Key Highlights

• Four-fifths of programmers now use an AI coding tool, and coding accounts for over half the combined annual Recurring Revenue of Anthropic and OpenAI.

• Total annual AI revenue must reach roughly USD 2.5 trillion by the end of the decade to cover the cost of the current Data Center buildout, up from around USD 150 billion today.

• AI legal, finance and Customer Service startups have posted rapid growth, but none has matched coding's pace of adoption.

• Four factors set coding apart: abundant Training data, ease of testing outputs, minimal human interaction, and software engineers' own appetite for experimentation.

Other Applications Growing, But Not Catching Up

Three other applications are frequently touted as AI's next major frontier: legal work, finance and customer service. Growth across all three has been substantial. A trio of AI legal tool developers has roughly doubled their combined annual recurring revenue over the past year to USD 1 billion, with one reporting that the amount of time lawyers spend on its platform is doubling monthly. An AI startup building financial services tools added 100 enterprise customers in a single recent quarter and grew its annual recurring revenue by 50%. In customer service, one AI company reached USD 200 million in annual recurring revenue in June, double its level from the previous November, and venture investors have poured more Capital into this category than any other AI application this year.

Despite this growth, none of these applications has matched coding's underlying adoption dynamics. Four structural factors set coding apart. Abundant open-source code published online has fed directly into training data, accounting for close to a fifth of all training material in some models, a Volume of public content that simply does not exist for customer complaints or contract negotiations. Code is also relatively easy to test and validate, since software checks itself, catching errors before they cause damage in ways that are much harder to replicate when assessing whether an irate customer was appeased or a financial model correctly captured a company's accounting nuance.

The Human Factor Matters Too

Software also has an inherent advantage in that much of the information needed to complete a task already exists in digital form within a program's code base, accessible through standardized interfaces connecting different pieces of software. In many other professional contexts, the crucial information instead resides in the judgment of a call center worker or the expression on a negotiating counterparty's face. AI companies are actively working to overcome these gaps, training tools on synthetic data generated by AI systems themselves, connecting professionals directly with model developers to transfer specialized expertise, and analyzing thousands of past interactions to infer institutional knowledge.

The final and perhaps hardest factor to replicate is the workforce itself. Software engineers, accustomed to frequent technological change in their field, have proven natural early adopters, with much of AI coding's uptake happening organically as engineers experimented independently. Customer service agents typically lack the same incentive or latitude to experiment on the job, tending instead to wait for tools to be deployed from above. Lawyers and finance professionals must navigate substantial regulatory constraints that programmers do not face to nearly the same degree.

Conclusion

The gap between coding's runaway adoption and the more gradual progress in law, finance and customer service suggests these differences run deeper than simple product immaturity, reflecting structural features of each profession that AI companies cannot easily engineer around. Whether other applications eventually close that gap as models improve and organizations grow more comfortable delegating tasks to AI agents will likely determine whether the industry's enormous infrastructure Investment ultimately pays off on the timeline its backers are counting on.