DeepSeek V4: How China’s Free AI Is Beating America’s $100 Billion Industry

DeepSeek V4, released by Chinese AI lab DeepSeek in mid-February 2025, is specifically targeting coding superiority over ChatGPT and Claude. The model was trained for approximately $6 million — 68 times cheaper than comparable American models that cost hundreds of millions to train. It is open source and free to use. And it is performing at or above the level of the most advanced American AI systems on key benchmarks. DeepSeek already commands 89% of China’s AI market share. Now it is going global.

This is not a story about one AI model. It is a story about a fundamental challenge to America’s assumption that massive capital investment is the only path to frontier AI. When DeepSeek’s earlier model, R1, was released in January 2025, it triggered a roughly $1 trillion selloff in U.S. tech stocks in a single day. Wall Street understood something: the moat protecting American AI companies may be smaller than anyone thought.

What DeepSeek V4 Actually Is

DeepSeek V4 is a large language model (LLM) — the same category of AI as GPT-4o and Claude 3.5 Sonnet. It was built by DeepSeek, a Chinese AI research company founded by Liang Wenfeng, who also co-founded the quantitative hedge fund High-Flyer Capital Management. The company is based in Hangzhou, China.

V4’s primary design focus is coding performance. DeepSeek’s benchmarks show it outperforming GPT-4o and Claude 3.5 Sonnet on HumanEval and SWE-bench — two standard tests of AI coding ability. It achieves this partly through architectural innovations described in DeepSeek’s published research on arXiv, including a technique called Manifold-Constrained Hyper-Connections, which improves how information flows between the model’s layers during training and inference.

Like its predecessor R1, V4 is released as open weights — the model parameters are publicly available for anyone to download, run locally, or fine-tune. This is a deliberate strategic choice, and understanding why China is giving away cutting-edge AI for free is the most important part of this story (Grokipedia: DeepSeek AI Overview).

The $6 Million Training Cost: What It Actually Means

The $6 million training figure for DeepSeek R1 (the figure cited for V4’s predecessor, which V4 builds on) caused significant controversy when it was first reported. American AI companies have spent hundreds of millions — in some cases over a billion dollars — training their flagship models. How can DeepSeek do it for $6 million?

Several factors explain the gap. First, DeepSeek uses a Mixture of Experts (MoE) architecture, which activates only a fraction of the model’s parameters for any given input rather than running the entire model on every token. This dramatically reduces the compute needed per training step. Second, DeepSeek has published novel training efficiency techniques that reduce redundant computation. Third, the $6 million figure refers to the final training run cost on owned hardware — it does not include the years of research, failed experiments, and infrastructure that led to that run.

Even with those caveats, the efficiency gap is real. DeepSeek has squeezed dramatically more performance per dollar out of its hardware than its American counterparts. This is partly a result of necessity: U.S. export controls restrict China’s access to the most advanced NVIDIA chips (H100s and A100s). DeepSeek has had to optimize heavily for older, less powerful hardware, and that constraint has driven innovation (DeepSeek-R1 arXiv paper).

Why China Is Giving Away Its Best AI for Free

This is the question American AI executives are most worried about, and it deserves a clear answer. DeepSeek’s open source strategy is not generosity — it is geopolitical and economic warfare by other means.

When DeepSeek releases a model that is free, open, and competitive with ChatGPT, it does several things simultaneously. It commoditizes the American AI stack: if a free Chinese model performs as well as a $20/month American subscription, the American product’s pricing power collapses. It increases global dependence on Chinese AI infrastructure: as developers worldwide build on DeepSeek’s open models, they integrate Chinese technology into their products, regardless of where they are based. And it undermines the investment thesis for American AI companies: OpenAI’s $157 billion valuation is built on the assumption that its model advantage justifies premium pricing. DeepSeek challenges that assumption directly.

Liang Wenfeng has framed DeepSeek’s mission in terms of Chinese technological sovereignty and global AI democratization. Both framings serve the same strategic end: making DeepSeek the infrastructure layer of global AI, the way TSMC became the infrastructure layer of global chip manufacturing.

For more context on open source AI dynamics, see our open source AI guide and our deep dive on the complete DeepSeek story.

The R1 Market Shock: A $1 Trillion Warning

When DeepSeek released its R1 reasoning model in late January 2025, the American stock market reacted as if it had been punched. NVIDIA lost approximately $600 billion in market capitalization in a single trading session — the largest single-day loss in market cap for any company in history at that point. Broader tech sector losses brought the total to roughly $1 trillion.

The market logic was straightforward: if AI training can be done efficiently on cheaper hardware, the massive infrastructure buildout that companies like NVIDIA, Microsoft, and Google have been driving may be partially unnecessary. The “GPU scarcity” narrative — which had justified enormous capital expenditure and sky-high AI company valuations — suddenly looked uncertain.

NVIDIA’s CEO Jensen Huang called DeepSeek’s work “a perfect example of scaling law being applied in a very thoughtful way” — a notably measured response to a development that cost his company hundreds of billions in market value in one day. Other American AI leaders were less diplomatic in private. The consensus emerging from the episode is that the AI efficiency frontier is more contested than the American industry had assumed.

What This Means for the AI Industry and You

For developers and businesses, DeepSeek V4 is simply a very capable free model that you can run locally or via API. If you need a coding assistant, it is worth evaluating alongside ChatGPT and Claude on your specific use cases. The open weights mean you can run it on your own infrastructure without sending data to any third-party API — a privacy advantage for sensitive enterprise use.

For the broader AI industry, DeepSeek’s success confirms that the era of a few American companies holding an unchallenged lead in AI capability is over. The efficiency innovations DeepSeek has published are already being studied and replicated by researchers worldwide. Training cost curves are falling faster than the industry’s investment narratives assumed.

For geopolitics, the DeepSeek story confirms that U.S. export controls on AI chips have not stopped China from advancing in AI — they have instead pushed Chinese researchers to innovate around the constraint in ways that may ultimately benefit the global AI ecosystem, including China’s competitors. The effectiveness of tech export controls as a tool of AI strategy is now a serious policy debate.

To understand where AI as a whole is going, see our complete history of AI and our guide to what artificial intelligence is.

Key Takeaways

• In one sentence: DeepSeek V4 is a free, open-source Chinese AI model trained for $6 million that performs at or above American models costing 68 times more to build.
• Key number: $1 trillion — the approximate market value erased from U.S. tech stocks when DeepSeek R1 launched in January 2025.
• Why it matters: DeepSeek proves the compute-intensive American AI model is not the only path to frontier AI, threatening the pricing power and valuations of American AI companies.
• What to do: Evaluate DeepSeek V4 for coding tasks in your workflow — it is free, open, and competitive with premium paid alternatives.
• Related: Complete DeepSeek story | Open source AI guide | History of AI

Frequently Asked Questions

Is DeepSeek V4 actually better than ChatGPT?

On coding benchmarks, DeepSeek V4 performs at or above GPT-4o and Claude 3.5 Sonnet. On general reasoning and conversational tasks, results are more mixed. The honest answer is: it depends on the task. For coding specifically, it is highly competitive and worth testing.

Is it safe to use DeepSeek given it is a Chinese company?

This is a legitimate question. The DeepSeek API sends data to Chinese servers, which raises data privacy concerns for sensitive enterprise use. However, DeepSeek’s open-weights models can be downloaded and run entirely on your own infrastructure — no data ever leaves your environment. For sensitive use cases, local deployment addresses the privacy concern.

What are Manifold-Constrained Hyper-Connections?

This is a DeepSeek architectural innovation described in their published arXiv paper. In plain terms, it is a technique that improves how information flows between layers of the neural network during training, allowing the model to learn more efficiently from each training example. It is one of several innovations that helped DeepSeek achieve high performance at low cost.

Why did DeepSeek trigger a trillion-dollar stock selloff?

Investors had built enormous valuations around the premise that AI required massive, ongoing investment in expensive chips and infrastructure. DeepSeek demonstrated that competitive AI could be built far more cheaply, which threatened the revenue projections for chip makers like NVIDIA and the investment rationale for data center buildouts by Microsoft, Google, and Amazon.

Who is Liang Wenfeng?

Liang Wenfeng is the founder of DeepSeek and co-founder of High-Flyer Capital, one of China’s largest quantitative hedge funds. He has an academic background in AI and has used High-Flyer’s substantial capital to fund DeepSeek’s research. He is low-profile by tech founder standards — few interviews, no major conference appearances — which makes DeepSeek’s rapid ascent even more striking.

This story was originally covered in our daily newsletter, Beginners in AI. Subscribe for free AI news delivered to your inbox every morning.

Get free AI tips delivered daily → Subscribe to Beginners in AI

You May Also Like

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