Taiwan index futures have held steady above 40,000 points, and US stocks hit fresh highs—yet the AI industry is only just entering its early stage?

Citrini Research researcher Jukan reposted a chart from Goldman Sachs Global Investment Research and said: “We’ve only just entered the early stage.” According to the chart, the key changes in the AI industry going forward may be not only improvements in model capabilities or an expansion in compute supply, but also consumer agent workloads—AI agent work tasks aimed at consumers—which would significantly boost global token consumption.

Consumer agents could increase token usage by 10x before 2030

The chart shows that consumer agent workloads could lead to token consumption increasing by more than 10x before 2030; and the red annotation in the chart goes even further, stating that token consumption will grow by over 12x by 2030. Goldman Sachs believes token growth is driven by three factors: broader user reach, higher daily usage frequency, and a shift in AI usage patterns from single chat sessions to on-demand agents and always-on agents.

So-called tokens are the basic unit of measurement used when large language models process text, instructions, responses, tool calls, and contextual memory. When users only occasionally ask chatbots questions, token consumption is relatively limited.

But once AI agents start continuously searching, monitoring, organizing, placing orders, scheduling, replying to messages, and handling workflows for users, the model is no longer just “answering questions”—it becomes a form of digital labor that runs continuously. This means the number of tokens each user consumes per day could shift from one-off conversations to high-frequency, long-duration, background-execution workloads.

In 2026, global token capacity is around about 7.5 trillion tokens per month; but after entering the first half of 2026, Goldman Sachs labeled it “Token Economics Turn Positive in 1H26,” meaning token economics could turn positive in the first half of 2026. In other words, as inference costs fall, compute infrastructure expands, and use cases mature, the unit economics model for AI companies to process tokens may start to improve.

Before 2030, non-agent workloads will continue to grow, but what truly drives the slope of the curve is consumer agents. Non-agent workloads mainly refer to existing AI usage patterns such as traditional chatbots, search, content generation, and general enterprise applications; consumer agents represent AI usage scenarios with higher frequency, longer duration, and more automation. By 2030, the chart estimates that monthly tokens processed will exceed 60 trillion, with consumer agents accounting for a substantial portion of the incremental source.

This is also the core meaning of what Jukan meant by “we’ve only just entered the early stage.” If Goldman Sachs’ forecasts hold, AI demand will not remain limited to today’s chatbots, code assistants, or enterprise copilots, but will further move into agentic workflows—the stage where AI agents continuously carry out tasks for humans. At that time, the focus of market discussion will shift from “how many people use AI” to “how many tasks people have AI execute every day.”

BlackRock CEO rebuts the AI bubble: The real problem is a shortage of compute supply

This also echoes recent pushback from Larry Fink, CEO of BlackRock, on the “AI bubble” narrative. As reported by Bloomberg, Fink said at a panel discussion at the Milken Institute Global Conference that the market is not facing an AI bubble right now, but a serious supply shortage. The pace of demand growth far exceeds what the market expects. It’s not only the United States that lacks production capacity in areas like computing power, chips, and memory—the world is only beginning to explore the enormous opportunity AI brings.

Fink even expects that as compute demand continues rising and the supply shortage can’t be resolved quickly, future markets may develop trading mechanisms for “compute futures,” enabling compute capacity to become a new asset class in futures markets. This view is important because it redefines AI infrastructure from “technology companies’ cost expenditure” into a scarce resource that can be traded, financialized, and priced over the long term.

In other words, the Goldman Sachs chart is about the demand side: consumer agents will drive token consumption to grow by over 12x before 2030; what Fink is talking about is the supply side: if token demand truly explodes, the market must confront shortages in compute, chips, memory, data center capacity, and power supply. Together, these form the core of the current bullish narrative around AI infrastructure buildout.

This article, “Taiwan stock index holds above 40,000 points and US stocks hit new highs for innovation, but is the AI industry only just entering the early stage?”, first appeared on Lian News ABMedia.