Goldman Sachs flags rising AI costs as use cases grow more complex

The first wave of AI adoption can hide the real expense. Once models move beyond simple prompts into reasoning, tool use and larger context windows, Commonwealth Bank of Australia chief executive Matt Comyn warned that token costs do not scale in a straight line, a problem that will matter to every bank trying to turn AI into a budgeting item rather than a science project.

That warning lands squarely inside Goldman Sachs, where AI is already being pushed into research, operations, engineering and client work. In October 2025, Goldman executives said the firm was looking at every process and asking how AI could interact with it, while also acknowledging that many workflows remain hard to automate because they are not easily verifiable by a machine. Jung Min has said the challenge is to reprocess workflows so they become verifiable, and Marco Argenti has argued that companies need to model processes in ways AI can check.

The economics are getting less forgiving as the use cases get more complex. Goldman Sachs Research said on May 11 that enterprises need structured data and efficiently orchestrated workflows to control AI costs, and it warned that most companies in the AI ecosystem have yet to make money from the technology. Just nine days later, Goldman projected that agentic AI could drive a 24-fold increase in token consumption by 2030, reaching about 120 quadrillion tokens per month, while chip supply could remain tight for 12 to 18 months as demand catches up.

For Goldman employees, that shifts the internal conversation from experimentation to discipline. A model that looks efficient in a proof of concept can become expensive when it is rolled out across thousands of users or embedded in complex work such as trade accounting and client onboarding. CNBC reported in February that Goldman was working with Anthropic on AI agents for those tasks, expecting efficiency gains and trying to constrain headcount growth rather than make near-term job cuts. That is the operating reality behind the enthusiasm: management wants productivity, but the bill has to be justified in a year when every budget line will be examined closely.

Comyn’s comments also point to a broader constraint that banks cannot ignore. AI buildout is feeding heavy energy and water demand for data centers, while businesses globally are likely to press harder on return on investment through 2026. Goldman has said generative AI adoption reached about 53% within three years of the first widely available product, but adoption alone does not solve the economics. For Goldman, the next test is not whether AI can be deployed. It is whether the firm can keep the costs from outrunning the value.