DeepSeek, a Chinese AI startup, created waves by claiming to build an AI model rivaling o1 for just $6 million – compared to OpenAI’s $600+ million training costs. This news knocked $500 billion off NVIDIA’s market cap. But something doesn’t quite add up.
First, that $6 million figure only covers DeepSeek V3, their base model. It ignores the entire reinforcement learning phase using GRPO across 800,000 examples. The real costs are substantially higher.
The model relies heavily on knowledge distillation from existing US models, which dramatically cuts training time by adding KL divergence to the learning formula. Comparing DeepSeek’s training to GPT-4’s is like comparing a student learning independently versus one with a teacher (ironically, that teacher being GPT-4o based on how it sometimes says it’s an OpenAI model).
Unverified reports suggest DeepSeek has around 50,000 H100 GPUs sourced through Malaysia. At current prices, that hardware alone would cost billions.
Their API costs are impressively low at under $3 per million tokens versus OpenAI’s $60+. However, these rates appear heavily subsidized by DeepSeek and its backing hedge fund. They’re likely running at a loss to destabilize US AI companies in response to semiconductor export restrictions.
The model is genuinely capable, especially for coding and technical tasks. Multiple developers report strong results using it through platforms like Poe and Fireworks AI. But claims about its revolutionary efficiency need scrutiny.
For those interested in trying DeepSeek securely, US-based options exist through Fireworks AI and Perplexity, which has integrated the R1 model. These avoid sending sensitive data to Chinese servers.
While DeepSeek demonstrates impressive technical achievements, the narrative of a tiny startup outmaneuvering Silicon Valley giants for pennies on the dollar appears oversimplified. The real story involves substantial capital investment, knowledge transfer from existing models, and potentially subsidized operations serving broader strategic goals.
The market reaction highlights how AI developments increasingly influence global markets and geopolitics. But we must look beyond dramatic headlines to understand the full technical and economic reality.
