DeepSeek’s Cost Claims Under Fire: New Analysis Accuses Chinese AI Firm of Misleading Investors

A recent deep-dive report has ignited controversy in the artificial intelligence community by alleging that DeepSeek, the high-flying Chinese AI startup, has intentionally misrepresented key details about its training costs and operational expenses. The explosive claims, published by industry analysts at SemiAnalysis, suggest that the firm’s touted “$6M training cost” for its flagship V3 model is misleading—concealing hundreds of millions in hardware investments and R&D expenditures.

The $6 Million Myth

DeepSeek recently made headlines by claiming their V3 model cost only $6 million to train, a figure that has now been exposed as highly misleading. According to detailed analysis from SemiAnalysis, this number represents only a fraction of the actual investment required to develop the model.

“The pre-training number is nowhere near the actual amount spent on the model,” states the report. “We are confident their hardware spend is well higher than $500 million over the company history.”

The True Cost Breakdown

The investigation reveals several critical costs that DeepSeek omitted from their public statements:

• Research and Development expenses
• Total Cost of Ownership (TCO) for hardware
• Employee salaries, including competitive compensation packages reaching $1.3 million USD for top talent
• Extensive GPU infrastructure, including approximately 50,000 Nvidia Hopper GPUs
• Server CapEx totaling approximately $1.6 billion
• Operational costs of $944 million for running computing clusters

Infrastructure and Resources

DeepSeek’s actual computing infrastructure is substantially larger than previously reported. The company maintains access to:

• 10,000 H800 GPUs
• 10,000 H100 GPUs
• Additional orders for H20 GPUs
• Shared resources with High-Flyer, their parent company

Industry Implications

The revelation of DeepSeek’s misleading cost claims raises important questions about transparency in AI development costs. While DeepSeek has made impressive technical achievements, including innovations in Multi-Token Prediction and Multi-head Latent Attention, their attempt to downplay the true investment required for state-of-the-art AI development has drawn criticism from industry observers.

For context, other major AI labs like Anthropic spend tens of millions of dollars on training individual models, with total development costs running into the billions when accounting for all associated expenses.

“This is akin to pointing to a specific part of a bill of materials for a product and attributing it as the entire cost,” notes the analysis. “The pre-training cost is a very narrow portion of the total cost.”

Industry experts suggest that DeepSeek’s misleading cost claims may be part of a strategy to gain market share, potentially offering services at below-cost prices while obscuring the true scale of their investment and operations.