AI押注出错，他们的钱会保护他们免受潜在市场崩盘的影响；而我们其余的人可能就没那么幸运了。
Whatever happens, most of Silicon Valley’s wealthy true believers will be fine. Their money will shelter them from a potential market meltdown if their AI bet goes wrong; the rest of us may not be so lucky.