To The Moon

DeepSeek’s AI Model Threatens $NVDA’s Premium GPU Strategy

A little known AI lab from China has caused a stir in Silicon Valley by releasing AI models that surpass America's top performers, despite being developed more cost-effectively and with less powerful hardware.

The lab, known as DeepSeek, introduced a free, open-source large-language model in late December. According to the lab, the model was developed in just two months at a cost of under $6 million, utilizing Nvidia's lower-capability H800 chips.

China's DeepSeek could pose a significant threat to US equity markets, as the company appears to have developed a groundbreaking AI model at an exceptionally low cost without relying on cutting-edge chips. This raises questions about the value of the hundreds of billions of dollars being invested in the industry’s capital expenditures.

"The new DeepSeek model is incredibly impressive. They’ve effectively created an open-source model that excels in inference-time compute and is highly compute-efficient,” said Microsoft CEO Satya Nadella at the World Economic Forum in Davos, Switzerland, on January 22. “We need to take the advancements coming out of China very seriously.”

It’s worth noting that DeepSeek still trained its model using NVDA GPUs, albeit the less expensive H800s. Initially, NVDA stock rose on news of META’s increased capital spending but later reversed due to concerns from the Davos discussions. Investors are now questioning whether there’s a continued need for massive investments in the latest NVDA Blackwell GPUs if superior models can be achieved using more affordable alternatives like the H800s

Resource Utilization: DeepSeek has demonstrated an ability to train advanced AI models with significantly fewer resources than typically required. They managed to train their DeepSeek-V3 model using only about 2,048 NVIDIA H800 GPUs, which is a stark contrast to the 16,000+ GPUs often used by major U.S. tech companies.

Cost-Effectiveness: The total cost to train DeepSeek-V3 was reported to be around $5.6 million, much lower than the costs associated with training models by competitors, showcasing a remarkable efficiency in AI development.

$NVIDIA Corp(NVDA)$ 

$Micron Technology(MU)$ 

$Credo Technology Group Holding Ltd(CRDO)$  

$Vertiv Holdings LLC(VRT)$ 

$Astera Labs, Inc.(ALAB)$ 

@TigerStars  @Tiger_comments  @Daily_Discussion  @TigerObserver  @TigerPM