DeepSeek, Explained: Why Those Hardware Stocks Dropped

A friend asked me: “Hey Mike! What’s up with the news about this DeepSeek thing? And why did some hardware stocks drop so much the other day?” Let’s break it down.

What is DeepSeek?

It’s a new generative AI model out of China. Depending on the test and use cases, initial independent evaluations show DeepSeek is comparable to established generative AI models being used today in accuracy and functionality. And it is more efficient than current models; it can run with less overhead, and can even run on some powerful laptops, without needing servers. Because of that efficiency, some hardware stocks dropped, with some expecting that those companies won’t have to pump out as much hardware as originally anticipated (see the last question for more on that).

Is the efficiency gain that new?

There was always a belief that the transformer concept underlying generative AI could be made more efficient than what’s currently being used. Most providers have been more concerned about the content and the power of GenAI to this point, not the efficiency. Now that the cat is out of the bag, I’d expect others to follow suit and continue the efficiency gains. It’s similar to Moore’s Law: technology over time trends to smaller, faster, cheaper, better.

I’m hearing that DeepSeek is pro-China?

GenAI systems in general have an inherent bias based on their training data, the same way history books are written by the victor. DeepSeek gives pro-China answers. In fairness, AMER- and Europe-based GenAI systems give answers that fit the beliefs of those creators.

Does DeepSeek take your data?

Users should know that if you run the DeepSeek model off their website, the initial read is that DeepSeek can take and store your queries for processing and training. If you download the DeepSeek model and run it locally, that concern is mitigated (but local models in general need more security, monitoring, and more for safe running). It should be noted that OpenAI struggled with data usage at their onset. (Remember when Italy blocked ChatGPT on privacy grounds?) OpenAI quickly added privacy controls to give users more fine-grained control over the use of their data, based on the Italian government’s (and others’) feedback.

Now about the hardware: does this efficient model reduce the need for GPUs?

Anyone’s guess is as good as mine. My take: it depends on the balance of efficiency vs. demand. Yes, transformers will get more efficient over time, which will reduce the hardware needs. But we as users will expand the use, power, and scale/concurrency of running GenAI, so that could increase the need for hardware. We’ll also add GenAI (if we so desire) to “the edge,” more places than just data centers, perhaps machinery running models locally, which requires more hardware.