In January 2025 the Chinese AI startup DeepSeek released its reasoning model DeepSeek-R1 — and within a week it had shaken the entire tech world. In performance, this Chinese AI model stands on a par with OpenAI’s ChatGPT, yet the development cost was radically different. While OpenAI’s spending runs into hundreds of millions of dollars, DeepSeek reportedly spent just $5.6 million on training. What is more, DeepSeek did not use Nvidia’s most advanced AI chips — US export controls bar them from being sold to China — but their cut-down versions. That raised an uncomfortable question: if a powerful, cheap AI can be built in China this way, are the astronomical budgets of American AI companies justified?
DeepSeek-R1 vs OpenAI at a Glance
Before going into the details, here is a side-by-side comparison of what made the DeepSeek launch so unusual compared with the established leader of the AI market.
| DeepSeek-R1 | OpenAI | |
|---|---|---|
| Reported training cost | About $5.6 million | Hundreds of millions of dollars |
| Hardware | Nvidia H800 — export-compliant, cut-down chips | Nvidia H100 — top-tier AI chips |
| Model access | Open weights and code, free to download and modify | Proprietary, API and subscription only |
| App Store (late January 2025) | No. 1 free app in the US and China | Overtaken by DeepSeek |
| Answer filtering | State censorship on politically sensitive topics | Content policy set by the company |
DeepSeek Crashes the Market: Nvidia Loses $600 Billion
The stock market reaction was immediate. Nvidia, the world’s biggest supplier of AI chips, saw its shares dive 17% in a single day, wiping out roughly $600 billion in market value — the largest one-day loss for any company in US stock market history. The Nasdaq 100, which comprises the largest technology companies, shed almost $1 trillion in one of the most significant tech crashes of recent years.
Venture capitalist, engineer and inventor Marc Andreessen called it AI’s “Sputnik moment”, comparing the event to the launch of the Soviet Union’s first satellite in 1957. Back then, the world realised that the USSR had overtaken the United States in the space race. Now China has made a similar leap in artificial intelligence, forcing the West to acknowledge that technological superiority is no longer its exclusive advantage.
The Secret Behind DeepSeek’s Success: Cheap Chips, Open Source and a Viral App
Three factors explain DeepSeek’s success. First, the company managed to cut the cost of training AI dramatically: while OpenAI spends hundreds of millions on developing its models, DeepSeek achieved a comparable level of performance for just $5.6 million. This was possible because the company used the more affordable Nvidia H800 chips instead of the high-end H100 processors its competitors rely on. Optimised algorithms then allowed DeepSeek to maintain high efficiency with far less computing power.
The second advantage was DeepSeek’s decision to make its artificial intelligence open source by publishing the model weights and code. That means anyone can download, adapt or enhance the technology — which is particularly attractive for businesses, researchers and startups that cannot afford proprietary AI at scale.
The third factor behind DeepSeek’s rapid rise was its instant and immense popularity among end users. Within days, the DeepSeek chatbot app became the most downloaded free app in the Apple App Store in both the US and China, overtaking ChatGPT.
Distillation Allegations and US Chip Sanctions
OpenAI and Microsoft launched an investigation, suspecting that DeepSeek used their technology to train its model. According to them, DeepSeek allegedly relied on knowledge distillation — a method in which one AI model imitates another to adopt its style of responses and reasoning. If proven, DeepSeek could face serious legal trouble.
Then there are the US sanctions on China. Beijing currently cannot purchase the most advanced chips from Nvidia, and tougher export controls could obstruct China’s efforts to develop the next generation of AI models. Yet the fact that DeepSeek has come this far shows how self-sufficient the Chinese tech industry has already become.
Despite the accusations, OpenAI CEO Sam Altman acknowledged that DeepSeek is a strong competitor and said the Chinese model was impressive, particularly given its relatively low cost.
What the DeepSeek Shock Means for the AI Industry
DeepSeek is likely to overturn the established view of what AI should be and how much it should cost. Until recently, it was assumed that building powerful large language models requires enormous investment and a highly complex infrastructure with thousands of state-of-the-art chips. The Chinese company has shown that comparable results are achievable with far fewer resources. This challenges the industry’s traditional practices and is forcing the major players to reassess their strategy.
This success may also shift the balance of power in the technological race. Not long ago, China was considered an outsider in AI; now it lags only a few years behind, if that. Feeling the threat of falling behind, US President Donald Trump announced the Stargate initiative — a programme worth up to $500 billion aimed at strengthening the position of American AI developers.
OpenAI has also begun cutting the prices of its services in an effort to retain users. If DeepSeek continues to release open-source AI solutions that compete with the market leaders at almost no cost, the pressure on the entire industry will only grow.
The Other Side of DeepSeek: Censorship and Cost Doubts
Chinese AI is far from flawless, however. DeepSeek’s answers are filtered in line with state censorship — the chatbot declines to respond to questions about Xi Jinping or Taiwan. Some analysts also question the reported development costs, suggesting that the Chinese government may have quietly subsidised electricity, salaries and computing resources for the project. Verifying this, though, is difficult.
There is also the question of data. The DeepSeek app stores user data on servers in China, which has already prompted regulators in several countries to look more closely at the service — a reminder that a cheap AI model can carry compliance costs of its own.
What’s Next for DeepSeek and the AI Race?
DeepSeek has already broken the rules of the game. From now on, the world’s leading IT companies will have to play by different rules. The technological race will continue — with Stargate-scale investment on one side and open-source models on the other — and we may truly be standing at the threshold of a new era of accessible artificial intelligence, or merely at the start of an intense contest between the United States and China in high technology.
What this means for startups and founders
Cheaper, open-weight AI lowers the barrier to building AI products: a small team can now run a top-tier model on its own infrastructure instead of paying for proprietary access. The trade-off is that the legal and regulatory side — data protection, content filtering, licensing terms — becomes the founder’s responsibility.
Eesti Firma provides corporate and legal support for AI projects and helps businesses keep pace with a constantly changing technological and regulatory environment. Feel free to contact us for professional consultation and support.
Frequently Asked Questions
DeepSeek-R1 is a reasoning-focused large language model released by the Chinese startup DeepSeek in January 2025. It performs at a level comparable to the models behind OpenAI’s ChatGPT while being far cheaper to train, and its weights are published under an open licence.
Investors feared that if a top-tier AI model can be trained on cheaper, restricted chips for a few million dollars, demand for Nvidia’s most expensive hardware may be lower than expected. The shares dropped 17% in one day, erasing roughly $600 billion in market value.
The model weights and code are publicly available, so developers can download, run and fine-tune the model themselves. The training data, however, has not been released, so it is more accurately described as an open-weight model.
DeepSeek reported a final training run cost of about $5.6 million. Some analysts note that this figure excludes research, earlier experiments and hardware purchases, so the total programme cost was likely much higher.
Yes — the open licence allows commercial use, and the model can be hosted on your own or European cloud infrastructure. Companies should still assess data protection, content filtering and the terms of the licence before deployment.
