DeepSeek R1 and DeepSeek AI Statistics, Facts and Trends (2025)

Key DeepSeek R1 and DeepSeek AI Statistics, Facts and Trends for 2025

• DeepSeek R1 is priced at 1/30th of similar OpenAI models. It costs $2.19 per million output tokens versus OpenAI’s 01 model which costs $60.00 per million output tokens. (Source)
• DeepSeek R1 is currently number #1 in the Apple App Store. It took Deep Seek R1 5 days to take the top spot in Apple App Store and get ahead of ChatGPT app (currently #2). (Source)
• Currently 6 million+ people use DeepSeek R1, after only a week of it going live. (Source)
• There are currently 3 million+ DeepSeek R1 app downloads. (Source)
• Deep Seek AI is only 20 month old (founded May 2023 in China). DeepSeek launched as a spin-off from High-Flyer hedge fund, prioritizing fundamental AI research over quick profit—similar to early OpenAI which once was a true nonprofit company. (Source)
• High-Flyer’s peak at ~$15 billion in assets gave DeepSeek robust funding. This enabled the team behind DeepSeek AI to do high-level experimentation without immediate pressure to generate revenue. (Source)
• There are ~200 employees at DeepSeek compared to OpenAI’s 3,500. (Source)
• DeepSeek R1 offers models that go up to 671B parameters easily rivaling the best models from OpenAI and Anthropic AI. (Source)
• The cost to train DeepSeek R1 was ~1/10 of comparable Western models. (Source)
• The estimated cost to train DeepSeek R1 is only $5.5 million. This is dozens of times lower than the money need to train comparable Western AI models. For example, it took ~100 million to train GPT-4. (Source)
• Most DeepSeek R1 models are MIT-licensed. (Source)
• DeepSeek’s latest AI model triggered a global tech selloff, jeopardizing over $1 trillion in market capitalization. (Source)
• January 2025 launch of DeepSeek R1 drew international attention and caused a 13% pre-market drop to Nvidia stock). (Source)
• There are currently 68 research papers on arXiv. (Source)
• Chatbot Arena Rank for DeepSeek R1 is #4 with a score of 1357, just after ChatGPT-4o on the #3 place. (Source)
• DeepSeek relies on RL over extensive supervised fine-tuning, producing advanced reasoning skills (especially in math and coding). (Source)
• Multi-Head Latent Attention (MLA): This subdivides attention mechanisms to speed training and enhance output quality, compensating for fewer GPUs. (Source)
• DeepSeek preemptively gathered 10,000 Nvidia H100 GPUs , then focused on software-based efficiency to compete with larger Western labs when export controls from the US tightened. (Source)
• Predominantly Recent Graduates: Most DeepSeek researchers finished their degrees in the past two years, fostering rapid innovation through fresh perspectives and minimal corporate baggage. (Source)
• DeepSeek R1’s largest model boasts 671 Billion Parameters (DeepSeek-V3). It easily rivals top-tier Western LLMs, with far fewer costs to operate it. However, only a targeted set of parameters is activated per task, drastically cutting compute costs while maintaining high performance. (Source)
• The “R1-Distill” model compresses large models.. This makes advanced AI accessible to those with limited hardware. (Source)
• Countering US Export Controls: Despite chip embargoes, DeepSeek innovated with custom GPU communication and memory optimizations, challenging the policy’s effectiveness and US export controls and chip bans. (Source)
• DeepSeek’s continued push in RL, scaling, and cost-effective architectures has the potential to reshape the global LLM and AI market if the model proves a viable alternative to its wester counterparts. (Source)