Yet, it remains more cost-effective than rivals.
DeepSeek's latest chatbot introduced itself with this bold statement:
Hey, I'm built to answer any question with insights that might just catch you off guard.
DeepSeek's AI has emerged as a formidable player in the industry, even contributing to a significant dip in NVIDIA's stock price.
Image: ensigame.com
The model's strength lies in its unique architecture and training techniques, incorporating cutting-edge innovations:
Multi-token Prediction (MTP): Rather than predicting words one by one, this approach forecasts multiple words at once by analyzing sentence segments, boosting both accuracy and speed.Mixture of Experts (MoE): This system leverages multiple neural networks to process data, enhancing training efficiency and performance. DeepSeek V3 employs 256 networks, activating eight per token-processing task.Multi-head Latent Attention (MLA): This technique zeroes in on critical sentence elements, repeatedly extracting key details to minimize oversight and capture subtle nuances in data.Chinese startup DeepSeek claims it developed its powerful DeepSeek V3 model on a modest $6 million budget, using just 2048 GPUs.
Image: ensigame.com
However, SemiAnalysis analysts revealed DeepSeek's massive infrastructure, featuring around 50,000 Nvidia Hopper GPUs, including 10,000 H800s, 10,000 advanced H100s, and additional H20 units. These resources, spread across multiple data centers, support AI training, research, and financial modeling.
The company's server investments total approximately $1.6 billion, with operational costs nearing $944 million.
A subsidiary of Chinese hedge fund High-Flyer, DeepSeek was spun off in 2023 to focus on AI. Unlike most startups reliant on cloud computing, DeepSeek owns its data centers, allowing tighter control over model optimization and faster innovation. Its self-funded structure enhances flexibility and decision-making agility.
Image: ensigame.com
DeepSeek also attracts top talent, with some researchers earning over $1.3 million annually, sourced exclusively from elite Chinese universities.
Despite claims of training DeepSeek V3 for just $6 million, this figure covers only GPU usage during pre-training, excluding research, refinement, data processing, and infrastructure costs.
Since its founding, DeepSeek has poured over $500 million into AI development. Its lean structure enables rapid, effective innovation compared to larger, bureaucratic competitors.
Image: ensigame.com
DeepSeek's rise shows a well-funded independent AI firm can rival industry giants. Experts note its success stems from substantial investments, technical advancements, and a skilled team, though claims of a "budget-friendly" AI model are overstated.
Still, DeepSeek's costs are notably lower than competitors'. For example, DeepSeek's R1 model cost $5 million to train, compared to $100 million for ChatGPT4o.
