The internal source code of Claude Code, an Anthropic AI coding agent, has been made public, prompting industry expectations that demand for AI memory chips could be larger than previously estimated. The leaked code revealed a blind spot in AI agent infrastructure demand that the market has yet to reflect in prices.

On March 31, Chaofan Shou (차오판 쇼), CTO of U.S. security firm Fuzzland, found and shared that a package for Claude Code version 2.1.88 distributed on the npm registry included an internal debugging source map file intact. The file exposed about 512,000 lines of code to the outside. Anthropic said no customer data or credentials were included and that it was human error in the packaging process, not a security breach.

Memory usage seen in the leaked code is 15 GB when idle with no tasks running, and up to 8 to 9 times that when active with a user working. The figures are for a single developer PC, not a server.

The reason for the large memory consumption is shown in the function architecture designed in the code. Claude Code takes a user’s natural-language request, decides which tools the AI should use, checks permissions, executes the tools and then feeds the results back to the AI, repeating the process.

The code also designs a feature called KAIROS, an always-on agent mode intended to keep the process running continuously in the background without user involvement. Because Claude Code continues running while the user is away, memory remains occupied at all times.

The AutoDream feature additionally runs a separate Claude session to organise accumulated work logs while the user is resting. The structure runs separate inference operations for self-organisation even when the user is not working. As this organisation work becomes more sophisticated, the required computation and memory consumption also rise.

In addition, the code already implements a feature that increases by five times the amount of conversation Claude can handle at once. As conversation volume rises, the server memory needed for temporary storage increases by the same proportion. If a multi-agent structure in which one user operates 5 to 15 agents at the same time is combined with KAIROS’s always-on sessions, memory consumption per user increases far faster than the number of users.

◆Matches Nvidia and Arm moves to strengthen CPUs

Until now, analysis of AI memory demand has focused on server infrastructure, such as high-bandwidth memory for H100 and B200. Contrary to estimates that AI memory demand would rise in proportion to growth in user numbers, it ends up requiring more.

Early generative AI had a simple response structure that generated answers to user inputs, concentrating computational bottlenecks in GPU-centric compute sections. But an AI agent workload that takes actions directly, like Claude Code, has a multi-layer execution structure. As seen in the leaked code, Claude Code interprets user requests, calls and runs external tools, and reflects the results to repeat further decisions.

In this process, the CPU handles task sequencing, database access, external tool calls, and session and memory management. If the GPU processes the AI computation itself, the CPU manages the overall flow so that the computation runs in the right order with the right data. KAIROS and AutoDream in the leaked source code can be seen as cases where this multi-layer execution structure is implemented in practice.

The background to Nvidia and Arm’s recent emphasis on CPUs has been specifically confirmed through the leaked code. According to Eugene Investment & Securities, Nvidia presented the Vera Rubin Pod as an optimal data centre cluster rack system. The CPU share rose to 49 percent under the VR Pod standard, from 33 percent under the NVL72 rack standard. The separate launch of the Vera CPU rack is also interpreted as a change that reflects the rising importance of CPUs to prepare for memory bottlenecks in line with the spread of AI agents.

Arm also announced plans to provide AGI CPUs as rack systems. According to Eugene Investment & Securities, Arm projected that the market size for data centre CPUs would reach $100 billion by 2030 as the AI agent market emerges. Arm’s own projection that annual AGI CPU sales will reach $15 billion within the next 5 years can also be seen in this context.