Goldman says AI agents could drive a cash flow inflection

Goldman Sachs is pushing the AI debate past model adoption and toward a harder test: when usage turns into cash flow. In a May 20 note, the bank said AI agents could lift token consumption 24 times to 120 quadrillion tokens a month by 2030, while falling compute costs could create a margin inflection for AI players.

That framing changes the conversation inside Goldman’s research and banking teams. The question is no longer whether AI matters, but where the economics land, in semiconductors, cloud infrastructure, software, platforms or applications. It also sharpens the way employees pitch clients, because more AI-linked financing may flow through convertibles, structured products and other hybrid instruments rather than only through straight equity or debt.

Goldman had already drawn that line in a May 11 note. The bank said consumer AI adoption has been “spectacular,” but most users are still on free versions, which makes enterprise adoption critical for the economics to work. It also said most companies outside semiconductors have yet to make money from AI and called the current pattern of chip-company outperformance versus the broader ecosystem “unprecedented and unsustainable.” For analysts building investment cases, that is a direct warning against confusing engagement with monetization.

The same cash-flow test runs through Goldman’s broader AI work on energy and infrastructure. The bank expects U.S. data-center power demand to rise from 31 gigawatts in 2025 to 66 gigawatts in 2027, with data centers’ share of peak summer demand climbing to 8.5% from 4.1%. Goldman also said only about 50% to 60% of scheduled capacity over the next one to two years is likely to come online on time, which means execution risk will shape winners as much as raw demand.

Goldman has also been backing that thesis with capital and capex math. On May 4, Goldman and Anthropic announced a $1.5 billion venture to speed AI adoption across hundreds of companies, and Marc Nachmann said there is a “big shortage” of people who know how to apply AI tools inside businesses. That shortage suggests the bottleneck is implementation, not model access. Goldman’s May 1 report added that the AI build-out depends on assumptions such as silicon replacement cadence, data-center costs, chip mix and bottlenecks, with small changes moving cumulative spend by hundreds of billions.

The bank’s 2026 capex estimate for AI hyperscalers also rose to $527 billion from $465 billion at the start of the third-quarter earnings season, underscoring how quickly the market has shifted from spending levels to return on that spending. Goldman’s latest AI work is making one point harder to ignore: the winners will be the companies that can turn usage into durable revenue, and eventually into free cash flow.