"AI Is Infrastructure"… Cost Structures Changing Platform Strategy
As AI startup Anthropic changed the usage policy of its coding assistant 'Claude Code' beginning April 4, 2026, debate is spreading across the industry. The core of this measure is that use through third-party tools like OpenClaw is excluded from existing subscription limits and a separate usage-based pay-as-you-go billing is applied.
In particular, since this policy is scheduled to expand not limited to specific tools but to all external tools, it is interpreted as a platform strategy change rather than a simple function restriction. The 'harness' mentioned here means an external program that makes AI automatically and repeatedly run. That is, rather than a person using it directly, a method where the program continuously calls AI on their behalf.
This measure, coinciding with the OpenClaw developer's move to a competitor, also has analysis emerging that it is a signal of the competitive landscape beyond simple technical policy change.
The essence of the debate lies not in technology but in economic structure. The existing AI subscription model was designed based on the premise that 'most users don't use up to the maximum limit.' Flat fees were possible thanks to a structure where a few heavy users' costs are supplemented by the many who use less.
However, the situation changed with the emergence of automation tools like OpenClaw. With AI being called repeatedly, usage surged and a structure formed where virtually all users become high-usage users. Evaluations emerge that the premise of existing subscription models is being shaken as token usage has exploded and GPU computation costs have increased.
Subscription models originally stand on the assumption that most users utilize only at a limited level. Even if some users use a lot, the structure maintains overall balance. However, automation tools break this balance and rapidly increase costs that must be borne from the corporate perspective. Ultimately the sustainability of the subscription model itself cannot help but be shaken.
In this process, the essence of AI services is also drawing renewed attention. Internet services don't incur large additional costs even when usage increases, but AI is different. Computation is needed to process each request, generating power and server costs accordingly. That is, it is a structure where increased usage directly leads to increased costs. Because of this, AI is evaluated as an industry closer to 'infrastructure' than simple software.
Interpretations of this policy are divided. One side sees it as a rational measure protecting all users by restricting excessive use by some users. On the other hand, criticism is also raised that it is a strategy to block external tools and induce use centered on one's own platform. Ultimately it reveals a fundamental difference in recognition about "how far subscriptions are permitted."
Another problem revealed is 'heavy users.' Existing subscription models stood on the balance of some heavy users and many general users, but automation tools have the effect of breaking this structure and making all users heavy users. This is evaluated as a case revealing the structural limits of subscription models.
Trust issues are also being raised among users. The concern that initially providing services in an inexpensive subscription form then gradually increasing restrictions and transferring costs. With external tool restrictions added on top, the perspective is also spreading that AI companies are prioritizing their own interests over users.
Ultimately this case shows the direction of the AI industry beyond simple policy change. Going forward, usage-based billing is likely to expand more than flat-fee subscriptions, and companies are highly likely to strengthen control centered on their own platforms. Simultaneously, open-source technology that can be used more freely in response to this is also expected to continue growing.
The question this debate poses is clear. Whether we will consume AI as a 'monthly subscription service' or must accept it as 'infrastructure that pays by usage amount' is a matter of choice.
