The Chinese startup Deepseek recently found a central place in the world of technology with surprisingly low employ of computing resources for the advanced AI model called R1, a model that is considered competitive in the case of O1 O1 AI, despite the company’s claims that Deepseek cost only $ 6 million and 2048 GPU for training. However, the company is an industry analytics Semianization He reports that the company standing behind Deepseek incurred $ 1.6 billion and has a 50,000 GPU NVIDIA Hopper Fleet, which undermines the idea that Deepseek for the fresh AI training and inference with dramatically lower investments than the leaders of the AI industry.
The report claims that Deepseek conducts extensive computing infrastructure with about 50,000 GPU Hopper. This includes 10,000 H800 and 10,000 H100S, with additional purchases of H20 units, according to semianization. These resources are distributed in many locations and serve, such as AI training, financial research and modeling. The company’s total capital investment in servers is about $ 1.6 billion, and about $ 944 million has spent on operating costs, in accordance with semianization.
Deepseek conquered the world of AI when he revealed the tiny hardware requirements of its AI deep V3 (MOE) mixture, which is much lower compared to the requirements of American models. Then Deepseek shook the world with an advanced technologically open model AI AI. However, the renowned semi -spatial Intelligence Company revealed its arrangements, which indicate that the company has about $ 1.6 billion of hardware investments.
Deepseek comes from High-Flyer, the Chinese Hedge Fund, which he accepted AI early and invested heavily in the GPU. In 2023, the high Fllyer fired Deepseek as a separate undertaking only focused on artificial intelligence. Unlike many competitors, Deepseek remains self -financed, which gives flexibility and speed of decision making. Despite the claim that this is a tiny branch, according to semianalysis, the company has invested over $ 500 million in its technology.
The main distinguishing feature of Deepseek is his ability to conduct his own data centers, as opposed to most other AI startups, which are based on external cloud suppliers. This independence allows full control over experiments and optimizations of the AI model. In addition, it enables quick iteration without external bottlenecks, thanks to which Deepseek is highly effective compared to established players in the industry.
There is also something that cannot be expected from a Chinese company: taking over talents from continental China, without poaching from Taiwan or USA Deepseek only employed from China, focusing on the skills and abilities of problem solving, not formal certificates, according to semianization. Recruitment efforts are addressed to institutions such as Peking University and Zhejiang University, offering highly competitive salaries. According to research, some AI researchers from Deepseek earn over $ 1.3 million, exceeding compensation in other leading Chinese AI companies, such as the moon.
Due to the influx of talents, Deepseek is a pioneer of innovation, such as multiple, hidden attention (MLA), which required months of development and significant employ of the GPU, reports semianization. Deepseek emphasizes the efficiency and algorithmic improvements in relation to the scaling of brutal strength, transforming expectations regarding the development of the AI model. This approach for many reasons led some to the belief that rapid progress can reduce the demand for high -class graphic processors, affecting companies such as NVIDIA.
The recent claim that Deepseek has trained its latest model for only $ 6 million, fueled a significant part of the noise. However, this number refers only to the part of the total training cost-in particular the GPU time required for initial training. It does not take into account research, improving models, data processing or general infrastructure expenses. In fact, Deepseek has spent over $ 500 million on the development of AI from the very beginning. Unlike larger companies burdened with bureaucracy, the slim Deepseek structure enables aggressive support in AI innovation, believes that semianization.
The creation of Deepseek emphasizes how a well -financed, independent company AI can challenge industry leaders. However, public discourse could be driven by noise. The reality is more convoluted: semianization claims that the success of Deepseek is based on strategic investments of billions of dollars, technical breakthroughs and competitive working strength. This means there are no miracles. As Elon Musk noted about a year ago, if you want to be competitive in artificial intelligence, you must spend billions a year, which is supposedly in the released scope.
