AI Agents: 24x Token Growth Projections, Enterprise Cost Pressure, and the Agentic Business Thesis

Goldman Sachs Research projects AI agent token usage will grow 24-fold by 2030, driven by agents executing multi-step workflows that consume far more tokens per task than single-turn chatbot interactions [1][2]. Eric Schmidt argues publicly that founding an agentic AI company is the most direct path to profit right now, calling this 'the agentic period in AI' in which every business will eventually build competing agents [4]. On June 25, OpenAI published research showing agents enabling longer, more complex tasks across professional roles — and internal data suggests Codex, its coding agent, now accounts for roughly 99.8% of OpenAI's own token output [8][9]. Against this optimism, enterprises including Uber and Microsoft are already pulling back on expensive agent deployments due to cost pressure [1].

A 24x increase in token demand would make AI inference cost a primary enterprise budget item, not a secondary tooling expense. Whether token price declines keep pace with volume growth determines whether the agentic AI thesis produces durable business value or drives a cost-driven consolidation in which only a narrow set of use cases — possibly just software development — survive at scale.

Goldman Sachs Research issued a forecast, publicized widely from late May 2026, projecting that AI agent token usage will multiply 24-fold by 2030 [1][2]. The underlying logic is structural: agents do not return a single response but instead plan, call tools, verify outputs, and iterate — consuming orders of magnitude more tokens per task than a conventional chatbot. Goldman's accompanying analysis also concluded that token cost declines are expected to outpace price reductions, with cloud providers approaching a turning point in gross margins [3]. This positions the 24x volume growth not purely as a cost risk but as a potential revenue driver for infrastructure providers, assuming demand can be monetized before pricing deteriorates.

The supply-side investment thesis has a prominent advocate in Eric Schmidt, who has stated plainly that founding an agentic AI company is currently the easiest path to making money in AI [4]. His framing — 'this is the agentic period in AI; everyone's going to build agents; the agents are all going to compete' — is simultaneously an endorsement of the category and an implicit warning about commoditization. The quote circulated heavily in the week of June 19-25 across Twitter, Reddit, Instagram, and LinkedIn [5][6][7], functioning as a rallying signal for entrepreneurs and investors positioning around agents. OpenAI added institutional weight to this framing on June 25 by publishing research documenting how agents are enabling longer and more complex tasks across diverse professional roles [8]. One observer noted that within OpenAI itself, approximately 99.8% of internally generated output tokens now come from Codex, its coding agent, rather than from ChatGPT [9] — a data point that both validates the scale of agentic consumption and narrows its apparent current concentration to software development.

The demand side carries a countervailing signal. Despite bullish projections and executive endorsements, both Uber and Microsoft have been reported to be reconsidering expensive AI agent deployments because of cost concerns [1][2]. A social media commentator observed in June that the AI industry is simultaneously promoting agents as transformative and confronting cost realities that constrain their deployment [10]. Practical content on controlling agent costs and building total-cost-of-ownership frameworks for enterprise agents has emerged as a distinct category in parallel [11][12], indicating that practitioners are actively working cost management problems — not just accepting the productivity narrative at face value. The specific deployments being scaled back and the magnitude of any pullback at Uber and Microsoft remain undetailed in available sources, leaving open whether this is tactical recalibration or evidence of a wider cost ceiling.