[Weekend Money] How ChatGPT Astra Will Change the Landscape

From Simple Conversational Chatbot to Always-On AI Workforce
"Exponential Growth in Memory Demand"

It has been forecast that OpenAI's recently released ChatGPT Astra will accelerate the proliferation of artificial intelligence (AI) agents, which in turn will significantly drive up memory demand.

Junyoung Park, a researcher at Hanwha Investment & Securities, stated in his recent report titled "Astra Signals Explosive Memory Demand Growth" that the spread of AI agents will substantially increase memory demand.

He assessed that ChatGPT Astra marks a turning point where the core nature of generative AI shifts from a simple question-and-answer chatbot to an autonomous AI agent capable of completing work on its own. While previous models required users to instruct every step, Astra can be assigned a goal and autonomously completes complex, multi-step work streams—moving between web browsers and enterprise resource planning (ERP) systems—without constant user intervention.

Astra’s mechanism resembles that of an office worker. Instead of staying idle during file downloads, much like an employee multitasking and working on other documents for approval, Astra performs multiple tasks by invoking asynchronous tools. It also manages long-term context, retaining previous instructions and decisions during extended work sessions. Astra records key points as work notes and retrieves previous context as needed. Here, context refers to instructions, reference data, and intermediate results needed by the AI for its current work. Instead of a single person handling all tasks sequentially, the system allows sub-agents—such as data analysts or reviewers—to work concurrently under the direction of a lead agent, resembling a project manager overseeing a team.

Besides Astra, Anthropic’s Claude and other large AI models are evolving into background infrastructures that run 24/7 in the cloud, even when users step away or close their laptops. The trend is shifting from one-off training to constant inference.

As a result, Park projected that token consumption will increase sharply and memory demand will soar. Compared to standard chat, a single agent consumes four times as many tokens, while a multi-agent system consumes up to fifteen times more. Previously, computing demand was estimated by multiplying user numbers by the average number of questions asked, but now it is determined by multiplying the number of agents per user by their active work time, which drastically increases required infrastructure runtime. In other words, every variable multiplies, so memory demand growth will far outpace the growth rate of AI users.

For example, high-bandwidth memory (HBM) is expected to see increased module capacity per graphics processing unit (GPU) and a shift to higher bandwidth standards, since it is used both for real-time operation of AI models and for the KV Cache (temporary memory) required during processing. Server DRAM holds context and intermediate results waiting for further processing, which cannot all be accommodated by costly HBM. Solid-state drives (eSSD) will also see growing demand as they store long-term accumulated work logs, original contracts, and large multimedia datasets.

This content was produced with the assistance of AI translation services.