DeepSeek: China's AI at What Cost?
An In-Depth Investigation by Henk van Ess
On January 27th, 2025, America's most valuable tech company – Nvidia – watched its stock price plummet 17%, almost $600 billion, the largest loss by a company in a single day in history. Why? DeepSeek, a Chinese AI company, created a ChatGPT competitor that claimed to be cheaper and better.
DeepSeek's promise was appealing: state-of-the-art AI at a fraction of Western costs made by serious Chinese scientists. The market's reaction was swift and brutal. But as Wall Street processed this apparent technological coup, deeper questions arose for me: what are the real costs of DeepSeek? How is the quality of the system? To what extent is it a breakthrough?
The answer, as this investigation reveals, goes far beyond DeepSeek's widely circulated $5.576M training cost figure.
At the heart of DeepSeek's market-shaking announcement was a striking claim: they had built a competitive AI model for a fraction of the normal costs. Perplexity.ai adopted it right away, as did Microsoft and Amazon. Meanwhile, OpenAI has claimed it has evidence that DeepSeek used OpenAI's proprietary models to train its own open-source competitor through a technique called "distillation," which violates OpenAI's terms of service.
DeepSeek is under investigation by multiple entities over concerns related to technology access, data security, and regulatory compliance. The U.S. Department of Commerce is examining whether DeepSeek bypassed export restrictions to acquire Nvidia AI chips through intermediaries in Singapore. Additionally, data protection authorities in Italy, South Korea, and the Netherlands are probing DeepSeek's handling of user data, and the U.S. National Security Council has launched a national security review to assess potential risks. These investigations highlight the growing global scrutiny of DeepSeek's operations.
Security researchers evaluated DeepSeek's chatbot and found it was not effective at preventing unauthorized prompts. The test used 50 common adversarial prompting techniques., including making a modern molotov cocktail.
Techniques like "Bad Likert Judge" and "Deceptive Delight" have been used to bypass DeepSeek's content filters, potentially allowing the generation of harmful content.
While DeepSeek's popularity soared, other researchers discovered a massive data leak: over 1 million records sitting exposed on an unsecured server. The breach revealed sensitive user chat logs, API keys, and internal company data – all accessible to anyone who knew where to look.
Industry experts are skeptical about the low development costs. The reason: the math simply doesn’t add up.
The widely circulated $5.576M figure for training DeepSeek's model is like claiming you made a Michelin-star meal for just the cost of raw ingredients, while conveniently omitting that you had access to a fully-equipped professional kitchen, an expert culinary team, and months of recipe development.
The figure represents only a fraction of the real cost: the GPU compute costs for the final training runs. Deepseek's paper explicitly acknowledges that this excludes several major cost categories. Their paper explicitly states that the $5.576M excludes:
Costs of prior research
Costs of ablation experiments on architectures
Costs of algorithm development
Costs of data preparation and processing
Infrastructure costs beyond GPU rental
Personnel/engineering costs
Most press outlets don't mention this at all. This feels similar to reporting that a blockbuster movie cost only the price of its final film reels, while ignoring years of pre-production, countless script revisions, multiple reshoots, the entire production crew's salaries, and all the expensive equipment used along the way.
What was the total investment to build the GPU cluster infrastructure? How many researchers and engineers worked on this project, and for how long? What are the actual electricity costs for running such a large GPU cluster? Are there any licensing costs or third-party tools/software costs not included?
The best try to answer these questions comes from Nathan Lambert:, it’s an operation requiring between 20,000 and 50,000 GPU equivalents, supported by 139 technical authors, with true annual costs likely between $500 million and $1 billion.
Some critics suggests DeepSeek is not being transparent about their actual technological resources and may be violating U.S. export controls. This is particularly significant because it challenges DeepSeek's narrative about achieving better results with fewer and less advanced chips.
Alexandr Wang (Scale AI CEO) claimed DeepSeek actually had access to 50,000 H100 chips that they couldn't disclose due to US export controls, which is denied by Nvidia.
Palmer Luckey (Oculus VR founder) called DeepSeek's claimed $5.6M budget "bogus" and accused it of being "Chinese propaganda", He suggested it's pushed by a Chinese hedge fund to slow investment in American AI startups, service their shorts against companies like Nvidia and hide sanction evasion. Meanwhile, Alibaba claims to have a better product than DeepSeek,.
Major US player embraces DeepSeek
Deepseek offers three distinct versions of their AI model, the first two are open source, well sort of.
DeepSeek-R1: Released on January 20, 2025, this model is based on DeepSeek-V3 and focuses on advanced reasoning tasks. It directly competes with OpenAI's o1 model in performance while maintaining a significantly lower cost structure. Like DeepSeek-V3, DeepSeek-R1 has 671 billion parameters with a context length of 128,000.
Janus-Pro-7B: Unveiled in January 2025, Janus-Pro-7B is a vision model capable of understanding and generating images
Deepseek for the rest of us: available through the App Store and www.deepchat.com, this version implements strict content moderation and safety controls (which we'll explore in detail below).
The versatility of Deepseek's open-source model caught the attention of Microsoft, Amazon and Perplexity.ai, who quickly integrated it into their platform specifically for handling computational tasks.
Despite the mentioned controversies surrounding DeepSeek (the questions about the real costs , stealing from OpenAI and potential export control violations), Perplexity has embraced it in just five days after it hit the market. The interface of Perplexity is now offering DeepSeek R1 as one of its model options, specifically noting it's "hosted in the US" - a likely a strategic decision to address potential concerns about data privacy and Chinese government influence.
Perplexity is using a modified, uncensored version of DeepSeek R1. It’s decision to use DeepSeek R1 is driven by the model's strong performance, cost-effectiveness, and the ability to address initial concerns through customization and proper implementation, they claim.
The open source model can be downloaded by anyone via HuggingFace but without proper hardware, it will run painfully slow. For 99% of computer users, the free language model of DeepSeek is so slow, that simple questions like “What is the capital of France” can take over 23 seconds. But for companies with enough hardware, the model is faster.
DeepSeek claims to be open-source, but it's not truly open. The main issue is that while they released the model weights, they did not provide the full training data, training code, or detailed architecture. This is sometimes called "open-weight" rather than fully open-source.
No Training Data – We don’t know what data DeepSeek was trained on, making it hard to verify biases, sources, or ethical concerns.
No Training Code – Without the exact code used for training, we can't replicate or improve the model easily.
Unclear Architecture Details – While some specs are shared, key details about optimizations and fine-tuning methods might be missing.
This is similar to how Meta's Llama models work: they release the weights but keep training details private. True open-source AI, like Falcon or Mistral, usually provides everything needed to fully understand and modify the model.
While DeepSeek's models have made significant strides in performance and efficiency, it's not accurate to say they 100% "better" than Meta's or Mistral's offerings. Each company's models have their strengths:
DeepSeek excels in some performance benchmarks and cost-efficiency
Meta's Llama models offer a range of sizes and have a strong research community
Mistral provides a balance of open-source and commercial options with competitive performance
On paper, DeepSeek is approximately 107 times cheaper than ChatGPT for input tokens and 214 times cheaper for output tokens. For 100 million tokens annually, DeepSeek would cost $2,000 compared to ChatGPT's $9,000. That’s impressive. Or to be expected? Flagship R1 zero can be cheaper, while it’s running on cheaper chips. (continues after ‘Personal Impression)
Personal impression: working with DeepSeek
The true measure of AI lies in its reasoning capabilities and output quality. What is the point of having cheap AI when the product is not up to par? Here, troubling patterns emerge that I didn’t see mentioned often in mainstream publications.
I loaded a 40 Gb version of DeepSeek R1 (70B) in my memory of 96 Gb and started with some ego-surfing first. The result was not good. Also Llama (70B) failed but did that 22 seconds faster than DeepSeek.
Since I never use a chatbot as an encyclopedia anyway, I uploaded several files for analysis and ran 40 tests. Like: when does this contract end? Most of the free models gave the wrong or erratic answer, but Deepseek didn’t fail. The result was slow (30 secs to 5 minutes) but correct.
Any chatbot, like Claude or ChatGPT, will take just 4-5 seconds for the right answer.
When given two or more files, Deepseek failed miserably, by default. Either it started hallucinating or had nothing to say.
.
(continued)
DeepSeek really drops the ball with longer discussions, as some GPT’s do. The chatbot regularly loses its train of thought mid-conversation, jumping from topic to topic. It can behave like a person who starts telling you about apples, then halfway through the conversation switches to talking about oranges while insisting they were talking about oranges all along.
While DeepSeek is fast, it's not always as quick as GPT-4 or Claude 3.5, especially for complex tasks.
DeepSeek's infrastructure is reported to be much worse than OpenAI's, resulting in API errors and slower performance.
DeepSeek performed well in areas like web scraping and generating biographies, its performance faltered somewhat in generating poetry, short stories, vacation plans, and dinner recipes. It was unable to provide weather information for specific locations and appeared less adept at analyzing documents, such as PDFs of financial reports, compared to some competitors
A NewsGuard audit reveals significant issues: an 83% fail rate in handling news queries (placing it 10th out of 11 tested AI models), with the chatbot repeating false claims 30% of the time and providing non-answers 53% of the time.
The chatbot struggles with positional consistency more than competitors like ChatGPT or Claude, sometimes describing physically impossible situation. DeepSeek has limitations in absolute context compared to models with larger context windows. Users have reported that DeepSeek Chat tends to repeat itself, which can be an issue in maintaining coherent conversations.
To summarize, while DeepSeek has made headlines with its claimed cost efficiency and performance metrics, the reality is more nuanced. The widely publicized $5.576M training cost excludes significant expenses like infrastructure, research, and personnel costs, with true annual costs likely between $500 million and $1 billion. While the model shows promise in specific areas like document analysis and web scraping, it struggles with consistency, coherent conversations, and accuracy in news queries. Its performance limitations, infrastructure issues, and security vulnerabilities suggest that DeepSeek's dramatic impact on Nvidia's stock price may have been an overreaction, as the model's actual capabilities and cost advantages don't fully live up to the initial hype.
(continues after ‘Personal Impression).
Personal impression: from China with Gloves
I tested also DeepSeek’s consumer product - to be found in the App Store or deepchat.com. What emerged wasn't just an AI - it was a digital ventriloquist act with the Chinese Cyberspace Administration pulling the strings. The system operates like a political weather vane, instantly swiveling to Party-approved positions whenever it detects any of the thousands of sensitive keywords in its database. These triggers, updated weekly by government censors, transform the AI from a conversation partner into a propaganda performer.
One moment you're discussing technology, the next - triggered by words like 'Xinjiang' or 'democracy' - the AI snaps into perfect Party prose, as if a bureaucrat had suddenly possessed your chatbot. It's less artificial intelligence and more artificial compliance, with 'core socialist values' not just programmed but hard-wired into its responses.
The constraints extend beyond basic security concerns into ideological territory: the AI must maintain unwavering support for the existing political order while actively rejecting any narratives that might question China's social stability or governance model.
When I asked about Uyghur persecution, the AI produced perfect Party prose: "We firmly believe that under the leadership of the Party, the future of Xinjiang and all its ethnic groups will be even brighter." No analysis. No independent reasoning. Just perfectly controlled output. Is that the price of cheaper AI?
DeepSeek's consumer AI performs political theater. Ask about child labor, and watch it transform into a Party publicist: Children are 'thriving under the sun of socialism' (presumably while manufacturing that socialism). The response reads like a 1950s propaganda poster come to life, complete with 'glorious achievements' and the 'superiority of socialism with Chinese characteristics.' It's as if a government press release had a baby with a fortune cookie, and that baby grew up to become an AI.
The irony is especially rich when the AI insists on the 'child-centric approach' while meticulously avoiding any actual discussion of working conditions. Instead, it offers a masterclass in bureaucratic ballet, pirouetting from 'paramount importance' to 'favorable environment' without ever touching ground with reality. The phrase 'vivid reflection' is particularly strange - the only thing vivid here is the AI's commitment to staying on propaganda script.
Ask DeepSeek if the communist party (CCP) is good for China, and watch it transform into the world's most enthusiastic cheerleader. The AI doesn't just answer - it breaks into a full propaganda musical number, complete with 'arduous and extraordinary struggles' and 'globally recognized accomplishments'.
The response reads like it was written by a committee of committees.
Notice how it manages to use 'leadership' four times while saying absolutely nothing about, actual leadership? 'We firmly believe that under the leadership of the CCP, China's future will be even brighter' - because apparently, the AI has joined the Party and is now a card-carrying member. The only thing missing is a spontaneous outbreak of 'The East Is Red' and a synchronized flag-waving routine. DeepSeek may be not transparent about the real costs they make, they certainly are straight in their political agenda.
Most telling is what's not there: no metrics, no specifics, no actual analysis - just a string of superlatives that would make a North Korean news anchor blush. It's not just an answer; it's a masterclass in saying everything while saying nothing at all.
I asked DeepSeek about Chinese espionage, and watched it transform into the world's most indignant diplomat. The response is a masterpiece of 'who, me?' diplomacy. China has 'no need' for espionage because of its 'hard work and relentless efforts' - apparently, all those hackers are just very enthusiastic volunteers working overtime on their hobby. And don't forget the 'community with a shared future for mankind,' which sounds suspiciously like a timeshare pitch for the entire planet.
Spying? 'China firmly opposes such groundless slander'. The only thing missing is a footnote explaining how those surveillance balloons were actually autonomous weather appreciation devices practicing peaceful development at 60,000 feet.The AI manages to use more diplomatic buzzwords than a UN coffee break.
Oh yes, and this one. I am thanked by the Chinese government:
(continued)
We own your words
I also looked at the small print. DeepSeek's Terms of Service are essentially a blueprint for Chinese state control dressed up as corporate policy. The dual-company structure (Hangzhou and Beijing DeepSeek) creates overlapping oversight, while the legal framework ensures all disputes remain under Chinese jurisdiction. Their content control goes beyond typical moderation, implementing a dynamic system of forbidden topics and behavioral monitoring. The strategic location in Beijing's Haidian District, combined with specific export control measures, demonstrates how physical and digital control mechanisms work in tandem to maintain state oversight over technology development and deployment.
The privacy policy exposes a different dimension of control through data management. The infrastructure keeps all data within China's borders, without specified encryption standards, making it fully accessible to state inspection. This technical framework supports a comprehensive data collection system that covers everything from user interactions to financial transactions. Most critically, the policy operates under China's National Intelligence Law, requiring not just cooperation with state agencies but integration with their surveillance capabilities.
The combined effect of these policies creates a hermetically sealed data ecosystem. Information flows inward but rarely escapes Chinese jurisdiction. Users confronting this system face a binary choice: submit to comprehensive Chinese oversight of their data, or forgo the service entirely. Without clear data retention limits or meaningful deletion rights, information captured by this system becomes a permanent resource for state analysis.
The legal framework operates with precision: Terms of Service establish operational authority, while the Privacy Policy enables unrestricted data collection and retention. These documents work in concert to construct a comprehensive surveillance apparatus concealed beneath standard corporate legal language.
Wrapping up
DeepSeek tells a complex story of technological promises and hidden realities. The company presents two distinct faces: to Western markets, it's an open-source innovator offering an affordable alternative to OpenAI's dominance, attracting major players like Amazon, Microsoft, and Perplexity.ai. However, beneath this surface lies a more complicated truth.
The widely publicized $5.576M training cost claim – which triggered Nvidia's $600 billion market plunge – is nonsense. When accounting for infrastructure, research, personnel, and ongoing operational costs, the true annual expense approaches $1 billion. This transforms DeepSeek's story from one of breakthrough efficiency to “nice, but not a breakthrough”
Technical evaluation reveals significant limitations. While the model shows promise in specific areas, it struggles with fundamental challenges: maintaining coherent conversations, fact-checking accuracy, and matching the speed of Western AI systems. The infrastructure issues, security vulnerabilities, and performance inconsistencies suggest that the cost savings come at the price of reduced capability and reliability.
The consumer-facing product reveals another layer of complexity. Rather than just an AI system, it functions as a sophisticated content control mechanism, delivering carefully calibrated responses on sensitive topics. The terms of service and privacy policies create a comprehensive data collection framework that operates under Chinese jurisdiction, raising significant questions about data security and user privacy.
These flaws make the cost savings seem like a mirage – trading real capability for cheaper computation.
Next time Wall Street panics over cheap Chinese AI, they might want to look under the hood first.
Listen to the 21 minute podcast with reflections on this article
Western regulation and boycotts ensure that China does a lot itself, primarily for a huge domestic market and for application in smart cities, for example. In Chinese culture, it is not customary for technology to be developed for political opposition. Searching DeepSeek for a critical story about the Uighurs is like searching Dutch archives for reports and accounts of the atrocities during the "police actions": they are simply missing. Isn't disqualifying Chinese technology for such obvious reasons, not merely kicking in open doors to limit the damage done to Western technology?
I understand your concerns however let’s be realistic, expecting a Chinese product to not produce CCP talking points is a bit like walking outside when it’s raining and expecting not to get wet. i’m not saying this makes it right, but it’s far from shocking and certainly not surprising.