The Chinese startup Deepseek recently took the scene in the world of technology with its surprisingly weak use of calculation resources for its advanced AI model called R1, a model which is considered competitive with the O1 of AI Open despite the allegations of the company that Deepseek costs only $ 6 million and $ 2,048 to train. However, the industry analyst company Semianalysis The reports according to which the company behind Deepseek has initiated $ 1.6 billion in material costs and has a fleet of 50,000 GPU Nvidia Hopper, an observation which undermines the idea that Deepseek reinvented training and inference with investments considerably lower than AI managers.
Deepseek operates a large IT infrastructure with around 50,000 hopper GPUs, says the report. This includes 10,000 h800 and 10,000 h100, with additional purchases of H20 units, according to semiianalysis. These resources are distributed on several locations and serve objectives such as training, research and financial modeling of AI. The total capital investment of the company in the servers is around $ 1.6 billion, with around $ 944 million spent on operating costs, according to semiianalysis.
Deepseek drew the attention of the world of AI Tiny material requirements of its AI model of the depth-V3 mixture (MOE) which are much lower compared to those of models based on the United States. Then Deepseek rocked the world of high technology with an open competitive R1 model. However, semianalyysy semiianalysis compartmentalistes for the lessee of Market revealed its results which indicate that the company has approximately $ 1.6 billion in material investments.
Deepseek comes from High Flyer, a Chinese coverage fund that adopted early AI and strongly invested in GPUs. In 2023, High-Flyer launched Deepseek as a distinct company only focused on AI. Unlike many competitors, Deepseek remains self -funded, which gives it flexibility and speed in decision -making. Despite the claims that this is a minor driving, the company has invested more than $ 500 million in its technology, according to semianalysis.
A major DEEPSEEK differentiator is its ability to manage its own data centers, unlike most other AI startups based on external cloud suppliers. This independence allows total control over the experiences and optimizations of the AI model. In addition, it allows rapid iteration without an external bottleneck, which makes Deepseek very effective compared to traditional players in industry.
Then, there is something that we would not expect from a Chinese company: the acquisition of talents from continental China, without poaching of Taiwan or depth Éntoises in the United States in China, concentrating On skills and problem solving capacities rather than formal references, according to semianalysis. Recruitment efforts target institutions such as the University of Peking and Zhejiang University, offering very competitive wages. According to research, some researchers from Deepseek AI earn more than $ 1.3 million, going beyond compensation in other leading Chinese AI companies such as Moshot.
Due to the influx of talents, Deepseek launched innovations such as multi-head latent attention (MLA), which required months of development and the substantial use of GPUs, reports semi-analysis. Deepseek emphasizes algorithmic efficiency and improvements in relation to scaling up by brute force, reshaping expectations concerning the development of the AI model. This approach has, for many reasons, some to believe that rapid progress can reduce high -end GPU demand, an impact on companies like Nvidia.
A recent statement that Deepseek has formed his latest model for only $ 6 million has fueled a large part of the media threshing. However, this figure refers only to part of the total cost of training – in particular, the GPU time required for pre -training. It does not take into account the research, the refinement of the model, the processing of data or overall infrastructure. In reality, Deepseek has spent much more than $ 500 million for the development of AI since its creation. Unlike large companies overwhelmed by the bureaucracy, the Lean structure of Deepseek allows it to advance aggressively in AI innovation, estimates semianalysis.
Rise de Deepseek underlines how a well -funded independent AI company can challenge industry leaders. However, public speech may have been motivated by the beaten media. The reality is more complex: semianalysis maintains that Deepseek’s success is based on strategic investments of billions of dollars, technical breakthroughs and a competitive workforce. This means that there are no wonders. As Elon Musk noted it about a year ago, if you want to be competitive in AI, you have to spend billions a year, which would be in the range of what has been spent.