the-envelope calculation by The Economist finds that covering the cost of the data-centre build-out will require total annual revenue from the use of AI to reach $2.5trn per year by the end of the decade, up from around $150bn now. So far uptake has been strongest by far among software developers. Four- fifths of programmers say they use an AI coding tool, according to Stack Overflow, an online forum. SemiAnalysis, a research firm, reckons that for Anthropic and OpenAI coding now accounts for over half their combined annual recurring revenue—a trendy measure of sales that extrapolates from the most recent month. In June 2025 the combined annual recurring revenue of Cognition, Cursor, Lovable and Replit, four AI coding startups, was roughly $800m. Today it stands at $6bn. Coders’ enthusiasm explains why AI has diffused more quickly in industries where they account for a larger share of employment (see chart). Optimists may see this as evidence that the AI boom is just getting started, with spending on the technology reaching far greater heights in the years ahead as other applications mature. Yet there may also be structural reasons why coding is an outlier.

Three candidates often touted as the next runaway application for AI are lawyering, finance and customer service. Growth in all has been impressive. Harvey, Clio and Legora, three developers of AI legal tools, have roughly doubled their combined annual recurring revenue in the past year to $1bn. Harvey says the amount of time lawyers spend on its platform is doubling each month. Bankers and other finance types are also starting to embrace the technology. Rogo, an AI startup building tools for the financial-services industry, added 100 enterprise customers in the past quarter and increased annual recurring revenue by 50%. AI is also taking off in customer service. Sierra, which builds tools for the purpose, reached an annual recurring revenue of $200m in June, double the figure in November last year. Venture capitalists are bullish. So far this year they have poured $3bn into AI startups in this area, more than for any other category of AI application. (Investment in AI coding startups has stalled, in part because AI labs’ own coding tools have grown so quickly.) These three applications bear some similarities to coding. They are big potential markets, and the work is often formulaic. But four factors set coding apart: the availability of training data; how easy it is to test a model’s output; the low amount of human interaction involved in the work; and software engineers themselves. AI companies are trying to overcome the differences in other markets, but doing so will not be straightforward. Part of why AI has proved so successful as a programming buddy is that a huge amount of open-source code has been published on the internet. It has thus made its way into models, accounting for almost a fifth of all training data in some cases. The same volume of publicly available content simply does not exist for customer complaints or contract negotiations. Code generated by AI is also relatively easy to validate. Software tests are a routine part of engineering: one piece of code checks another. That increases the likelihood that any errors in AI-generated code are caught before they wreak havoc. Moreover, testing helps to improve the performance of models by alerting them when a mistake is made. Assessing whether an irate customer was appeased or a valuation model appropriately reflected the nuance in a company’s accounts is trickier.

Software-writing tools have the added advantage that much of the information required to complete a task already exists in a program’s code- base, or can be reached through the digital pipelines that connect software, known as application programming interfaces. In many other contexts crucial information is not available in digital form, instead residing in the brain of a call-centre worker or the facial expression of a counterparty’s lawyer. AI companies are working on all these challenges. Winston Weinberg, the boss of Harvey, says that advances in AI mean his firm is now able to train its tools in part by relying on “synthetic” data, which are generated by AI systems. Startups such as Mercor are connecting AI companies with professionals in various fields who can transfer their expertise into models. Ping Wu, boss of Cresta, an AI customer-service startup, says that by analysing thousands of call-centre transcripts his firm can infer some of the knowledge lodged in the heads of staff, such as how refunds are handled. Legions of so-called forward-deployed engineers are now being dispatched by AI companies to their customers’ offices to troubleshoot problems and improve their tools. But the last factor behind AI’s success in coding may be harder to replicate: software engineers themselves. Being both nerdy and accustomed to change, given that programming languages are updated every few years, they are natural adopters of new technology. Much of the uptake of AI coding has thus happened in a bottom-up way, with engineers discovering and experimenting with the tools of their own accord, notes Anish Acharya of Andreessen Horowitz, a venture-capital firm. Call-centre agents, by contrast, typically have neither the incentive nor the freedom to experiment on the job, and tend to wait for technologies to be deployed top-down. Lawyers and finance workers, for their part, must tread carefully to avoid falling foul of the many regulations that govern their occupations. Other applications of AI will certainly continue to grow, and may well accelerate as models improve and companies become more comfortable with handing over tasks to AI agents. But it could be a while before the majority of office drones spend their days whispering into microphones. ■

This article was downloaded by zlibrary from https://www.economist.com/business/2026/08/30/will-anybody-use-ai-as-much-as- coders-do

Business · Business | Reaching for the sky

Neoclouds like CoreWeave are getting much bigger—and riskier The mountain of debt funding their expansion is growing rapidly Sep 3rd 2026 New technologies often create good conditions for middlemen. In the 1960s entrepreneurs snapped up mainframes from IBM and leased them to banks and defence companies. In the dotcom era “domainers” bought website addresses and sold them at a higher price. Artificial intelligence is no different. Consider the neoclouds, such as CoreWeave, which buy specialist AI chips and rent them out. They are among the fastest-growing companies in the AI industry. But they are also some of the most vulnerable. Neoclouds have thrived over the past few years, for two reasons. The first is that demand for AI has grown more quickly than America’s “hyperscalers” have been able to expand their computing capacity. Many neoclouds started

off as crypto miners, and subsequently repurposed their data centres. And because they have focused exclusively on AI, rather than a broader range of cloud services, they have been able to concentrate their investments. As a result, their biggest customers are model-makers, such as OpenAI, and the hyperscalers themselves, who often lease extra capacity from the neoclouds. The second reason for their rise is that Nvidia, the leading maker of AI chips, which is eager to reduce its reliance on the hyperscalers’ custom, has provided various forms of financial support to help them buy more of its chips. The combined annual revenue of the five largest listed neoclouds— CoreWeave, Nebius, IREN, Applied Digital and Core Scientific—has risen to $18bn over the past 12 months, up from $3bn in 2024. Although they are still far behind the hyperscalers, they are gaining ground. Dell Oro, a research firm, estimates that in total neoclouds have 1.5-2.5 gigawatts (GW) of data-centre capacity between them, compared with a combined 20-24GW for Alphabet, Amazon, Microsoft and Meta. But they are roughly doubling their capacity each year, which is about triple the rate that the hyperscalers are managing. That growth, however, is being funded by mountains of debt. The five listed neoclouds now have $61bn-worth of borrowing on their balance-sheets,

quadruple the amount just 12 months ago, as well as $18bn-worth of operating leases (see chart). None is generating an operating profit, and all are deeply in the red once interest bills are added. In the most recent quarter Nebius’s interest payments amounted to a fifth of its revenue; CoreWeave’s reached a quarter. The neoclouds are betting that, if they continue their vertiginous growth, profits will eventually arrive. Perhaps. But the competition has shown no sign of easing. Besides the established cloud giants—as well as Meta, which is considering launching its own third-party service—there are now some 200 neoclouds around the world duking it out for contracts. Add to that a growing cohort of mid-sized providers. Oracle, originally a database company, has spent a fortune trying to join the hyperscalers’ club with a cloud business of its own. Its capital expenditure, much of which is financed by borrowing, is expected to hit $77bn this year. In July SoftBank, a Japanese conglomerate up to its eyeballs in debt, unveiled a cloud business called SB Neo. The group also owns a majority stake in SB Energy, an infrastructure business that has teamed up with OpenAI to build a giant data centre in Ohio, and which filed on September 1st for an initial public offering. Fierce rivalry compounds the two big risks facing neoclouds. The first is a credit-market shock that pushes borrowing costs higher, further squeezing margins. Investors already seem jittery: the spread in yields between neoclouds’ bonds and junk bonds as a whole is widening. A second—and even bigger—danger is that supply and demand rebalance. Right now, vast amounts of infrastructure are being built on the assumption that use of the technology will continue to soar. If the pace of adoption slows, however, model-makers and hyperscalers alike may quickly decide they no longer need the neoclouds. Either event would be cataclysmic for the industry. Only those neoclouds that offer something more to customers than sheer capacity would avoid going bust. To that end, some are focusing on positioning themselves as reliable infrastructure providers for Western governments that have grown suspicious of America’s tech giants. Others, including CoreWeave and

Nebius, are trying to offer added software services to customers. To survive, the neoclouds may need to be more than middlemen. ■ This article was downloaded by zlibrary from https://www.economist.com/business/2026/09/03/neoclouds-like-coreweave-are-getting- much-bigger-and-riskier