Last week, social media was awash with a flurry of comments and reactions. The announcement of the launch of a China-based open-source language model christened DeeepSeek caused a storm. What appears to be fascinating to a lot of people is DeepSeek’s performance benchmarks. It is on par with that of the grandmaster of LLMs —ChatGPT. In fact, in some cases, it proved to be even better despite costing a fraction of the cost to build.
But as one internet user aptly puts it,
«The emergence of DeepSeek shouldn’t have been a surprise to anyone, if you know what China has been doing in the last 20 years.”
So, in this article, we want to go over everything you need to know about DeepSeek and the reactions that have followed its launch around the world, particularly in North America and Europe.
What Is DeepSeek R1 And Why Is It Causing Controversy?
On the surface, DeepSeek was founded by the Chinese edge fund, High Flyer. The company had built a 100 billion yuan ($13.79 billion) investment portfolio using proprietary artificial intelligence models to make decisions. However, it decided to pivot in 2023 from the financial market to focus on developing cutting-edge AI models in hopes of ushering in the artificial general intelligence era.
«High-Flyer will concentrate its resources and strength, wholly devote itself to serve AI technology that benefits all of humanity, create a new independent research group, and explore the essence of AGI,» the company once said in a post on its official WeChat account.
The independent research group envisaged by High-Flyer was DeepSeek. According to various insiders in the US, people knew China was coming for the AI industry.
However, not a lot of people realize how advanced China’s AI models have become. Especially since Nvidia’s powerful chips needed to run these AI models had long been banned from being exported to China by the US government in September 2022.
How Does Chain-Of-Thought AI Work?
But then, as with a lot of things with China these days, the launch of DeepSeek was a clear indication that the attempt to slow China’s AI advancement by the West was futile — again. This was the reason for the frenzy from tech insiders.
To demonstrate how impressive this feat is, at the time of DeepSeek’s launch on the IOS store in January, ChatGPT was the most downloaded app. It had an impressive 110 million downloads. However, within days of launch, DeepSeek not only displaced OpenAI from the top of the app store. It also got influential Silicon Valley figures praising its release as an MIT-licensed open-source project, as well as the approach it took to the development of its model.
This approach is what insiders call the chain-of-thought approach. It is particularly distinct from the approach taken by OpenAI’s ChatGPT.
To understand the chain-of-thought approach to training an AI model, think of it like this. To train regular large language models like OpenAI’s, you feed the model with data. You then query it about that data and rank the performance of answers to said query appropriately. This is what is called ‘supervised’ fine-tuning.
In contrast, what makes DeepSeek special is that it uses direct reinforcement learning instead. Meaning that it teaches itself to correctly answer a query and shows all the ‘thinking’ steps it did to arrive at an answer.
This is why it is called a ‘reasoning or chain of thought’ model and not just a large language model.
Although the details of direct reinforcement learning are complicated and won’t be covered here, what is important to know is this: it is the approach closest and most appropriate toward a world where AI models would be independent of human input. Or what’s commonly called artificial general intelligence (AGI).
Deepseek R1 vs OpenAI O1
That said, while DeepSeek’s success is impressive on account of the amount of obstacles against its creation alone, what’s more impressive was its performance benchmark against the established giants.
This benchmark evaluates DeepSeek’s performance on the American
Invitational Mathematics Examination (AIME) contest
and others. Source: DeepSeek
According to various reports, DeepSeek achieves performance comparable to OpenAI’s December 2024-launched O1 model across math, coding, and reasoning tasks.
The O1 version of the popular GPT model uses the same reinforcement learning and spends time «thinking» before answering, making it better at certain tasks than GPT-4o.
However, up until the release of DeepSeek, it was only available as a premium service to OpenAI subscribers.
How DeepSeek Wiped $1 Trillion from US Stock Price
The surprising launch of DeepSeek caused industry insiders and observers to ask questions. Who are the people behind this AI model, and how does it work? What is the prospect of this new model? How much did it take to build? Etc.
Naturally, some of these questions were asked with respect and in comparison to OpenAI’s GPT models. And they were also being asked by everyone in tech, including investors in AI and other related Silicon Valley stocks.
Given these questions and the surprise surrounding its release, it’s easy to see why and how the launch of DeepSeek caused a massive selloff in tech stocks.
Already, before the launch of the model, mixed reactions had trailed the news that OpenAI had secured $500B in investment to build AI data centers around the US, with concerns hovering over OpenAI’s profitability and currently limited model use case.
Thus, critics believe the addition of a new competitor in Deepseek caused stock prices to drop because investors might have felt that they had been oversold on AI’s future. Critics observed that the stock value, difficulty, and the needed investment for AI development and other related businesses —such as semiconductor manufacturing- are perhaps «overloaded» since there were reports that DeepSeek cost just $6 million to build –an amount significantly less than the reported cost for ChatGPT.
In addition, DeepSeek’s launch also means hobbyists do not need to buy Nvidia chips that are traditionally used in training and running these AI models.
The project’s creators made it so that everyday processors such as Apple M2s can be used instead. DeepSeek showed there was no need for the purchase of specialized microchips or graphics cards for AI development.
As one insider puts it in an appraisal video on YouTube, “In order for the stock value of OpenAI and its associated businesses like Nvidia to remain high they need people to believe AI is hard but as DeepSeek has shown, a better AI model can be built for way less and faster than OpenAI did and [they] don’t need to raise trillions of dollars like [OpenAI] recently did.”
Controversies
Nonetheless, following the surprising DeepSeek launch and the even more surprisingly good reception to it in most places in the world, including the US, it started to circulate that DeepSeek stole OpenAI training data.
DeepSeek Accused Of IP Theft By OpenAI
This allegation was first made by “AI sza” and the chair of the US President’s Council of Advisors on Science and Technology, David Sacks, who was a former colleague and co-founder with another of President Trump’s associates—Elon Musk.
According to Sacks, DeepSeek stole output from OpenAI to train its model using a technique known in the AI industry as ‘Distillation,’ which is generally accepted but which OpenAI forbids in its terms of service.
DeepSeek Chinese Censorship
Furthermore, several users of the China-hosted web interface for DeepSeek have alleged that the model sometimes refuses to talk about censored events that relate to China’s history, such as the controversial protest of 1989 in Tiananmen Square.
However, others have pointed out that this issue of censorship only exists when using the China-hosted interface. Since DeepSeek is open-sourced, censorship issues don’t exist when self-hosted.
Why is DeepSeek banned in Italy?
In an instance highlighting the controversy surrounding the model, Italy announced a ban on DeepSeek outright and removed it from app stores across the country at launch.
This was followed by bans by several other countries and government agencies, including in Australia and Taiwan, due to what they called concern for data privacy.
20 Years In Prison For Using DeepSeek in The US
Furthermore, in another instance of deterrence, this month, U.S. Senator Josh Hawley proposed a bill that could criminalize downloading DeepSeek AI and penalize offenders with up to 20 years in prison for violations.
Chinese Tech Boom Explained
Following charades that have followed the launch and reception of DeepSeek and several other Chinese AI models that launched alongside it, such as Janus, Hunyuan3D 2.0, Alibaba’s Qwen2.5 Max, and Kimi K1.5, critics are tempted to think that the AI race is between competing Chinese companies rather than US ones due to how good these Chinese models are.
As a matter of fact, tech commentators have muttered that the idea of banning and cutting off China from acquiring AI chips in 2022 might have led to the country becoming more resourceful in that regard.
Even further, some wonder why DeepSeek was revealed just after the US invested $500B in AI data centers and why several of these models were unveiled at almost the same time.
Others have pointed out that this is part of a broader trend. Chinese apps like TikTok, Shein, and Temu, as well as electric vehicle brands such as BYD and Drone manufacturing giant DJI, have all become household names around the world.
As an illustration, there are reports that Chinese researchers now publish more papers on Quantum computing than anywhere else in the world. With EVs, too, the Chinese recently became the largest consumers of electric vehicles, and their manufacturers have recently become some of the largest.
According to analysts at the BBC, China’s tech dominance was part of a long-term plan that was put in place in 2015 to move away from its history of low-cost goods manufacturing in what the country termed the China 2025 manifesto. A goal which has largely been achieved.
Nvidia Vs AMD & Apple
Apart from concerns regarding China’s increasing tech dominance, another area of concern for tech industry insiders and investors in the US is the semiconductor industry.
This is because the stocks that were most affected after DeepSeeks’ launch were those of chip manufacturing companies such as Taiwan Semiconductor, Broadcom, and Nvidia.
For years, Nvidia has been the primary beneficiary of the AI boom. The company has had a monopoly on AI training tech for the last decade because it has better Linux drivers than the next-biggest microprocessor chip manufacturer, AMD. Because of this, almost all machine learning algorithms have been optimized to run on the CUDA library, which is proprietary and only runs on Nvidia-produced GPUs.
This has been the source of a longstanding question that has been asked by industry insiders for 10 years, but has now come to a head.
With the DeepSeek launch, people noticed that it achieved 10X efficiency in part by bypassing the use of Nvidia’s proprietary CUDA library that only runs on Nvidia-produced GPUs.
What this means is that the future of AI training is tilting towards low-cost general-purpose processors rather than expensive proprietary ones such as those found in Nvidia GPUs.
Read more on this from the attached X thread above. Or you can view a recent discussion about Nvidia Vs AMD chips here.
Musk Vs Sam Altman
“ [Elon Musk’s] whole life is from a position of insecurity, I feel for the guy… I don’t think he’s, like, a happy person.” —Sam Altman in a recent Bloomberg interview.
Finally, OpenAI’s CEO Sam Altman revealed that he has recently refused to sell OpenAI to Elon Musk for $9.74B.
Sam shades Elon after the latter proposed
buying OpenAI
Now, you may be wondering, isn’t Elon a major competitor with XAI? Well, yes. However, critics alleged that Sam’s refusal is likely an attempt to sell OpenAI to himself for way below market valuation.
This is coming months after OpenAI confirmed in December 2024 that it has plans to restructure its operations. This could mean the company would move from a nonprofit to a for-profit business model.
There are arguments that since OpenAI is a nonprofit, its primary duty is to humanity, not shareholders, and as such, it shouldn’t have the incentive to make money.
Downloading DeepSeek
In conclusion, to try out DeepSeek as an everyday user, you could use it on the web at chat.deepseek.com. Conversely, DeepSeek can also be deployed locally using the following hardware and open-source community software according to the project’s GitHub repo:
- DeepSeek-Infer Demo: We provide a simple and lightweight demo for FP8 and BF16 inference.
- SGLang fully supports the DeepSeek-V3 model in both BF16 and FP8 inference modes, with Multi-Token Prediction coming soon.
- LMDeploy: Enables efficient FP8 and BF16 inference for local and cloud deployment.
- TensorRT-LLM currently supports BF16 inference and INT4/8 quantization, with FP8 support coming soon.
- vLLM: Support DeepSeek-V3 model with FP8 and BF16 modes for tensor parallelism and pipeline parallelism.
- AMD GPU: Enables running the DeepSeek-V3 model on AMD GPUs via SGLang in both BF16 and FP8 modes.
- Huawei Ascend NPU: Supports running DeepSeek-V3 on Huawei Ascend devices.
Alternatively, you can follow Alex Chema’s tweet above to learn how to run DeepSeek R1 on 7 M4 Pro Mac Minis and 1 M4 Max MacBook Pro.
Looking for help building a product idea? Reach out to us through the form below. We help businesses like yours build and deliver big ideas. See our case studies for more.
