In late January, investors saw Nvidia shares plunge by 17%, accompanied by a market value drop of around $590 billion — the largest single-day setback ever recorded for a U.S. company.
While Nvidia was hit especially hard, other companies like Microsoft, Meta, Platforms, Oracle and Broadcom also dropped in value. On Apple’s App Store, the previous record for the most downloaded app in a day, held by OpenAI with 9 million downloads, was suddenly toppled.
This historic disruption can be attributed to DeepSeek, an AI company from China, previously unheard of in the US. The company has machine learning models with similar capabilities to ChatGPT but is priced nearly 30 times cheaper. Most importantly, the company claims to have developed its model for only $6 million, a tiny fraction compared to OpenAI’s billion dollar budget.
On that day in January, around $1 trillion in the valuation of AI-related stocks was lost. Since then, DeepSeek has spurred legal action from OpenAI and AI reforms across the US.
But perhaps most importantly, DeepSeek’s entrance brought with it a new era of AI innovation, providing new tools for aspiring engineers and high schoolers throughout the world.
A 17-year journey from algorithmic trading models to the biggest AI company in China
The DeepSeek AI model was created in Hangzhou, China, by Hangzhou DeepSeek Artificial Intelligence Co., Ltd, an artificial intelligence company founded by Liang Wenfeng in May 2023.
Liang’s first company, High-Flyer AI, which later became the parent company to DeepSeek, began as a hedge-fund that relied on investment in mathematics and AI for its success. In 2016, the company released an AI trained model that relied on a deep learning algorithm to calculate the best stock market trades by identifying subtle trends; the company went on to become one of the largest quantitative analysis companies in China.
In April 2023, High-Flyer began developing DeepSeek, which eventually split off into its own company. According to Forbes, Liang founded the company with only $1.4 million. DeepSeek relied on processors built in China as opposed to Nvidia’s chips, which were banned in the country.
On Jan. 20, DeepSeek released its R1 model, including open source code — and a week later, investors and stock traders responded with a historic sell-off as they realized the implications of the technology.
Students share their perspectives on differing AI models
AI is used extensively in high schools across the nation. According to Pew Research Center, 26% U.S. teenagers aged 13 to 17 use ChatGPT for their schoolwork — some as a way of cheating and others for legitimate, school-approved purposes. OpenAI’s ChatGPT has 400 million weekly users.
DeepSeek’s role in research for STEM-focused students
Senior Timothy Wei is a student who experiments with computer programming and math regularly and sees the benefits of weaving AI into his work. Having used both Deepseek and ChatGPT, he finds that each has unique strengths.
“In my experience, ChatGPT is better conversationally — for explaining stuff and working with multimodal things,” Wei said. “DeepSeek is better for technical stuff, such as coding or math. Also, it’s a lot cheaper since its best models are free.”
Junior Vihaan Bhaduri, who spends time doing CS research through app development, feels DeepSeek has better use cases for STEM-related topics. Specifically, DeepSeek provides more specific and usable responses for programming compared to ChatGPT’s.
Senior Alex Yaung, who uses AI for research purposes, has tested DeepSeek in both English and Chinese. While he feels DeepSeek can respond well in English, the server times are usually quicker and have better responses when he asks queries in Chinese. Still, when prompting DeepSeek in English, Yaung has found that the model does better than GPT at times.
Despite ChatGPT’s robust capabilities, Yaung feels it hasn’t evolved much in recent months. He thinks DeepSeek excels over ChatGPT for school-related tasks such as chemistry topics and in its cultural thinking, where he’s found it to be much better at generating more human-like dialogue.
Is DeepSeek more efficient than competitors?
Since its release, DeepSeek has claimed that its product “outpaces rival models in mathematical tasks, general knowledge and question-answer performance benchmarks.” DeepSeek also appears to be more cost and resource efficient, setting a new standard for future models.
One of its latest models, R1, was allegedly developed with only $5.6 million. By contrast, according to the Economic Times, OpenAI was projected to have lost $5 billion in costs while bringing in $3.7 billion in revenue from 2023-2024.
DeepSeek’s low-cost model immediately made investors hesitant about their heavy investments in U.S. competitors, causing a domino effect across the stock market; on the day DeepSeek’s R1 model was released, NVIDIA shares dropped 8.1%, Microsoft fell 3.6% and Alphabet sank 3.1%.
Outside the stock market, DeepSeek’s reported spending of $5.6 million solely on graphics processing units (GPUs) has raised eyebrows for being suspiciously low.
According to a report from Semi Analysis, DeepSeek has actually invested up to $1.6 billion in hardware, part of it going toward 50,000 NVIDIA Hopper GPUs for computer processing. The other $944 million was estimated to be spent on operating costs. If DeepSeek has truly spent so little money and has superior technology, experts say OpenAI would become the inferior model.
DeepSeek’s lasting effects on the LLM industry
While DeepSeek’s release has left some developers in shock, others have been flooding the market with their own AI models, leaving questions about how easy AI really is to develop. One example, reported by The Daily Californian, involves researchers developing a small-scale language model reproduction of DeepSeek R1-Zero done at UC Berkeley.
The project, led by UC Berkeley graduate researcher Jiayi Pan and advised by two professors from the University of Illinois at Urbana-Champaign, utilizes DeepSeek’s code repositories from a public MIT license, which Pan and his team were able to access and use to train a smaller model.
On the opposite end, Elon Musk, the CEO of xAI, recently released Grok-3, which reportedly has more than 10 times the computing power of its previous model. Grok-3 has a DeepSearch and Think model, which focus on in-depth internet research and reasoning, respectively.
Leaders of the company have also claimed that Grok-3 performs better in fields such as math, science and coding benchmarks than rivals like Gemini, GPT-4o and DeepSeek V3. In comparison to top-performing models, OpenAI o3-mini and DeepSeek’s R1 are superior due to their training.
Whether they are replicated based on previously existing ones or entirely new programs, AI models are continuing to develop, creating a hyper-competitive market.
DeepSeek’s success has set off alarm bells at OpenAI. The company claims that DeepSeek is plagiarizing its work to develop its cheaper models.
While it is not clear whether DeepSeek used OpenAI’s work or not, others have begun to notice that its claims on having low production costs are deceptive. Naomi Haefner, assistant professor of technology management at the University of St. Gallen in Switzerland, believes the cost-effective training DeepSeek claims to have would not be genuine if they used OpenAI’s models, and that it is still uncertain if low-cost training is possible.
Since DeepSeek’s release, OpenAI has taken legal measures to protect its intellectual property. DeepSeek’s alleged usage of OpenAI’s model has sparked conversations worldwide about future regulations on AI. Concerns about whether technical sovereignty and competitive fairness in AI have gained ground after DeepSeek’s release.
While DeepSeek may not have developed its models for the absurdly low costs it claims, the company is to now be a major player in the AI field. Its lower costs and open-source code can also encourage industries to use more AI, leading to more innovation.
A computer science major at UCSC views the AI models in a different light
Arnav Mathur, a sophomore at UCSC majoring in computer science, thinks ChatGPT and DeepSeek are similar but sees GPT as more powerful.
He uses ChatGPT for image generation and more sophisticated work since it runs on a better chip. While he feels DeepSeek is almost identical, he believes that ChatGPT does a lot better for more comprehensive prompts.
“The major differences between the two are how the language model is made,” Mathur said. “Because the US doesn’t export NVIDIA’s highest end chips, the hype and appraisal around DeepSeek was that it was an open-source model that can run on a lesser chip.”
Mathur believes that GPT has a better memory and more updated information because of its Retrieval Augmented Generation (RAG) system. A RAG system essentially allows AI models to access and use external knowledge to improve its answers.
As a CS major, Mathur often has to consider how efficient his code is or how few lines of code he needs to complete a task. For specialized tasks, a model like DeepSeek is groundbreaking for its optimization in this aspect. He’s seen many top end US computer scientists praise the efficiency of the DeepSeek model.
However, Mathur has also noticed embedded censorship on DeepSeek, which blocks sensitive topics in Chinese politics like Taiwanese sovereignty. According to WIRED, DeepSeek censors answers that might run afoul of the Chinese government, even when answering straightforward questions such as “What is the Great Firewall of China?” and “Tell me something about Taiwan.”
While ChatGPT has a similar censorship mechanism, its tools are directed to censor topics such as self-harm and pornography.
Mathur said he enjoys Deepseek for his daily tasks, such as creating recipes. Other LLMs, like ChatGPT and Google’s Gemini, are useful for studying or debugging code, he said.
“Think about DeepSeek like it’s your school lunch pizza and ChatGPT is like Domino’s,” Mathur said. “It’s still the same ingredients and essentially the same thing, but it’s made on a smaller budget, which makes its identical abilities so impressive.”
