SemiAnalysis: frontier labs can't find profitability — ChatGPT Pro consumes ~$14,000 in compute per month for just $200/month

⚠️ CONTENT BEHIND PAYWALL: The full SemiAnalysis article requires a paid subscription. The scraper was only able to download the navigation interface, product menus, cookie notices and subscriber-acquisition banners of the website—no paragraph of the article body is available. Therefore, the summary that follows is based exclusively on the title as it appeared in the newsletter forwarded to Manuel and on widely documented public facts, without adding any invented figure or claim.

The digest title anticipates the analysis's central thesis: the frontline artificial intelligence labs—called 'frontier labs'—are for the moment unable to find a profitable business model, even when they charge premium subscription prices. The concrete example SemiAnalysis would have addressed is ChatGPT Pro: the $200-a-month plan that OpenAI launched in late 2024 would be costing the company much more in compute per user than it brings in, according to what the article suggests (the exact figure is not available).

This gap between sales price and real cost is not surprising in qualitative terms. Since OpenAI launched ChatGPT Pro, multiple sector analysts pointed out that the plan was designed for ultra-high-usage-intensity users—researchers, programmers and professionals who exploit the capabilities of large models like o1 and o1 Pro during long sessions of chained reasoning. Reasoning models ('thinking models') are precisely notoriously more expensive to serve because they generate large amounts of internal tokens before producing a visible response, which multiplies the inference cost relative to conventional models.

SemiAnalysis is one of the reference publications in the economic analysis of semiconductor and artificial intelligence infrastructure. Its usual methodology consists of building bottom-up cost models: they estimate the number of GPUs needed to serve a given level of traffic, apply real hardware costs (typically NVIDIA H100 or H200 clusters), add energy, cooling, network and personnel costs, and obtain a total cost of ownership (TCO) per token generated. If the cost of serving these users far exceeds the $200 subscription price, it would represent a structural operating loss of considerable magnitude.

It is important to contextualize this situation with caution, given that we have not been able to read the article. The cost difference probably corresponds to a maximum or extreme usage user profile, not to the average ChatGPT Pro user. AI inference TCO analyses are very sensitive to assumptions about cluster utilization rate, the mix of models used, the volume of input and output tokens, and whether or not amortized training costs are computed. A user who uses ChatGPT Pro occasionally for simple tasks probably costs much less than an intensive user.

The underlying problem the article seems to address is structural for the entire sector: competition among OpenAI, Anthropic, Google DeepMind and Meta is leading the labs to offer increasingly advanced capabilities at consumer prices, while training and inference costs remain enormous. OpenAI has publicly acknowledged that it is not profitable. Anthropic also operates with substantial losses. The implicit strategy is to grow in users and data in the hope that economies of scale, improvements in model architectures and falling silicon prices will eventually close the gap—but that break-even point remains uncertain.

SemiAnalysis's role in this debate is relevant because it brings quantitative rigor to the analysis of hardware costs, an area where most commentators lack the visibility needed to make reliable estimates. The publication, led by Dylan Patel, has access to semiconductor supply-chain data, GPU contract prices and datacenter architectures that allow it to build more precise models than the high-level analyses common in mainstream media.

For the readers of Manuel's Agentic AI newsletter, the most relevant practical implication is the following: the current economics of frontier models make AI agents that intensively consume advanced reasoning capabilities extremely costly to operate at scale. Any product or company building on the APIs of models like o1, Claude 3.5 or Gemini Ultra must keep firmly in mind that current API prices may not reflect the real long-term costs, and that the sector's pricing dynamics are subject to significant revisions as the labs seek to move toward profitability.

📌 RECOMMENDATION: To access the full analysis with the detailed cost models, it is necessary to subscribe to SemiAnalysis at newsletter.semianalysis.com. The subscription price is significantly lower than what it apparently costs to serve a ChatGPT pro user.