News that artificial intelligence (AI) assistant DeepSeek can compete with models like ChatGPT and Gemini for a fraction of the cost and computing power made headlines on Monday (Jan. 27). The reporting caused technology and energy stocks to sink as much as 21% and raised questions about AI strategies.
These large language models generate text and images in response to user queries, processes that require significant energy consumption. Technology companies are increasingly incorporating them into internet search engines, social media platforms and productivity applications like Microsoft Word.
Penn State experts across the AI and business landscapes explained in the following Q&A what DeepSeek is and what it means for the future of AI.
Q: What is DeepSeek?
Shomir Wilson, associate professor of information sciences and technology, studies natural language processing and AI, such as the technology underlying large language models like ChatGPT, as well as security and privacy issues.
Wilson: DeepSeek is an artificial intelligence assistant along the lines of OpenAI's ChatGPT or Google Gemini. The assistant is designed to accomplish a broad variety of tasks, but DeepSeek is advertised to be particularly strong at formal reasoning tasks like math and logic problems. This is a significant claim because many contemporary AI assistants struggle with formal reasoning tasks.
What surprised me most was how rapidly observers, including the stock market, reacted to DeepSeek's claims. A big part of the advantage DeepSeek claimed is performance at "benchmarks," standard tests that people administer to AI assistants to compare them. Those comparisons don't necessarily show how useful an AI assistant is for practical work. I was also surprised that DeepSeek appeared to be much more efficient than its peers, in terms of computation and energy consumption, but researchers will need more time to assess whether these early claims translate to real-world advantages.
Q: How did stock markets react to the news?
Akhil Kumar, professor of supply chain and information systems, studies blockchain technology, business analytics, deep learning and AI systems, health IT, business process management and process mining.
Kumar: The news shook up Wall Street. Nvidia, the largest player in the semiconductor market, dropped by 17% and lost nearly a staggering $600 billion in market value on Monday. The other six stocks in what's known as the Magnificent 7 - Alphabet, Apple, Amazon, Meta, Microsoft and Tesla - were also hit but to a smaller extent. The market reaction was puzzling. On the one hand, DeepSeek is some 10 times more efficient than current products and needs fewer chips - 2,000 compared to 16,000 for its competitors. So, that may drive down the demand for Nvidia and other specialized chips. But on the other hand, the new technology is more efficient and more affordable. This means its use could explode, thereby creating enormous new demand for chips and hardware. This breakthrough could also accelerate progress towards AGI, or artificial general intelligence, a type of AI that matches or exceeds human intelligence capabilities.
Q: How might these AI assistants affect workers around the globe?
Mehmet Canayaz, assistant professor of finance, studies economics of agentic AI, technological and ethical standardization of AI, and chip manufacturing workforce design.
Canayaz: The business impact of AI assistants primarily comes from their role in enabling "agentic AI," where AI agents act as autonomous digital employees. Think of large language models (LLMs) as a chef who writes a recipe, while an AI agent is the chef who autonomously cooks the meal from start to finish. Like human employees, AI agents manage business tasks independently and interact with a company's internal data, software systems and personnel. They use a variety of tools, including but not limited to LLMs like DeepSeek and ChatGPT. The potential of these AI agents took center stage at this year's Davos conference, where Nvidia CEO Jensen Huang declared, "The age of agentic AI is here." This aligns with Meta CEO Mark Zuckerberg's earlier prediction that by 2025, AI agents at Meta and similar companies will have skills comparable to midlevel engineers.
In my research, I show how AI agents can lower costs compared to human employees while maintaining similar levels of task accuracy. Their effectiveness increases when businesses implement clever workforce designs and strong oversight. For example, when AI agents collaborate in a well-monitored environment, they demonstrate a clear advantage in autonomously performing business tasks traditionally done by humans (and solo AI agents). This is good news for firms financially and for workers who design and operate them from a career perspective. But it raises concerns for workers whose roles may be replaced.
Q: How does DeepSeek's approach to generative AI differ from its competitors? What does it mean for the future of AI?
Dana Calacci, assistant professor of information sciences and technology, studies crowdsourced AI audits and AI harms, data tools for workers, data rights as labor rights and commercial surveillance.
Calacci: I think the approach the DeepSeek team takes is good for AI development for a number of reasons. The first is that right now, many models are evaluated against a "global" idea of what a "good" response is to a given query or prompt. This creates biases and makes models less useful for marginalized groups and specific tasks.
My understanding is that DeepSeek's approach allows for more fine-grained, task-specific training in a way that is much cheaper and faster than current approaches. This means that their strategy could be used to make models that, for some prompts, are more accurate or more useful to specific communities. This is an under-reported, interesting benefit of their approach.
The second aspect is that this approach can likely cut training costs at least in half, train models faster and make smaller models. DeepSeek's approach uses half as much compute as GPT-4 to train, which is a major improvement. The size of the final DeepSeek model also means probably over a 90% reduction in the energy cost of a query compared to GPT-4, which is huge. Right now, GPT-4 queries are run on big cloud server infrastructure. DeepSeek can run on tinier, energy-efficient devices, potentially making things like GPT-4 deployable almost anywhere without a bunch of cloud computing owned by large technology companies. This makes developing consumer-facing apps with these models much more democratized and may reduce the monopolistic stranglehold that the big tech companies have on the market.
