Ideas beneath
the surface
What Does DeepSeek Mean for AI?
What Does DeepSeek Mean for AI?
Not as much as the market would have you believe.
This past Monday, most publicly traded AI companies saw a sharp devaluation from the emergence of a new AI model called DeepSeek R1. DeepSeek is a Chinese AI research company, and the story was that they created their AI model for far less than American companies building comparable models.
The prevailing narrative was that DeepSeek proved that building and operating AI models would be far cheaper than previously expected. Further, a significant reduction in the investment required for AI infrastructure would threaten the companies benefiting from that buildout.
While DeepSeek did discover some new techniques to improve efficiency in several ways, they did not, in our opinion, redefine how frontier AI models will evolve:
- The all-in cost to train DeepSeek’s top-end model was almost certainly more than the $6 million that’s been reported and shared widely.
- It seems highly likely that DeepSeek used a process called “distillation” whereby they copied OpenAI’s work that did require significant compute and resources.
- DeepSeek didn’t release a model that moved the bar for performance ahead. They merely met existing model performance more efficiently.
While DeepSeek did discover some new techniques to improve efficiency in several ways, they did not, in our opinion, redefine how frontier AI models will evolve.
In our view, frontier AI models will continue to move forward dependent upon access to compute, large amounts of data, and model improvements. That was our view before DeepSeek and that remains our view today.
What’s new is that DeepSeek R1 offers a solid reasoning model at a significantly lower cost than current models. This means AI will quickly get cheaper for all of our portfolio companies, accelerating their ability to use AI to deliver great experiences to customers.
From a thesis standpoint, we’ll continue to look for alien founders solving important and challenging problems with AI. And we’ll continue to benefit from AI models getting better, driven largely by US innovation and reliant on US infrastructure.