NextFin News - The economics of the generative AI boom are facing a quiet but severe reckoning as the cost of "agentic" computing—where AI models perform complex, multi-step tasks like software engineering—threatens to outpace subscription revenues by a factor of ten. While Silicon Valley giants Anthropic and OpenAI have seen their annualized revenues soar to $30 billion and $25 billion respectively by June 2026, a growing body of evidence suggests that their most advanced users are being heavily subsidized by venture capital and enterprise reserves.
Gerben Wierda, an IT architect and analyst known for his critical examinations of enterprise technology, recently detailed a research project involving Anthropic’s "Claude Code" that exposes a staggering gap between retail pricing and operational reality. Wierda, who has long maintained a cautious stance on the long-term viability of current LLM architectures, found that a user on a $100-per-month "Claude Max" plan could easily consume tokens that would cost over $1,000 at standard API rates. This discrepancy arises when models are pushed into "high effort" modes, where they engage in recursive "thinking" processes that remain largely invisible to the end user.
The financial strain is most visible in the rise of "dark tokens"—computational cycles used for internal reasoning, trial-and-error, and tool-calling that do not appear in the final chat output but consume massive amounts of GPU power. According to Wierda’s analysis, a single complex task performed by a top-tier recursive model can cost upwards of $75 at current API rates. For a heavy user, the $100 monthly subscription fee is exhausted within days, leaving the provider to eat the remaining costs of the compute cycle. This dynamic suggests that the current "all-you-can-eat" subscription model for frontier models may be fundamentally unsustainable in its current form.
This "token pricing crisis" comes at a time when the competitive landscape is shifting toward enterprise dominance. Anthropic recently reported that 80% of its revenue now stems from business customers, a stark contrast to OpenAI’s more consumer-heavy mix. However, even these enterprise contracts are under pressure. While OpenAI expects to spend a staggering $121 billion on compute in 2026, the efficiency of that spend is being questioned. Anthropic’s move toward Opus 4.7 and 4.8 versions appears to be a strategic attempt to "stop the bleeding" by optimizing token usage, though Wierda warns this may signal the beginning of a plateau in model capability as providers prioritize cost-cutting over raw performance.
The broader market remains divided on whether these costs will eventually be solved by hardware breakthroughs or if they represent a structural flaw in the LLM business model. Google recently claimed its Gemini 3.5 Flash model could save enterprises $1 billion annually in AI costs, yet the "frontier" models required for high-level reasoning continue to demand premium compute. From a buy-side perspective, the risk is that the "killer app" of AI-driven coding is being built on a foundation of artificial pricing that cannot survive the transition from growth-at-all-costs to a profit-driven environment.
As the industry moves toward 2027, the reliance on massive infrastructure deals—such as Anthropic’s multi-gigawatt agreement with Google and Broadcom—indicates that the scale of investment is still accelerating. Yet, if the cost of a "resolution" remains ten times higher than the price paid by the subscriber, the music may stop long before the AI can truly build itself. The current era of subsidized intelligence offers a window of high productivity for those who can leverage it, but the underlying ledger suggests a correction is inevitable as providers move to align retail prices with the brutal reality of inference costs.
Explore more exclusive insights at nextfin.ai.
