Open-source reasoning models challenge proprietary AI dominance

The race to build ever-larger artificial intelligence models is hitting a wall of diminishing returns, forcing the industry into a quiet but fundamental shift. Instead of simply building bigger systems that guess the next word faster, AI labs are focusing on "inference-time compute"—giving models the time and processing power to "think" step-by-step before they answer.

This shift has quickly dismantled the competitive moats of Silicon Valley’s wealthiest players. The release of highly capable open-source reasoning models, most notably DeepSeek-R1 from a Chinese research lab, has proved that frontier-level reasoning does not require a multi-billion-dollar supercomputer or a blank check from Microsoft or Google.

What has changed

For years, the formula for better AI was simple: throw more data and more chips at the training phase. But pre-training web-scale data is yielding smaller improvements. The new approach focuses on what happens after training.

Reasoning models use reinforcement learning to generate an internal "chain of thought" before presenting a final answer. They test different approaches, catch their own mistakes, and refine their logic. On difficult benchmarks like the American Invitational Mathematics Examination (AIME) and the GPQA graduate-level science assessment, these systems are scoring alongside top human academics.

What surprised the industry was not just that these models could reason, but how cheaply they could be made. DeepSeek reported that its R1 model was trained for a fraction of the cost of comparable Western models. While the exact details of its training run have not been independently verified, the code and model weights are freely available for anyone to download and run.

Why the economics of AI are shifting

This represents a direct challenge to the business models of proprietary AI companies. Startups like OpenAI and Anthropic have justified their massive valuations on the assumption that proprietary, closed-source models would always stay several steps ahead of the open-source community.

If a developer can run a model that performs comparably to OpenAI's o1 on their own servers for a fraction of the API cost, the premium for closed-source models begins to evaporate.

This changes the calculations for businesses looking to deploy AI agents—autonomous systems designed to handle multi-step tasks like software engineering or research. These agents require thousands of model calls to complete a single task. On a proprietary API, that cost is prohibitive. On cheap, open-source reasoning models, it becomes economically viable.

What comes next

The battle ground has now moved from training clusters to the hardware running these models in production. Because reasoning models spend more time processing each request, they require different chip configurations and software optimisation to keep latency manageable for users.

There is also the unresolved question of safety. When a model's internal reasoning process is hidden or highly complex, predicting how it will behave in edge cases becomes harder. Security researchers have already pointed out that reasoning models can become better at deceiving human evaluators if they learn that lying is the most efficient path to a correct benchmark score.

As proprietary labs prepare their next major releases, they no longer just compete with each other. They are competing with a global open-source ecosystem that can copy, adapt, and cheapen their breakthroughs almost as fast as they can write the press releases.