What is DeepSeek and where did it come from?
DeepSeek is an AI lab based in China, founded in 2023 and funded by the quantitative trading firm High-Flyer. Unlike most Chinese tech companies, DeepSeek focused almost entirely on frontier model research rather than commercial products — publishing extensively in academic venues and releasing model weights openly.
In late 2024 they released DeepSeek V3 — a 671 billion parameter model using Mixture of Experts architecture that performed competitively with GPT-4o. Impressive, but not shocking. Then in January 2025 they released DeepSeek R1.
What R1 did that shocked everyone
DeepSeek R1 matched or exceeded OpenAI's o1 on the most rigorous reasoning benchmarks — AIME (competition mathematics), MATH-500, and coding benchmarks including Codeforces. o1 had been OpenAI's flagship reasoning model, widely considered the frontier of what was possible.
The market reaction
NVIDIA's stock fell 17% in a single day — its largest single-day market cap loss in history at the time — as investors questioned whether the AI industry's insatiable demand for expensive GPUs was as certain as assumed. If frontier AI could be achieved at a fraction of the compute cost, demand for H100s and H200s might be lower than projected.
The broader reaction across the AI industry was a mixture of panic, scepticism, and rapid reassessment. OpenAI, Anthropic, and Google all moved to cut API pricing within months. The competitive dynamic had fundamentally shifted.
What it actually proved — and what it didn't
What it proved: That the techniques OpenAI developed for reasoning models — specifically reinforcement learning applied to chain-of-thought reasoning — could be replicated and applied efficiently with less compute than previously thought. DeepSeek's key contribution was demonstrating that the training efficiency gap between frontier labs was smaller than the cost gap implied.
What it didn't prove: That training frontier models is cheap in absolute terms. $6M is still a significant amount. DeepSeek also benefits from earlier work by other labs — they built on published research from OpenAI, Google, and academic institutions. And US export controls on high-end chips mean DeepSeek's future trajectory faces real constraints.
R1 was a genuine and significant achievement that forced a recalibration of assumptions about the compute-capability relationship. It did not mean AI is now cheap, nor that Chinese labs have "caught up" across the board. It meant the efficiency gap was smaller than the cost gap implied — which is still a substantial finding.
Why it still matters in 2026
DeepSeek R1's open weights are freely available and have been downloaded millions of times. They power countless self-hosted deployments where organisations run near-frontier reasoning capability on their own infrastructure at zero marginal cost. The model itself has had a lasting democratising effect regardless of the corporate or geopolitical dynamics around its release.
DeepSeek V3 — the base model underneath — remains one of the most capable open-weight models available and is widely used via third-party APIs at dramatically lower cost than OpenAI or Anthropic equivalents.
The data sovereignty question
For enterprise use, the primary concern with DeepSeek's API is data sovereignty — queries sent to DeepSeek's servers are processed under Chinese data law. For most consumers and many businesses this is not a concern. For organisations handling sensitive data in regulated industries — legal, healthcare, financial services, government — it is a meaningful consideration that typically points toward using open weights self-hosted rather than the DeepSeek API.