DeepSeek R1 reasoning model challenges OpenAI with low-cost open-source AI

DeepSeek, an artificial intelligence company based in Hangzhou, has released DeepSeek-R1, an open-source reasoning model that matches OpenAI’s proprietary o1 on key technical benchmarks. The launch has triggered a sharp reassessment of the AI industry's competitive dynamic, proving that state-of-the-art logical reasoning no longer requires massive Silicon Valley budgets or unrestricted access to the latest chips.

For months, the leading frontier labs operated under the assumption that building smarter AI required exponential increases in compute power and capital. DeepSeek-R1 challenges this thesis. It matches or exceeds US models in mathematics, coding, and logical reasoning, but its developers claim to have trained it for a fraction of the cost of its American counterparts.

How DeepSeek R1 works

Traditional large language models generate answers by predicting the next most likely word instantly. This works well for creative writing or summarisation, but fails at complex problem-solving.

Like OpenAI's o1, DeepSeek-R1 uses a technique called "test-time compute" or "inference-time scaling". Before outputting an answer, the model generates an internal chain of thought, evaluating different approaches, correcting its own errors, and reasoning through the problem step-by-step.

What makes R1 unique is how it was trained. DeepSeek first created a model called DeepSeek-R1-Zero, which used pure reinforcement learning without any initial human instruction data. The model discovered reasoning behaviours, such as self-correction and double-checking, entirely on its own. DeepSeek then refined this process for R1 by adding a small amount of high-quality human data to make the outputs more readable and stable.

The economics of the AI race

According to DeepSeek's technical report, the training of its base model, DeepSeek-V3, on which R1 is built, cost roughly $5.6 million in chip rental fees. By comparison, industry estimates put the training costs of models like OpenAI's GPT-4 or Anthropic's Claude 3 at well over $100 million, with next-generation training runs expected to cross the billion-dollar threshold.

DeepSeek achieved these savings through algorithmic efficiency. The model uses a Mixture-of-Experts architecture, meaning only a fraction of its total parameters are activated for any given query. It also employs a novel mathematical approach to attention mechanisms, reducing the memory footprint of the model during both training and inference.

Crucially, DeepSeek trained its models using Nvidia H800 graphics processors—older, slower chips designed to comply with US export restrictions on China. This suggests that software-level efficiency can bypass hardware limitations that Western policymakers expected would hold Chinese AI back.

What this means for Silicon Valley

The release has sent shockwaves through the technology sector. Because DeepSeek has open-sourced R1 under an MIT licence, developers worldwide can inspect, run, and modify the code for free. This lowers the barrier to entry for building complex AI agents and reasoning applications.

It also places immense pressure on proprietary software companies. If developers can access frontier-level reasoning capabilities for free, or run them locally on their own servers, the premium subscription models of OpenAI, Google, and Microsoft become much harder to justify.

While the US government has focused its strategy on restricting hardware access, DeepSeek-R1 suggests the defining factor in the next phase of AI development may not be who has the most chips, but who writes the most efficient algorithms.