AI labs pivot to reasoning models as traditional scaling hits limits

The race to build larger artificial intelligence models is hitting a physical limit, forcing the industry's leading labs to pivot to a new strategy: teaching their systems how to think before they speak.

For years, the formula for better AI was simple. Developers poured more data and more computing power into the initial training phase, creating larger foundation models that could instantly predict the next word in a sentence. But as high-quality human data runs dry and the cost of building massive data centres climbs into the billions, this brute-force approach is yielding diminishing returns.

Instead, researchers are focusing on "inference-time compute"—giving a model extra computational room to process, plan, and self-correct while it is generating an answer.

What is inference-time compute?

Traditional large language models operate on what psychologists call "System 1" thinking: fast, instinctive, and automatic. When you ask a standard chatbot a question, it begins generating its response immediately, token by token, without the ability to plan ahead or revise its thoughts.

The new generation of reasoning models introduces "System 2" thinking. When presented with a complex prompt, these models generate an internal "chain of thought" before delivering a final answer. They try out different angles, spot their own mistakes, and try again—all behind the scenes.

This shift means a model might spend several seconds, or even minutes, calculating an answer to a difficult mathematics or coding problem. While this makes the interaction slower, the accuracy of the output is significantly higher.

Why the industry is shifting

The pivot addresses two urgent problems facing AI developers.

First is the data bottleneck. The internet is finite, and tech companies have already scraped much of the high-quality text available. By shifting the computing burden from the initial training phase to the moment the user asks a question, labs can improve a model's reasoning capabilities without needing trillions of new words to train on.

Second is cost. Training a frontier model from scratch requires an upfront investment of hundreds of millions of pounds. Inference-time compute allows developers to scale a model's intelligence dynamically. A simple query about a recipe requires a fraction of a second of thought, while a complex software engineering task can be allocated an hour of computational reasoning.

Who is leading the transition

OpenAI pioneered this transition publicly with the release of its o1 and o3 models, which demonstrated sharp improvements in competitive programming and PhD-level science benchmarks.

However, the competitive dynamics changed when Chinese firm DeepSeek released its R1 model. DeepSeek published its training methodology, showing that similar reasoning capabilities could be achieved at a fraction of the cost using open-source architectures. This has forced rivals to reconsider how much they charge for high-end reasoning.

Google has also integrated reasoning chains into its Gemini family, while Meta is reportedly training its next Llama models to use similar multi-step planning techniques.

What happens next?

The ultimate goal of this research is the creation of reliable AI "agents"—systems that can be set a broad goal, such as "research this market and write a report," and left to work independently.

Without the ability to self-correct and plan, early agents frequently got stuck in infinite loops or hallucinated incorrect information. By grounding agents in reasoning models, developers hope to build systems that can work through multi-step tasks over hours or days, checking their own work as they go.

It remains to be seen whether this approach can entirely compensate for the slowing progress of traditional training. But for now, the smartest AI models of the near future will not necessarily be the biggest ones—they will be the ones that know how to take their time.