Photo: DeepSeek.
Well, all it took was one week following the demonstration of its capabilities for Chinese AI company DeepSeek, to make its presence felt and for us to see NVIDIA lose nearly US$600 billion in its market value.
But what is DeepSeek and why are US companies so afraid of it?
Part of the reason behind all the brouhaha is that DeepSeek was able to deliver a comparable AI experience, US$5.6 million in computing power for its base model, especially when compared to its US counterparts. Additionally, its low-cost approach makes it seem as though ChatGPT and the like, have poured too much money into developing AI.
Another factor that causes concern is that DeepSeek was able to do all of this despite a ban from the United States Government citing national security concerns, to restrict the supply of high-power AI chips to China. This also means that DeepSeek delivered its AI platform without needing to use the latest and more powerful AI chips.
According to reports, DeepSeek claimed to have used only 2,048 NVIDIA H800 chips and US$5.6 million to build what is known as a reasoning-focused model. For comparison, to build its Llama 3.1 model, Mark Zuckerberg’s Meta used 16,000 NVIDIA H100 chips. This is relevant because H800s, are Hopper GPUs and have constrained memory bandwidth compared to H100 processors, designed specifically for export to China due to export regulations.
Photo: DeepSeek.
DeepSeek’s R1 release is said to have outperformed Meta and Anthropic (Amazon-backed) AI models, as well as OpenAI’s ChatGPT-4o, in some benchmarks such as accuracy, coding and complex problem-solving.
And by doing so, DeepSeek showed that despite US technology curbs, China could still challenge its AI dominance.
“DeepSeek, however, just demonstrated that another route is available: heavy optimisation can produce remarkable results on weaker hardware and with lower memory bandwidth; simply paying NVIDIA more isn’t the only way to make better models,” one report added.
But not everything is all rosy for DeepSeek. It said yesterday that it was “temporarily limiting registrations” due to a cyberattack after the company's AI assistant amassed sudden popularity. It also remains to be seen what other actions the United States Government has in mind to further slow China’s AI moves.
Source: Reuters, Stratechery, DeepSeek, The Telegraph
Our articles may contain affiliate links. If you buy through these links, we may earn a small commission.