The shift to reasoning models: why AI labs are changing how they build intelligence

AI development has hit a sharp turn. For years, the recipe for building smarter artificial intelligence was simple: gather more data, plug in more microchips, and train a larger model. But as those traditional scaling laws begin to plateau, the world’s leading AI labs are pivoting to a different strategy. Instead of focusing entirely on how a model is trained, they are shifting computational power to the moment the model actually answers a question.

This approach, known as inference-time compute or "test-time compute", allows models to generate, check, and refine their own internal chain of thought before showing a final response to the user.

What is test-time compute?

In traditional large language models, generating a response is a straight line. The model predicts the next word, then the next, based on patterns it learned during training. It cannot pause to reconsider an error mid-sentence.

Reasoning models change this dynamic. When presented with a complex mathematics or coding problem, the model allocates extra processing power to run internal search algorithms. It tries different paths, recognises its own mistakes, and corrects them in a private scratchpad before delivering its output.

The technique was thrust into the spotlight by OpenAI with the release of its o1 and o3 models. More recently, the Chinese research firm DeepSeek adopted a similar approach with its R1 model. Rather than relying solely on massive, expensive pre-training runs, these models use reinforcement learning to reward the system for finding correct reasoning steps.

How the open-source world caught up

The AI industry was shaken when DeepSeek released the weights of its R1 model under an open-source licence. For the first time, researchers and developers worldwide could inspect, run, and modify a top-tier reasoning model on their own hardware.

Crucially, DeepSeek claimed it trained its underlying models for a fraction of the cost typically associated with Silicon Valley's frontier systems. While American tech giants have spent billions of dollars on clusters of Nvidia graphics processing units (GPUs), industry analysts estimate DeepSeek’s training costs were in the low millions of dollars.

While these exact training budgets have not been independently audited, the release has forced a reassessment of the barriers to entry in high-end AI development. It suggests that algorithmic efficiency, rather than brute-force capital expenditure, may dictate the next phase of the AI race.

What this means for the AI industry

This shift changes what we can expect from AI assistants and agents. Traditional models are notoriously poor at tasks requiring multi-step planning, such as writing complex software or verifying scientific proofs. By allowing a model to "think" for seconds or even minutes before responding, developers are beginning to build autonomous agents capable of executing multi-stage workflows without human intervention.

However, the transition also introduces new challenges. Giving a model extra time to process a query makes inference far more computationally expensive for the servers hosting it. It also raises questions about transparency: if a model's reasoning occurs in a hidden scratchpad, verifying why it reached a specific conclusion becomes much harder for developers and auditors alike.

For now, the era of simply building bigger models is giving way to an era of building smarter reasoning loops. How companies manage the costs of this real-time computation will likely decide who leads the market in the coming years.