DeepSeek: The Chinese AI That Shook the Industry

What Is DeepSeek?

DeepSeek is a Chinese AI research laboratory founded in 2023 by Liang Wenfeng, also the co-founder of quantitative hedge fund High-Flyer Capital Management. In the span of roughly 18 months, DeepSeek went from a relatively obscure research outfit to an organization that shook the global AI industry with a model release that simultaneously challenged the assumption that frontier AI required billions of dollars in compute, and demonstrated that open-source models could match the performance of GPT-4.

The moment that crystallized DeepSeek’s global impact was the January 2025 release of DeepSeek-R1 and the simultaneous announcement of the training costs for DeepSeek-V3 — approximately $6 million USD, compared to the hundreds of millions estimated for comparable models from OpenAI and Anthropic. That disclosure sent Nvidia’s stock down 17% in a single day, erased roughly $600 billion in market capitalization from AI-adjacent companies, and triggered an urgent reassessment across Silicon Valley of whether the “capital-intensive compute race” model of AI development was as defensible as assumed.

Understanding DeepSeek requires understanding both its technical achievements and the geopolitical context. DeepSeek operates under US semiconductor export controls, meaning it cannot legally access Nvidia’s most powerful chips (H100s). The fact that it produced frontier-capable models under these constraints is either proof of extraordinary engineering ingenuity — or, in the more concerning interpretation circulating in Western security discussions, evidence that the export controls have less effect than assumed.

DeepSeek-R1: The Model That Changed Everything

DeepSeek-R1 is DeepSeek’s reasoning-focused model, released in January 2025 alongside its technical report and full model weights. It is a 671-billion parameter Mixture of Experts model that achieves performance comparable to OpenAI’s o1 reasoning model on benchmarks including AIME 2024 (mathematics olympiad problems), MATH-500, LiveCodeBench (competitive programming), and MMLU.

What made R1 technically distinctive was how DeepSeek achieved this reasoning capability. Rather than relying primarily on supervised fine-tuning on human demonstrations of reasoning chains, DeepSeek used a reinforcement learning approach called Group Relative Policy Optimization (GRPO) that taught the model to reason effectively through trial, error, and reward signals. The model learned to produce long, structured chain-of-thought reasoning traces that walk through complex problems step by step — similar to OpenAI’s o1, but developed independently and published openly.

The model weights for R1 were released under the MIT license — one of the most permissive open-source licenses available, with virtually no restrictions on commercial use, modification, or distribution. This stands in sharp contrast to Meta’s Llama license (which has commercial use restrictions for very large companies) and is genuinely remarkable for a frontier-class reasoning model. The release of R1 weights effectively gave every developer in the world access to a GPT-4-class reasoning model at zero cost beyond infrastructure.

DeepSeek-V3: The Cost-Efficiency Story

DeepSeek-V3 is DeepSeek’s general-purpose conversational and instruction-following model, also released in late 2024. Its technical report revealed the training run cost approximately $5.576 million using 2,048 Nvidia H800 GPUs over 55 days. To understand why this is remarkable: industry estimates for training comparable models like GPT-4 (OpenAI) or Claude 2 (Anthropic, the predecessor to today’s Claude Opus 4.7) ranged from $50 million to $100 million. DeepSeek’s reported cost efficiency of roughly 10-20x better than competitors was extraordinary.

DeepSeek achieved this through a combination of architectural innovations: Multi-head Latent Attention (MLA) that reduces the KV cache memory footprint substantially, a MoE architecture that keeps only 37 billion parameters active per token despite a 671 billion total parameter count, and an efficient FP8 mixed-precision training pipeline. The sum of these optimizations is a model that runs faster, trains cheaper, and serves inferences more economically than competing architectures.

The cost efficiency story matters beyond the headline number. If models of this capability can be trained for single-digit millions rather than hundreds of millions, it dramatically lowers the barrier to entry for AI development — for countries, universities, startups, and researchers without access to massive capital. It challenges the incumbent advantage that well-funded Western AI labs have accumulated through sheer compute spending. For context on the historical significance of this, see our History of AI.

How to Access and Use DeepSeek

The simplest way to try DeepSeek is through chat.deepseek.com — a web interface similar to ChatGPT that’s free to use and gives you access to both DeepSeek-V3 and DeepSeek-R1. The chat interface is clean and functional; the notable UI difference with R1 is a visible thinking trace panel that shows the model’s chain-of-thought reasoning before it delivers its final answer, which is illuminating both for understanding what the model is doing and for debugging complex prompts.

DeepSeek also offers an API at extremely competitive pricing — as of early 2025, significantly cheaper per token than GPT-4o for comparable capability. However, the API is operated from China, which raises data privacy considerations for business use: data sent to DeepSeek’s API passes through servers in China and is subject to Chinese data law. This is a meaningful factor for enterprises handling sensitive or regulated data, and many Western companies have chosen to self-host the open-source weights rather than use the API for exactly this reason.

Self-hosting is the privacy-safe route. DeepSeek’s weights are available on Hugging Face, and the model can be deployed on cloud infrastructure through AWS, Azure, or Google Cloud, or run locally on high-end hardware for the smaller distilled variants. Our explainer on Hugging Face Explained covers how to access DeepSeek models through these channels. For understanding the concepts behind these models, our guide on AI Tokens Explained is helpful.

DeepSeek vs. GPT-4o vs. Claude 3.7

On reasoning benchmarks specifically — complex mathematics, logic puzzles, competitive programming — DeepSeek-R1 performs at the level of OpenAI’s o1, which was the reasoning frontier model when R1 was released. On general capability benchmarks (MMLU, coding, instruction following), DeepSeek-V3 is competitive with GPT-4o and Claude 3.7 Sonnet for most tasks, with the occasional gap in instruction following on nuanced creative tasks.

The practical differences for users come down to: (1) availability and reliability — DeepSeek’s API has experienced outages during peak demand given the enormous global interest after launch; (2) content restrictions — DeepSeek applies Chinese regulatory content filters that result in refusals on politically sensitive topics related to China; (3) cost — DeepSeek’s API is significantly cheaper than GPT-4o; (4) data privacy — as noted above, using the hosted API means data goes to Chinese servers.

For users who self-host the weights, options (1), (2), and (4) largely go away — you control the deployment, the filtering, and the data residency. This is why the open-source release of R1 under MIT license was so significant: it gave the global developer community the ability to deploy frontier reasoning capability on their own terms. See our broader AI tools comparison in ChatGPT vs Claude vs Gemini.

For a broader map of the open-source AI landscape that DeepSeek now sits at the frontier of, our Open Source AI Guide covers the key models, platforms, and considerations.

The Geopolitical Implications of DeepSeek

DeepSeek’s emergence has intensified already significant geopolitical tensions around AI development. The US government’s export controls on advanced semiconductors to China were predicated on the assumption that compute access was a meaningful bottleneck — that restricting China’s access to H100s would slow the development of capable AI systems. DeepSeek’s results, achieved on H800s (the export-control-compliant but lower-spec alternative), challenged that assumption directly.

The response in Washington was swift. Congressional hearings were called. The Biden and then Trump administrations reviewed and tightened export control frameworks. The Department of Commerce opened an investigation into whether DeepSeek had circumvented controls through third-party chip access. The US Navy issued guidance prohibiting use of DeepSeek on government devices. Several Western nations followed with their own advisories.

From a purely technological standpoint, DeepSeek represents a healthy development — more competition, more open research, more efficient models that benefit developers everywhere. From a geopolitical standpoint, it raised legitimate questions about technology transfer, AI alignment across different regulatory and value systems, and whether the assumption that frontier AI was a Western domain would hold as the decade progresses. The full implications will take years to resolve, but DeepSeek firmly ended the comfortable assumption that AI’s frontier was a closed club.

The DeepSeek Controversy and What It Means

DeepSeek’s emergence wasn’t just a technical milestone — it triggered a genuine geopolitical and economic shock wave that’s worth understanding, because it reshaped how the entire AI industry thinks about the relationship between investment, compute, and capability.

The cost question. DeepSeek reportedly trained its R1 model for approximately $5.6 million in compute costs — a fraction of what OpenAI, Google, and Anthropic spend on their frontier models (estimated at $100 million to $1 billion+). This claim, if accurate, challenged a core assumption of the AI industry: that building competitive models requires massive capital expenditure on training compute. The implication was uncomfortable for companies that had raised billions specifically on the premise that more money equals better AI. When the news broke in January 2025, NVIDIA’s stock dropped over 15% in a single day, wiping out roughly $600 billion in market capitalization — one of the largest single-day value drops in stock market history.

The export control question. DeepSeek built R1 using NVIDIA H800 GPUs — chips specifically designed as export-compliant alternatives after the US government restricted sales of the more powerful A100 and H100 chips to China. The fact that a Chinese lab achieved frontier-class results using the restricted chips raised serious questions about whether export controls were achieving their intended goal of maintaining US AI superiority, or simply forcing Chinese companies to innovate around constraints — potentially making them more efficient in the long run.

The open-source question. DeepSeek released R1’s model weights openly, allowing anyone to download, modify, and deploy the model. This was a strategic choice: by making R1 open-source, DeepSeek ensured rapid global adoption, attracted developer mindshare, and made it essentially impossible to restrict through policy measures. Within weeks of release, R1 had been fine-tuned for dozens of specialized applications by developers worldwide, integrated into major open-source frameworks, and deployed by companies that would never have used a Chinese-hosted API. The open-source strategy transformed a potential national security concern into a globally distributed technology that no single government can control.

Using DeepSeek in practice. For individual users, DeepSeek R1 is available through the DeepSeek chat interface (chat.deepseek.com), through API access at extremely competitive pricing, or as a downloadable model you can run locally through platforms like Ollama and LM Studio. R1’s reasoning capabilities are particularly strong in mathematics, coding, and scientific analysis — tasks where the chain-of-thought approach excels. For everyday writing and conversation, Claude and ChatGPT generally produce more polished output. The practical recommendation for most users is to keep DeepSeek as a specialized tool for technical reasoning tasks while using other models for general-purpose work.

Frequently Asked Questions

Is DeepSeek safe to use?

For personal use and non-sensitive queries, DeepSeek is technically safe to use through its chat interface. For business use involving sensitive, regulated, or proprietary data, the recommendation is to self-host the open-source weights on your own infrastructure rather than sending data to DeepSeek’s API. Several government bodies and enterprises have restricted use of the hosted API on data privacy grounds; the open-source model weights themselves are widely used with no such restrictions.

What is DeepSeek-R1’s reasoning capability?

DeepSeek-R1 is specifically optimized for complex reasoning tasks: advanced mathematics, logical deduction, competitive programming, and multi-step problem solving. It uses a reinforcement learning training approach that taught it to produce detailed chain-of-thought reasoning before arriving at an answer. On standardized benchmarks like AIME 2024 and MATH-500, R1 performs comparably to OpenAI’s o1 model, which was the leading reasoning model at the time of R1’s release.

Why did DeepSeek cause Nvidia’s stock to drop?

DeepSeek’s V3 training cost report ($5.576 million) challenged the assumption that frontier AI development requires enormous compute expenditure — and therefore enormous purchases of Nvidia’s high-end AI chips. If capable models can be trained cheaply, the demand growth for expensive AI accelerator chips might be lower than projected. This reassessment of the long-term demand trajectory drove Nvidia’s stock down approximately 17% on January 27, 2025, erasing about $600 billion in market cap — one of the largest single-day losses in stock market history for a single company.

Can DeepSeek be used commercially?

DeepSeek-R1 and DeepSeek-V3’s weights are released under the MIT license, which is one of the most permissive open-source licenses available. Commercial use is permitted without restriction — you can build products on top of DeepSeek, fine-tune it, redistribute it, and use it in commercial applications. The hosted API at api.deepseek.com also permits commercial use under DeepSeek’s API terms of service, subject to standard content policy restrictions.

How does DeepSeek’s MoE architecture work?

DeepSeek-V3 and DeepSeek-R1 use Mixture of Experts (MoE) architecture. Despite having 671 billion total parameters, only 37 billion parameters are active at any given moment during inference — the model routes each token through a subset of specialized “expert” layers rather than activating all parameters for every computation. This architecture delivers the capacity of a much larger model while maintaining the inference cost of a smaller one, which is a core reason for DeepSeek’s reported cost efficiency.

DeepSeek R1: The Reasoning Model That Changed Everything

The January 2025 release of DeepSeek R1 sent shockwaves through the AI industry. Here is what makes it technically significant, and how to use it effectively.

How R1 Reasoning Actually Works

DeepSeek R1 uses a technique called chain-of-thought reinforcement learning. Before producing an answer, the model generates a long internal reasoning trace (visible to the user as a “thinking” block) where it works through the problem step by step — checking its logic, identifying mistakes, and revising its approach. This is similar to OpenAI’s o1 and o3 models, but R1 was developed at a fraction of the cost: DeepSeek reports training R1 for approximately $6 million, compared to estimates of $100M+ for comparable OpenAI models. The open-source release of R1’s weights allows anyone to verify these claims and run the model independently.

Benchmark Performance

• AIME 2024 (math olympiad problems): R1 scores 79.8%, compared to OpenAI o1’s 79.2%. A near-tie with the then-leading model.
• Codeforces programming competition: R1 achieves a 2,029 Elo rating, placing it in the top 96.3% of human competitive programmers.
• MMLU (general knowledge): 90.8%, comparable to the best closed-source models.
• SWE-bench (real-world software engineering tasks): 49.2% resolution rate — one of the best results on this benchmark at time of release.

The Controversy Explained

DeepSeek’s release triggered several controversies worth understanding:

• Export controls and hardware: DeepSeek trained R1 on NVIDIA H800 chips (a lower-spec variant of H100 sold to China under US export controls). This raised questions about whether export controls can actually slow Chinese AI development — a policy debate with significant geopolitical implications.
• Safety and censorship: R1’s default deployment refuses to answer questions about Tiananmen Square, Taiwan independence, and other politically sensitive topics in China. Researchers running the open weights locally without DeepSeek’s system prompt do not encounter these restrictions.
• Data origin questions: Several researchers published findings suggesting DeepSeek’s earlier V3 model showed patterns consistent with training on distilled outputs from OpenAI models — a potential violation of OpenAI’s terms of service. DeepSeek denied these claims.
• Stock market impact: Nvidia’s stock fell 17% in a single day on the day R1’s capabilities became widely known, as investors processed the implication that highly capable AI models might require far less expensive hardware than previously assumed.

How to Use DeepSeek R1 Today

• DeepSeek website (chat.deepseek.com): Free to use with a DeepSeek account. Toggle “Deep Think” mode to enable R1 reasoning. Best for math, coding, and complex analysis tasks.
• API access: DeepSeek’s API charges approximately $0.55 per million input tokens for R1 (with cache hits as low as $0.14) — dramatically cheaper than OpenAI o1 at $15/million tokens.
• Via third-party platforms: Perplexity, Poe, and Openrouter all offer R1 access, sometimes with free tiers.
• Running locally: Download the distilled versions (R1-Distill-Qwen-7B or R1-Distill-Llama-8B) via Ollama with ollama pull deepseek-r1:7b. The 7B distilled version runs on a laptop with 8GB RAM and still shows strong reasoning behavior.

You May Also Like

• What Is Artificial Intelligence
• Best AI Tools for Beginners
• How to Use AI
• AI Tools Directory
• Best Free AI Courses