Skip links

Inside Markets — DeepSeek

DeepSeek

January 27, 2025

DeepSeek is a ~2 year-old Hangzhou-based, venture-funded AI R&D startup that operates under its parent company ‘High-Flyer,’ which is a quant fund.  DeepSeek claims its distilled R1 model performs better at some tasks than GPT4 and Claude 3.5 Sonnet at a fraction of the cost ($2.19 per million output tokens on the API vs. $60 for OpenAI o1). Over the weekend, DeepSeek briefly became #1 in Top Free Apps on Apple’s iTunes store this morning but is now temporarily limited to customers with China-based phone numbers only. Weekend reports focused on how the company was able to obtain strong results given US restrictions on advanced GPU sales to China, but the company’s parent, High Flyer’ is believed to have a 50,000 H100 cluster, which is what GPT5 is believed to be training on.  The company claims to be using a novel form of ‘reinforcement learning’ that has lowered the cost of training compute but there’s even more excitement around its potential to reduce inferencing costs.

DeepSeek’s accomplishments are shining a brighter light on preexisting concerns around AI spend and ROI.  While it remains unclear how much cheaper it will actually make training and inference, the model’s cheaper API costs will likely create some industry-wide pricing pressures.  From a market perspective, the development has introduced a cyclical component to a very strong secular growth story that likely resulted in market imbalance.

Read more