Patrick Moorhead, Chief Analyst at Moor Insights & Strategy, explains why the DeepSeek selloff was an overreaction
Category
🥇
SportsTranscript
00:00So tell me, you know, we saw a big panic in the tech sector after this low-cost Chinese
00:07artificial intelligence model from DeepSeek emerged on the scene. Can you break down why
00:13we saw such a big sell-off and why is there such a perceived threat from DeepSeek?
00:20Yeah, so first of all, any perceived risk into the bull case for NVIDIA, given its
00:26appreciation of its stock, is always met with a giant sell-off. And that's what we saw here.
00:35And the fundamental thesis or where there was perceived risk was that a company of a few
00:44hundred people in China was able to achieve what it appeared that others took billions to invest
00:53in like the open AIs of the world and they spent 5.6 million dollars to do that. So one might
01:01conclude that that means that people don't need as much NVIDIA infrastructure or any company
01:10affiliated or in around NVIDIA like an AMD, a Broadcom or a Marvell.
01:17So is what we saw an overreaction? I'm just asking that because there's some
01:23people who are questioning whether the validity of what's coming out of DeepSeek. We're already
01:28hearing news that Alibaba has an AI model that beats out DeepSeek. Microsoft is weighing in that
01:35maybe the way DeepSeek obtained this was through open AI to begin with. So where do you stand
01:42on all this? So where I stand on it after putting my team and myself grinding through white papers
01:52and talking and doing channel checks is that very little has changed. If you look at the wheel
02:01of large language models and the innovations that go into it, there's always a cost reduction effort
02:09or a cost reduction move. I don't believe that it costs 5.6 million dollars to train this model.
02:20It takes hundreds of millions of dollars to do this and then you actually have to run the AI
02:28which is lovingly referred to as inference and that takes a tremendous amount of money
02:35to stand that up. So our belief is that until we get to AGI at some point in the future
02:44for nearly every single task that we do and this could be five to ten years when you extend that
02:52to the edge and things like robotics, this infrastructure build out will continue. So
03:00it's also it's less about training it's mostly about inference. So I do believe that this
03:08was a little overdone, a lot overdone, but we're still seeing elements of risk and fear
03:17out there which is understandable given the run-up of NVIDIA.