Large language models such as ChatGPT have transformed the debate about AI and the stock market since their release in mid-2022. While tech stocks were down and expectations for AI were at a 10-year low at the time, the release of ChatGPT revived predictions for a productivity revolution and $13 trillion in economic gains from AI by 2030. The market capitalizations of tech stocks have risen by more than $10 trillion. Or to use the old adage, the rest is history.

Unfortunately, history is again being rewritten. It is rapidly becoming clear that LLMs are not economical. Their huge training costs, along with required updates to the models, make their operating costs, sometimes called inference costs very high. Systems like RAG (Retrieval-Augmented Generation) are used to update information, but they are part of the higher costs.

Simply put, a core model that is not refreshed every few months quickly turns from a helpful assistant into a tech debt generation machine with outdated knowledge about news, libraries, interfaces and application programming interfaces (APIs).

OpenAI is at the heart of the problem

The firm had revenues of $3.5 billion and costs of $5 billion in 2024, and those losses continue to rise. It has recently forecasted $115 billion in cumulative losses before it becomes profitable in 2029.

Image Credit: TStudious - Adobe Stock

Those losses dwarf Amazon’s $3 billion in cumulative losses before it became profitable in 2024, its tenth year of existence. That’s the same age as OpenAI is now.

They also dwarf Uber’s almost $40 billion in peak cumulative losses before it became profitable two years ago. Uber only became profitable after the decimation of the public transport system during the lockdowns, which enabled Uber to raise prices, something that is much harder to do in today’s world of AI.

Behind OpenAI’s losses are its low prices, set far lower than the costs that it incurs when its users access cloud services from Microsoft, Google, and others in order to receive answers from ChatGPT. This arrangement, of course, benefits those who provide cloud services and those who construct new data centers to provide those cloud services.

The construction of such centers has become so large in 2025 that it creates huge risks for the U.S. economy. One risk is that the current AI system doesn’t provide more value than cost. That is because, as mentioned above, OpenAI and others have set their prices at several times less than their costs, something that is unknown in the history of economics. And any student who passes Econ 101 knows that just raising prices suddenly can kill the goose that’s laying the golden eggs.

Costs were supposed to fall, and they have. The problem is that demand for more accurate responses has also increased. Users are tired of fixing the inevitable hallucinations. They want more accurate models, including reasoning models. The result is that more AI resources, often called tokens, are consumed to provide a little more accuracy, even as token prices fall.

What difference does GPT-5 make?

OpenAI’s latest model, GPT-5, has barely changed the economics of AI. It does not provide much better accuracy. Instead, it routes simple questions to old models to reduce costs. This hasn’t satisfied users and further price increases seem like the only way forward, a solution that would likely dramatically reduce the demand for data centers.

One solution is small language models (SLMs). The term may be unimaginative, but it accurately captures their advantage. Companies, whether they are users or AI startups, can apply OpenAI’s models to a limited set of data and questions. Thus they create a smaller model that is cheaper to train, update, and run than OpenAI’s system. They can do this because LLMs are compressed storage and retrieval systems. Any data stored in the LLM is retrievable through interaction with it (that’s the point of the system, after all). With enough interaction data, any “emergent” model capabilities are retrievable too. 

Some claim that companies can switch over to a smaller model with as few as several tens of thousands of interactions with an LLM, thus speeding an exodus from OpenAI.

Many specialized corporate processes leverage only a small part of the overall knowledge stored in an LLM and thus they can be very small compared to LLMs.

DeepSeek and data centers

DeepSeek used this approach to develop a cheaper system, and the approach has since been copied by many American startups. Trained on the output of a creation process, transformer models have learned to replace their creator. And it doesn’t really matter in this context if the creator is human or a more expensive AI.

Image Credit: Gorodenkoff - Adobe Stock

On the bright side, these lower costs might make many more corporate applications economical and thus start Americans down the road towards using AI effectively.

On the dark side, the more efficient models would mean lower demand for cloud centers and thus less construction would be needed in the short run. Fewer cloud centers mean that many of the trillions that have been allocated for them in a multiyear planning and construction process would have to be written off. That would certainly cause the huge AI bubble to pop.

The more efficient models would also push OpenAI out of its central position. OpenAI will be the biggest loser because it tried to lock-in users with flat rate deals that have cost it billions. CEO Sam Altman has famously bragged about the company losing money even on its $200 per month subscriptions.

Copyright infringement?

Ironically, suppliers of LLMs have no effective legal defense. While the entire industry, from OpenAI to Anthropic and Grok, has extracted billions of dollars of value from authors, artists and video owners without providing compensation, companies such as DeepSeek are now doing this to the suppliers of LLMs.  Nobody has the clean hands required to sue for copyright infringement.

At the same time, data owners are now aware of the value they hold. Data suppliers like Reddit are aware that LLMs and SLMs need fresh data, and they are some of the few that hold that data.

A new value chain for generative AI is on the way, but it will likely not fully emerge until after the current bubble pops. Trillions will be lost before the real value creation begins.

Georg Zoeller is the Co-Founder of the Centre for AI Leadership, the VP of Technology at Novi Health and a former Meta/Facebook Business Engineering Director.