2026-08-20
Source: AI Agents Forecast to Boost Tech Cash Flow as Usage Soars
Main points:
- Agentic AI is expected to drive a 24-fold increase in token consumption by 2030 as consumers and enterprises adopt the technology, according to Goldman Sachs Research.
- AI chipmakers’ token unit costs are falling, setting the stage for gross margin improvements at hyperscalers as demand for agent applications rises.
- For the next 12 to 18 months, there is likely to be a shortage of chips as semiconductor makers build new plants to catch up with demand.
- Business adoption of agentic AI will leapfrog consumer use, but it is expected to take time as companies manage organizational challenges.
---
Cecile G. Tamura's post:
The agentic economy is beginning to take shape — and the numbers are getting enormous.
Goldman Sachs Research estimates that AI-agent token consumption could increase 24× by 2030, reaching roughly 120 quadrillion tokens per month.
Why?
Because AI is moving from chatting to doing.
Chatbots answer one prompt at a time. Agents can plan, retrieve information, call tools, execute tasks, check their own work and keep operating in the background. That makes them dramatically more compute-intensive.
And the transition is already underway.
Wave 1: Chat — humans ask, AI answers.
Wave 2: Agents— AI executes workflows for consumers and enterprises. Goldman expects enterprise agents to become the largest driver of incremental token demand, with enterprise and consumer agents together pushing usage to that 24× level by 2030.
Wave 3: Physical AI — the same intelligence begins moving into robots, autonomous machines and other physical systems.
That third wave could be the really interesting one for infrastructure.
Every digital agent needs inference. Physical AI potentially needs inference plus sensors, memory, accelerators, networking, batteries, motors and power.
And there is another important piece of the equation: the cost of inference is falling rapidly. Goldman estimates inference cost per token has been declining roughly 60–70% per year. If compute gets cheaper while the number of AI tasks explodes, entirely new workloads can become economically viable.
That is why the AI infrastructure story may be much bigger than chatbots.
The bottleneck is gradually shifting from "Can AI do it?" to "How much intelligence can we afford to deploy — and how much infrastructure will it require?"
The chart below captures the transition beautifully:
Chat → Agents → Physical AI.
The AI economy isn't simply adding more users.
It is adding more work.
And work consumes compute.
Không có nhận xét nào:
Đăng nhận xét