Imagine a world where machines not only calculate, but also reason in a way analogous to a human being: solving complex equations, proving theorems or generating flawless code instantly.
Today, this dream is a reality thanks to AI reasoning models, which have established themselves as essential tools in fields as diverse as mathematical research or advanced programming.
However, a recent report from Epoch AI, a non-profit organization dedicated to researching trends in artificial intelligence (AI), warns that the exponential pace of progress could stall sooner than expected.
While some experts envision a next computational and economic “ceiling,” others, like Sam Altman of OpenAI, continue to bet on new innovations. Are we on the verge of a slowdown or are we simply facing a temporary blip?
Keep reading and discover what could stop these advances, the promises that still maintain optimism and the avenues of research that aim to overcome current barriers.
Why might progress be slowed?
Epoch AI maintains that the escalation in resources needed to train reasoning models is reaching unsustainable levels.
For example, to move from model o1 to o3, OpenAI invested approximately ten times more computing power, concentrated in refinement phases through reinforcement learning.
If this trajectory continues, it is estimated that the development of the following generations could exceed 1 billion euros in 2027, a figure that puts the margins of any research laboratory in check.
Infrastructure limits
The production of semiconductors and the availability of electrical energy are two critical bottlenecks. Data centers already consume a high percentage of global energy, and the manufacturing of advanced chips faces supply constraints and rising costs.
According to Epoch AI, up to 2·10²⁹ FLOP would be needed to achieve the 2030 goals, a magnitude that far exceeds the current installed capacity today.
Diminishing returns
Traditional “scaling laws” indicate that each increase in power or parameters produces improvements in performance that follow a curve of diminishing returns.
That is, as the model and the computation grow, the gain per extra unit of resource decreases, making each advance more costly and less profitable.
Reasoning models have relied on scaling both size and computation (and also inference time) to break records in specialized benchmarks. Epoch AI and other analyzes project that it will soon no longer pay to scale just by adding computing power.
What do reasoning models promise us?
In the previous part we talked about why we could reach a brake (costs, infrastructure, diminishing returns). So it’s time to remember why we invest so much in these models.
Solutions to complex problems
Despite the shadows, the progress has been surprising. Models like o3 or DeepSeek-R1 can:
- Prove mathematical theorems of moderate complexity in minutes, a task that previously required weeks of human work.
- Generate code with a level of correctness that is close to 90% on platforms such as LeetCode or GitHub Copilot.
- Solve logic problems and process long chains of reasoning coherently.
These capabilities are transforming sectors such as academic research, software engineering, and quantitative analysis.
Cognitive automation
It is estimated that by 2027, these models will be able to outperform the best mathematicians in theorem proofs with a probability of around 60%. This opens the door to:
- Automate code audits and smart contracts.
- Advanced optimizations in quantitative finance.
- Research assistants who formulate preliminary hypotheses or experimental designs.
What have reasoning models NOT been used for?
When we leave the purely digital domain, the effectiveness of reasoning models drops.
In robotics, for example, training a mechanical arm to move precisely in an unpredictable environment requires huge volumes of real-world data and costly trial-and-error cycles.
The same goes for advanced video editing or the creation of high-quality generative art, where aesthetic intuition and creative “touch” come into play.
Degradation in very long chains of reasoning
Although models have improved their chain-of-thought, they continue to lose coherence when a task requires dozens of intermediate steps.
Prompt engineering techniques (chained prompts, task decomposition) help, but add complexity and do not guarantee robust results in all cases.
Is it worth continuing to invest in AI?
Companies like Google or OpenAI must decide whether to continue investing in reasoning models despite the rising cost of computing. An alternative could be to diversify your technology portfolio by targeting other more accessible approaches, such as federated learning or explainable AI (XAI).
A prolonged slowdown could chill venture investment and slow the ecosystem of startups focused on advanced logic AI.
Bid for talent and resources
The increase in operating costs also puts pressure on the salaries of specialized engineers. As Big Tech competes to hire elite researchers, the barriers to entry for smaller companies become virtually insurmountable.
How to overcome the “ceiling” in reasoning models?
Modular architectures and composite agents
Projects like Ember (Nvidia + Google + Foundry) promote pipelines where several specialized models collaborate: one manages the logical planning, another the search for external data and a third the verification of results.
This reduces the load on each person and optimizes the use of resources.
Mixture of Experts (MoE)
Instead of activating the entire model on each request, MoE architectures invoke only certain “experts” (subnets) relevant to the task at hand. This allows parameters to be scaled without linearly increasing the inference cost. This is one of the approaches applied in Deepseek.
Information retrieval and hybrid generation
Integrating external knowledge bases (retrieval-augmented generation) helps minimize hallucinations and reduce the need to memorize information in the model itself.
The system consults documents or databases in real time, thus combining memory and reasoning.
Hardware innovations
Specialized chips (next-generation TPUs, energy-efficient ARM-based GPUs) promise to multiply performance per watt. Designing sustainable data centers, with liquid cooling and renewable sources, can alleviate the energy bottleneck.
Paradigms beyond deep learning
Neuro-symbolic approaches are explored, which combine neural networks with traditional symbolic reasoning engines, and causal models, which aim to understand cause-effect relationships rather than pure statistical correlations.
Between worry and hope
AI reasoning models have raised the level of what we understand by artificial “intelligence”, conquering challenges that only a few years ago seemed unattainable.
However, reports from Epoch AI and signs of diminishing returns warn us that advances based solely on increasing size and compute could peak in the next 12 to 24 months.
However, the outlook is not completely bleak. Research in modular architectures, MoE techniques, hybrid external memory systems and the development of increasingly efficient hardware offer promising routes to extend the improvement curve.
The next chapter of reasoning AI is already underway: it will be written with smarter approaches, not just bigger ones.
The question is whether this additional intelligence will arrive in time to avoid the “bump” that is coming, or if we will have to adapt to a stage of slower, but perhaps more sustainable advances. Under this new logic, the real challenge will no longer be “can they think like us?”, but “how can we make them think better and with fewer resources?”
This post is also available in: