The shift to reasoning models is changing the rules of the AI race

What is the new reasoning paradigm?

For years, the formula for building better artificial intelligence was simple: gather more data, plug in more chips, and train a larger model. But as the supply of high-quality human text runs dry and the financial cost of training balloons, the industry is shifting its focus. Instead of just building bigger brains, researchers are teaching models to spend more time processing a prompt before they reply—a concept known as "test-time compute".

This approach, popularised by OpenAI’s o1 and o3 models and followed rapidly by open-source alternatives like DeepSeek’s R1, allows a model to generate an internal "chain of thought" before delivering its final answer. It is the digital equivalent of a human pausing to draft a solution on scratch paper before speaking.

How the competition is shifting

Until recently, the consensus in Silicon Valley was that frontier AI development would remain a closed club, restricted to a handful of American giants capable of spending billions of dollars on compute clusters.

That assumption was shattered by the release of DeepSeek's R1 model. Its creators claimed they achieved performance comparable to Western proprietary models in mathematics and coding benchmarks while spending only a fraction of the training budget—reportedly around $6 million. While competitors and researchers are still dissecting how much of this efficiency came from hardware optimisation versus clever data distillation, the release proved that reasoning capabilities are no longer the exclusive domain of trillion-dollar tech giants.

Why test-time compute matters

The shift to reasoning models changes how we measure AI capability. Traditionally, a model's intelligence was fixed after its training phase. Now, capability can be scaled dynamically at the moment of use. If a user asks a simple question, the model responds instantly. If tasked with a complex software engineering problem or a novel scientific hypothesis, it can "think" for minutes, running through thousands of internal reasoning steps to self-correct its errors before presenting a solution.

This change is forcing a reallocation of resources. Tech companies are pivoting their infrastructure plans to support heavy inference workloads, where the heavy lifting happens when a user asks a question, rather than solely during the initial training run.

What happens next

The ultimate destination for this technology is the creation of reliable AI agents. Current models struggle with multi-step tasks because a single error early in the process derails the entire output. By integrating reasoning chains, future agents will be able to plan, test their own code, recognise when they have made a mistake, and try a different approach without human intervention.

However, significant hurdles remain. Reasoning models are currently slow, expensive to run at scale, and prone to "overthinking" simple queries. Industry researchers are now focused on making these reasoning steps invisible to the end user and reducing the latency of the thought process, aiming to bring human-like deliberation down to conversational speeds.