The $200 paradox: why frontier labs lose money even on their premium plans

A note of honesty before we begin: the SemiAnalysis article is behind a paywall and it was not possible to read its body. What follows is based on the newsletter's headline and on widely documented public facts, without attributing to the publication specific figures we have been unable to verify. That said, the thesis the headline anticipates —that the frontier labs cannot find profitability even when charging premium prices— deserves comment, because it touches the economic nerve of the entire industry.

The emblematic case is ChatGPT Pro, the $200-a-month plan that OpenAI launched in late 2024. The intuition the analysis suggests is that, for the most intensive users, the cost of serving them in compute comfortably exceeds what they pay. And it makes qualitative sense: that plan is designed for those who squeeze reasoning models over long, chained sessions. It is precisely those 'thinking models' that are the most expensive to serve, because they generate enormous amounts of internal tokens before delivering a single visible line. Every polished answer hides a disproportionate inference bill.

Here a caveat that rigour demands must be kept in mind: these total-cost-of-ownership analyses are extremely sensitive to the assumptions. The utilisation rate of the clusters, the model mix, the volume of input and output tokens, or whether the training cost is amortised or not, change the result entirely. The most striking gap almost always describes the extreme-usage user, not the average customer who opens the tool now and then. Generalising 'every subscriber loses X' would be a mistake; the problem is structural, but its magnitude depends on the profile.

That SemiAnalysis is a reference on this is no accident. The publication run by Dylan Patel builds bottom-up models with real semiconductor supply-chain data, GPU contract prices and datacenter architectures —a visibility most commentators lack. That is why its hardware-cost estimates carry more weight than the usual high-level analyses. The backdrop, moreover, is well known and acknowledged by the companies themselves: OpenAI has admitted it is not profitable and Anthropic operates at substantial losses. Competitive pressure pushes them to offer high-end capabilities at consumer prices while costs remain enormous.

The sector's implicit strategy is the same as ever in tech: grow in users and data, trusting that economies of scale, better architectures and cheaper silicon will eventually close the gap. It may happen. But the break-even point remains uncertain, and hence the practical lesson for anyone building agentic products on these APIs: current prices may not reflect the real long-term cost. Designing an agent that consumes advanced reasoning intensively is, today, expensive to operate at scale, and the sector's pricing policy is subject to revisions. Building with that hypothesis in mind —and not assuming that today's rate will be tomorrow's— is simply sound engineering and business prudence.