Goldman projects 120 quadrillion tokens

Goldman Sachs said in a May 20 analysis that token consumption from AI agents could rise 24-fold between 2026 and 2030, reaching 120 quadrillion tokens per month. The forecast appeared in a Goldman Sachs Insights article titled “AI Agents Forecast to Boost Tech Cash Flow as Usage Soars,” published five days before the figure circulated widely on social media. Jim Schneider, a senior equity analyst covering U.S. semiconductor and IT services at Goldman Sachs Research, was cited in the article as saying consumer and enterprise adoption would drive the increase. (goldmansachs.com) The number drew attention on May 24 in posts on X that framed it as a marker of how much compute, networking and data-center capacity AI agents could require if adoption broadens. Goldman’s article paired the usage forecast with a second claim: that falling computing costs could improve economics for major AI companies even as total usage rises. Schneider said AI players were positioned for a period of “margin inflection” as computing costs fall. ### Where did the 120-quadrillion figure come from? Goldman Sachs published the figure in its own Insights article, not in an anonymous social-media summary. (goldmansachs.com) The article said “token consumption is expected to multiply 24 times, to 120 quadrillion tokens per month, between 2026 and 2030,” and attributed that estimate to Schneider. Goldman did not present the number in the source excerpt as a stand-alone market consensus figure; it was part of the bank’s research view on AI-agent adoption and tech-sector cash flow. May 24 X posts then amplified the figure, including one that linked the number directly to Goldman analysis. (goldmansachs.com) Those posts described the forecast as evidence of a coming jump in infrastructure demand tied to agentic AI. The underlying Goldman article is dated several days earlier than the social-media circulation. ### What exactly is Goldman projecting? The Goldman Sachs projection refers to monthly token consumption by AI agents, not to revenue, users or model count. The article says the increase would occur “between 2026 and 2030,” implying a baseline in 2026 and a projected endpoint in 2030. (goldmansachs.com) Goldman’s wording also says the growth would come from “consumers and enterprises adopting AI agents,” placing both personal-use and business-use software in scope. A separate Goldman Sachs article published in 2025 had already argued that a “generational infrastructure buildout” could hinge on AI agents, showing the bank has been building a broader research theme around agent-driven demand. (goldmansachs.com) Another Goldman article published in 2026 said data-center power consumption could jump 175% by 2030 from 2023 levels in its base case, though that forecast was presented separately from the 120-quadrillion-token estimate. ### Who at Goldman is attached to the forecast? Jim Schneider is the named Goldman Sachs Research analyst attached to the token-consumption estimate. (goldmansachs.com) Goldman’s article identifies him as the senior equity analyst covering U.S. semiconductor and IT services. That matters because the forecast was presented through the lens of infrastructure, component demand and cash-flow implications for technology companies, rather than as a consumer-product forecast. Gabriela Borges, another Goldman Sachs Research analyst, wrote in a separate 2025 Goldman article that AI agents could expand the customer-service software market by 20% to 45% by 2030. (goldmansachs.com) That article addressed software-market size rather than token volumes, but it shows Goldman has published multiple agent-related forecasts across different parts of the tech stack. ### Why did the figure spread beyond Goldman’s site? May 24 social posts used the 120-quadrillion figure as a shorthand for scale. The number is large enough that it became a talking point in discussions about semiconductors, cloud spending and AI operating costs, especially as investors debate whether current capital spending on AI infrastructure can be justified by future usage. (goldmansachs.com) Goldman has separately published work on “the scale of the AI build-out,” framing AI capital expenditure as highly sensitive to assumptions about infrastructure design and renewal. (goldmansachs.com) PYMNTS, citing Goldman’s report, also reported on May 25 that the bank expected autonomous AI agents to drive a 24-fold increase in global token consumption by 2030. That secondary coverage matched the headline numbers in Goldman’s own article. ### What is the next place to check for more detail? Goldman Sachs’ Insights site is the primary published source for the forecast, and the article names Schneider as the analyst behind it. Goldman’s reports and artificial-intelligence pages are the next public places where related assumptions, follow-up notes or adjacent infrastructure forecasts are likely to appear. (goldmansachs.com 1) (goldmansachs.com 2) (pymnts.com)