Detailed Analysis
The economics of frontier AI products remain deeply unfavorable for the companies building them, and reporting from 36 Kr brings renewed scrutiny to the financial architecture underpinning both OpenAI's ChatGPT and Anthropic's Claude. Despite charging users up to $200 per month for premium subscription tiers, the underlying cost structure of serving computationally intensive AI workloads can dwarf that revenue figure by orders of magnitude. The reported exposure of up to $14,000 per user speaks to a structural problem in the industry: inference costs — the computational expense of actually running a model in response to user queries — scale dramatically with usage intensity, meaning that the heaviest and most engaged subscribers can consume resources that vastly exceed their subscription fees.
This dynamic is particularly acute for models like Claude, which Anthropic has positioned at the frontier of capability, including features such as extended context windows, complex reasoning tasks, and multi-modal processing. The more capable a model becomes, the more compute-intensive each query tends to be, and the more attractive it is to power users who generate disproportionate costs. Anthropic's $200/month "Claude Max" tier, introduced in 2025, was designed partly to capture revenue from this high-usage cohort, but the math remains challenging when a single engaged user conducting research, coding, or analysis tasks for hours per day can generate inference bills that dwarf their subscription contribution.
The broader context is one of an industry betting that costs will fall faster than revenue pressure forces a reckoning. Historically, compute costs in AI have dropped significantly as hardware improves and model efficiency increases through techniques like distillation and quantization. Both Anthropic and OpenAI are wagering that the current period of losses represents a necessary investment phase, with unit economics improving as model serving becomes cheaper and as enterprise contracts — which carry higher margins and more predictable usage patterns — become a larger share of revenue.
The 36 Kr framing reflects a growing conversation in both Western and Asian technology media about whether the AI subscription model is fundamentally viable or whether it represents a subsidized land-grab for users that will eventually require either dramatic price increases or a pullback in model capabilities offered at consumer price points. Anthropic, which has raised tens of billions in capital from investors including Google and Amazon, has positioned itself as a safety-focused research organization, but its commercial sustainability is inseparable from solving this unit economics problem. The company's API business, where enterprise customers pay per token, offers more pricing flexibility and is likely where Anthropic sees a clearer path to profitability.
The reporting underscores a fundamental tension in the AI industry's current moment: the products most valued by users — deeply capable, always-available AI assistants — are precisely the ones most expensive to operate. Until inference costs fall dramatically or usage can be more precisely metered and priced, the gap between what consumers pay and what they cost to serve remains one of the defining financial challenges for both Anthropic and its competitors.
Read original article →