The promise of artificial intelligence has never been louder. In 2023 alone, global corporate investment in AI reached nearly $200 billion, according to data from McKinsey and PitchBook. Yet behind the headlines of “AI revolution” and “productivity breakthrough,” a quieter but more persistent question is being asked in boardrooms and research labs: does the economic equation actually work?
The most direct reason for skepticism is the massive mismatch between costs and revenue. Training a frontier model like GPT-4 is estimated by independent analysts to have consumed more than 10,000 GPUs running for months, costing upward of $100 million. When you add the annual operating costs for inference — the process of generating answers for users — the total can exceed $1 billion per year for a major deployment. By contrast, OpenAI’s reported annualized revenue in early 2024 was around $2 billion, still far from covering its total expenses when accounting for infrastructure, personnel, and research. The company was reportedly losing $5.4 per dollar of revenue on its ChatGPT API in 2023, according to leaked financial documents cited by The Information.
This cost structure is not unique to OpenAI. Google’s DeepMind has never disclosed a profit from its AI products. Microsoft’s Copilot, despite strong adoption, has been described internally as generating “negative margin” per seat in 2024, due to high compute costs per query. The core problem is fundamental: current deep learning methods require a staggering amount of energy and hardware for both training and each subsequent query. Unlike software of the past, which could be scaled at near-zero marginal cost, AI inference is tied to expensive compute time. This is a regression to a factory-like business model, not the high-margin software play investors hoped for.
Skeptics often miss a crucial nuance: the cost is not static. Hardware improvements and algorithmic innovations are steadily reducing the cost per token of inference. For example, the introduction of Mixture-of-Experts architectures (used in GPT-4 and Gemini) can cut inference costs by 40–60% compared to dense models. Open-source projects like Mistral are demonstrating that small, specialized models can match larger ones on many tasks with a fraction of the resources. The key question is whether these efficiency gains will outpace the growth in demand. As more tasks are automated, usage volumes explode, potentially keeping total costs high even as unit costs fall.
A second economic hurdle is the difficulty of monetization. Many AI applications today deliver value that is real but fragmented. A legal assistant that saves 30 minutes per day is worth perhaps $5,000 per year, yet companies find it hard to charge more than $30 per month in a competitive market. The result is a race to the bottom: by late 2024, dozens of companies offered unlimited AI chat for under $20 per month. This price compression mirrors what happened to cloud storage and web hosting, but with the added burden that the marginal cost of serving each customer remains positive and sizable. In contrast, pure digital goods like a music streaming license can be shared across millions with near-zero incremental cost.
A different reading of the landscape comes from those who argue we are still in an investment phase, analogous to the early internet. In 1995, it was equally hard to see how Amazon or Google would ever repay their initial billions. The difference is that internet infrastructure had a clear slope of declining costs — Moore’s Law — and a network effect that increased value with users. AI’s cost decline trajectory is promising, but it is not automatic. It depends on breakthroughs in chip design, cooling, and algorithms. Moreover, the network effect for AI is weaker: a smarter model does not necessarily become smarter when more people use it, unless that usage generates high-quality training data.
From a policy perspective, governments are pouring subsidies into domestic AI champions, distorting the true economic picture. The European Union’s Digital Europe Programme allocated €2.1 billion for AI-related R&D from 2021 to 2027. China invested an estimated $13 billion in AI startups in 2023, much of it from state-owned funds. These injections lower the apparent cost of capital for developers but do not erase the fundamental unit economics. If subsidies are withdrawn, many business models will collapse overnight.
What would it take for AI’s economics to finally close? Three conditions seem necessary, none of which are guaranteed today. First, inference costs must fall by at least one order of magnitude more, to roughly the per-query cost of a Google search (about 0.3 cents). Second, applications must be bundled into must-have services with high switching costs, rather than generic chatbots. Third, a small number of platform players must consolidate the market enough to stabilize prices, avoiding the commodity trap. Without these, the economic case for AI will remain — to borrow a phrase from Silicon Valley — “pre-revenue and post-hope.”
The market is not wrong to be excited; it is wrong to assume that excitement pays the bills. The difference between a technological marvel and a sustainable business lies in the ugly mathematics of cost and price. Until the industry either slashes the cost per intelligence by a factor of ten or finds a way to charge customers ten times as much, the economic ledger will remain stubbornly in the red. The challenge is not whether AI can reason; it is whether anyone can afford to let it.