DeepSeek

DeepSeek disrupted the AI industry's assumptions at the start of 2025 by releasing frontier-quality open-source models at a fraction of the cost Western AI companies were charging. The Chinese AI lab released DeepSeek V3 and R1 as open-weight models under MIT licences, achieved performance competitive with GPT-4o and Claude on major benchmarks, and offered API access at 50-90% lower prices than comparable Western models. The market reaction affected semiconductor stock prices and forced the industry to rethink training cost assumptions.

The consumer application is a free chat interface similar to ChatGPT providing access to DeepSeek's models without any subscription. It became one of the most downloaded AI apps in multiple markets.

The reasoning models, R1 and successors, are particularly notable. DeepSeek's approach produces visible step-by-step thinking, making it easier to understand how conclusions were reached and identify errors. For mathematical reasoning, coding, and complex analytical tasks, reasoning models perform comparably to OpenAI's o1 and o3 models at significantly lower cost.

For developers evaluating AI infrastructure costs, DeepSeek's API pricing represents a compelling alternative. Organisations spending hundreds of thousands annually on OpenAI or Anthropic APIs have reported 50-90% cost reductions by switching to comparable DeepSeek models.

The central limitation is data sovereignty. DeepSeek is subject to Chinese data laws, which differ significantly from GDPR and US frameworks. For European and US organisations with data governance requirements, routing production data through DeepSeek's hosted API is complicated. The open-source models can be self-hosted, removing this concern, but that requires GPU infrastructure.