Generative AI stalls: No GPT-5, soaring costs

Generative AI hits the cost wall: pogre slows down

We have been waiting for over a year for the generational handover that artificial intelligence promised, and that handover is not arriving. Meanwhile, major labs continue to train increasingly larger models, but the results no longer match up. The latest cycle of generative AI has stopped impressing: different systems are equalizing in performance, and no breakthrough stands out above the others. The escalation of GPUs and data, the recipe that has worked for years, is now yielding diminishing returns.

The scaling ceiling

Physics sets limits. Silicon has its limits, and each chip improvement costs more and consumes more energy. Reports circulating among investors suggest that training models with more and more GPUs and data is no longer effective. This is not just a technical issue: the marginal cost of each performance increase has skyrocketed, and gains have become marginal. Several industry sources speak of stagnation, and the delay of GPT-5, which was certain a year ago, is the most visible symptom.

Costs do not balance in the accounts

There is a brutal disconnect between investment and return. While tech giants burn billions on data centers, subscription revenues are not taking off. The reality is that no one knows how to turn generative AI into a sustainable business: users do not pay enough, and companies cannot find use cases that justify the expense. The promise of AGI (Artificial General Intelligence), which would change everything, has become a receding horizon. Even OpenAI’s top executive, Sam Altman, who advocates for AGI arriving in 2025, acknowledges its impact will be lower than expected. It is hard to interpret these words as anything other than an exercise in managing expectations.

Chinese competition tightens

In parallel, Chinese models have taken a step forward at a fraction of the cost. Models like Qwen offer more natural language than Western proposals, and at a much lower price. The technological advantage taken for granted has evaporated, and the gap has become a burden: if the rival does the same spending much less, the Western scaling strategy loses all logic. Attempts at renewal, such as the o1 or o3 mini reasoning models, have failed to take off; in fact, some analyses point out that o3 mini is not superior to Chinese proposals. The response from major companies has been to accelerate the hype machine: now it is time to sell “agents” as the next great revolution, even if it is just a variant of what already existed, and without a clear business model behind it.

AGI as an act of faith

The technological debate has become a tug-of-war between skeptics and believers. For some, generative AI is a useful but limited tool, lacking self-awareness or projection, and unlikely to lead to general intelligence. For others, stagnation is just a bump in the road to singularity. But the facts are stubborn: after years of promised exponential growth, available data points to a plateau. And while evangelists talk about robots and a future of abundance, the only business that has truly flourished around AI is selling expectations. The shocking fact: despite billions invested and a $10 billion bonus for OpenAI’s CEO, no one has yet explained how the technology will be monetized. The product has ceased to be artificial intelligence and has become the promise.