The AI industry is selling the weapon that broke its own subscriptions

The subscriber who broke the flat rate had a name: the inference whale. A customer paying a flat monthly fee for an AI coding tool, whose automated workflows — a model looping over a long task, re-reading its own context — could burn thousands of dollars of compute against a plan priced for one human's questions. Cursor watched it happen on its own unlimited tier and put the mechanism plainly.

We've identified extreme usage by a small number of customers that impacts capacity for our broader community. — Anthropic

The unlimited tiers died through 2025. By this year the whole industry had re-priced: OpenAI moved its federal OneGov program to token-metered billing, Google shifted Gemini paid plans from daily prompt caps to compute-based limits, GitHub Copilot went consumption-based. Flat rates, the providers concluded, had proved economically unsustainable ^[1]. No one tried to repair the subscription. They abandoned it. In April, Anthropic fenced its own door. It blocked third-party agents — OpenClaw and others — from using Claude's flat-rate Pro and Max subscriptions, keeping the subscription for humans while steering the automated users toward metered access. The company's own explanation was the confession.

We've been working hard to meet the increase in demand for Claude, and our subscriptions weren’t built for the usage patterns of these third-party tools. — Boris Chernyshov

The fence went up in April; Meta's Muse shipped in September. The labs had not repaired the subscription model. They had priced the agents that broke it out of it — and then begun selling those agents to everyone else's subscribers. The product is explicit about what it targets. Muse is built to cancel forgotten subscriptions, negotiate internet bills, and recover airline credits — automating the work against exactly the friction that subscription businesses, insurers, and airlines monetize ^[2]. The average American household carries $1,887 a year in subscriptions ^[3], and the cancellation friction those businesses lean on is worth 14% to 200% of their revenue ^[4]. Meta's own plan is to collect a small transaction fee from vendors for every sale the agent funnels — a toll on the commerce it disintermediates ^[5]. OpenAI is pushing the same logic into the bank account. ChatGPT Finance connects users' accounts, automates budgeting, recommends credit cards, flags fraud ^[6]. Analysts see the next step already: agents moving household cash out of low-interest checking accounts, draining the cheap deposits that subsidize consumer banking ^[3]. A senior leader at Amazon's cloud unit said the quiet part in public: agentic commerce can raise sales while weakening the economics of the businesses selling through it, bypassing advertising revenue, inserting new intermediaries, and shifting influence over the customer away from the provider ^[7]. Markets are pricing the direction, not the damage. In the days around Muse's launch, shares of Booking.com, Airbnb, and Planet Fitness all fell — a booking platform, a lodging marketplace, and a gym chain whose business is precisely the recurring membership an agent can cancel ^[2]. Amazon knows the cost of that friction: it blocked Muse from shopping its platform after paying $845 million in Prime Video refunds over its own cancellation process ^[2]. None of this has landed yet. Agentic commerce is still under 1% of all e-commerce ^[8]. In Worldpay's survey, 6% of Americans would grant an agent full automatic buying control, 44% would let one browse, and comfort clusters under $50 ^[9]. Consumers were already dismantling the inertia on their own before the agents arrived — household subscriptions fell 32% in a year, and 70% of consumers say they want usage-based pricing ^[10]. Incumbents have live defenses, too. Wix is shipping its own AI tools and leaning on customer lock-in to hold margins ^[11]. Walmart is running its own agents toward half its online revenue by 2027 ^[12]. Sezzle grew revenue 67% by pivoting toward subscriptions rather than away from them ^[13]. Even the agent economy rebuilds what it dismantles: the specialized shopping tools that actually work charge commissions or subscription fees of their own ^[14]. The loop closes where the money was supposed to come from. Goldman Sachs found that only 2% of S&P 500 companies have quantified an AI effect on their earnings, and those that did show no statistically significant difference from peers — even as AI infrastructure firms' earnings rose 54% and carried roughly half of all index earnings growth ^[15]. The wallets the agents are built to strip belong to the customers who have not yet earned what the buildout was financed against. Part of that infrastructure profit is itself an accounting choice: stretched depreciation schedules leave an estimated $854 billion gap between cash spent and expenses recorded ^[16].