Reasoning models rewrite the economics of frontier AI

The pivot to thinking in loops

The race for raw scale in artificial intelligence has hit a wall of diminishing returns, forcing the industry to pivot toward a new frontier: teaching models how to think before they respond. This transition, marked by the arrival of reasoning models like OpenAI's o1 and the open-source DeepSeek-R1, is rewriting the competitive rules of Silicon Valley.

For years, the formula for better AI was simple: throw more data and more chips at the training phase. But pre-training large language models on the entire public internet has yielded flatter performance curves. The new consensus, shared by top labs from San Francisco to Beijing, is that the next leap in capability will come from test-time compute—allocating more processing power while the model is answering a question, rather than just when it is being built.

These reasoning models do not just blurt out the most statistically likely next word. Instead, they use reinforcement learning to generate internal chains of thought, testing hypotheses, spotting their own mistakes, and correcting course before presenting a final answer.

The open-source disruption

This architectural shift has unexpectedly levelled the playing field. When China-based DeepSeek released its R1 model, it claimed to match the reasoning performance of proprietary American models at a fraction of the estimated training cost. While Western tech companies have historically spent hundreds of millions of dollars on training clusters, DeepSeek claimed its model cost roughly $6 million to train.

Industry analysts have spent weeks debating the exact economics of these claims, and independent verification remains difficult. However, the release has forced a reckoning. By open-sourcing the weights and the training methodology, the project proved that advanced reasoning is not a moat restricted to companies with infinite capital.

The implications for developers are immediate. Instead of renting access to expensive, closed APIs, software engineers can run highly capable reasoning models on their own hardware or cheaper cloud instances. This has triggered a rush to integrate these models into complex, multi-step workflows where accuracy matters far more than speed.

What happens next

As reasoning becomes cheaper and more accessible, the focus of AI development is shifting from chatbots to autonomous agents. Traditional models often struggle with multi-step tasks because a single error early in the process derails the entire output. A model that can pause, evaluate its progress, and self-correct is far better suited to writing complex software, conducting scientific research, or managing database migrations.

However, this new paradigm introduces its own bottlenecks. Running reasoning models is slow and computationally expensive at the point of use. While a standard system responds almost instantly, a reasoning model might pause for thirty seconds or more while it runs through thousands of internal "thinking tokens."

For the industry, the challenge is no longer just about building the biggest supercomputer. It is about optimising inference—making these internal reasoning loops fast enough, and cheap enough, for everyday applications. The companies that win the next phase of the AI race will not necessarily be those with the largest datasets, but those that can make thinking cheap.