DeepSeek: Smaller and cheaper may be better in AI arms race, says leading investment bank

Chinese startup DeepSeek’s meteoric rise is making many people wonder if adding more computing hardware is still the best way to boost artificial intelligence (AI), according to new research from Deutsche Bank.

Drawing an analogy with the EV sector, the paper asks: Do you do not need a high-end Tesla Model X car just to run quick errands, when perhaps the cheaper BYD can handle day-to-day needs?

DeepSeek’s low-cost R1 AI assistant, launched on 20 January, has soared to the top of download charts in the United States—sparking a wave of doubts about whether “bigger is better” remains a valid motto in AI.

Deutsche notes that the Chinese chatbot shot to the top of the charts despite using fewer advanced computer chips and a budget well below those of America’s AI powerhouses.

DeepSeek reportedly trained its model with only 2,048 Graphics Processing Units (GPUs), the specialised chips that accelerate AI tasks.

Cheaper, better

These GPUs were described as H800s made by Nvidia Corp (NASDAQ:NVDA, ETR:NVD), the world’s largest chipmaker, which are not most up to date technology.

Training a so-called “large language model” usually costs hundreds of millions—if not billions—of dollars. Yet DeepSeek claims it spent about US$5.6 million and still achieved performance comparable to bigger US models.

Nvidia, whose advanced GPUs are widely used to power AI systems, has seen its shares soar in the last couple of years.

Yet on the day DeepSeek hit the headlines, major US tech stocks, including Nvidia and Microsoft, fell sharply.

Many market watchers are now questioning whether unlimited spending on cloud computing and hardware will continue to be the main driver of AI innovation.

No need for brute force

The new research also flags how smaller, more efficient models could become the norm, says Deutsche.

DeepSeek’s sudden success suggests that if “small” systems can match a large model’s abilities, companies might not need to keep chasing ever-larger models that demand thousands of power-hungry GPUs.

Instead, smaller models could run on everyday devices like smartphones, lowering costs for businesses and consumers alike.

The German bank highlights that US restrictions on the export of certain high-end chips may have pushed Chinese companies to innovate in software rather than rely on brute-force hardware.

DeepSeek’s approach could be an example of how these constraints have forced Chinese developers to do more with less. The research calls it a reminder that China is not as far behind in AI as some might have assumed.

Rapid pace

Open-source AI has also been trending, which Deutsche Bank says may help newcomers like DeepSeek keep innovating at a rapid pace. If open models gain ground, businesses could switch AI providers easily, making AI a commodity service. That would potentially lower prices, draw in more users, and encourage even smaller models to flourish.

So far, DeepSeek has played coy about the attention. When asked for a comment, its chatbot often returned a simple message: “Oops! DeepSeek is experiencing high traffic at the moment. Please check back in a little while.”

It remains to be seen whether this lean and low-cost approach will permanently change the direction of the AI arms race, says Deutsche Bank.