DeepSeek has become an overnight sensation, rattled the US AI sector and may have single-handedly focused CxOs on cost of genAI.

We convened a call of Constellation Research analysts to outline the issues CxOs need to know about when it comes to DeepSeek.

Here are some recent headlines what you need to know about DeepSeek.

What is DeepSeek?

DeepSeek is a Chinese AI company that develops open-source large language models. It has launched a series of models that can compete with the likes of OpenAI's ChatGPT, Anthropic's Claude family of models and Meta's Llama. Constellation Research CEO Ray Wang said DeepSeek has "democratized the access to AI." Wang noted that the other benefit is that the model can run in private environments without top-of-the-line hardware. DeepSeek is censored as anyone who has asked the service about Winnie the Pooh or other sensitive topics in China.

What's the big deal about DeepSeek?

The hubbub surrounding DeepSeek in a nutshell is that the company "proved a point that you don't need gazillion dollars to train AI model," said Constellation Research analyst Andy Thurai.

DeepSeek has "proven not only that you can find cheap but also the fact that you can open source the entire thing, which means others can start using it or building it, which is going to challenge all those big guys," said Thurai.

DeepSeek also garnered a lot of attention because Wall Street decided a week after its latest model release that perhaps Nvidia customers didn't need the latest and greatest GPUs.

What did DeepSeek do that was different?

Holger Mueller, analyst at Constellation Research, said:

"Not having the best computing resource always makes for better models and software. China doesn't have availability of so many GPUs and people get creative. The distillation really worked. The second really important thing is that DeepSeek has been training about human intervention."

What's unclear is how much DeepSeek piggybacked off of larger models and IP from around the globe. "It's going to be interesting to see what kind of IP battle is going to unfold," said Thurai.

What should CXOs do?

For now, it's best to monitor DeepSeek, think through use cases and if you experiment make sure it's air gapped and sandboxed. Don't ignore the DeepSeek developments though. Constellation Research analyst Chirag Mehta said:

"If you're a CxO, the best analogy is what open source did to the industry. That's what this model is now doing to its competitors.

If you're a CxO, you have two options: Buy the Ferrari in the high-end platform as a service model or do smaller, specialized, narrow models that are cheaper to run, and almost free. Open source is not quite free. You still have to manage it, and you still have to run it, and you have to maintain it."

Mehta said CxOs need to keep their model options open and stay focused on what problem you're trying to solve with genAI.

Wang said focus on the cost curve. Wang said:

"At this point, we know that it's possible to do reasoning at a lower cost and lighter models are going to be available. We know that people are going to want to do this outside of the cloud and back on premises. The cost curve is coming down on AI, and I think you're going to see more of that. And I think those monetization models are important."

Will DeepSeek mean on-premises AI?

The jury is decidedly mixed on this one. Holger Mueller said AI workloads will reside in the cloud for the most part. "I see still larger models winning and cloud winning. If you might see a dip in revenue. That's totally possible," said Mueller.

Mehta noted that on-prem vs. cloud AI isn't zero sum, but the majority of workloads will go to the cloud with the exception of edge computing use cases.

Should Wall Street be this concerned about DeepSeek?

Thurai said concerns are overblown. If you are building an LLM or using one for inferencing you're still likely to use a Nvidia stack. Where it gets interesting is if DeepSeek used AMD GPUs. "This is a knee jerk reaction and it's going to continue for a while," said Thurai.

Wang said the concerns are more about big spending tech giants and whether the capex will be questioned. He said:

"We have to figure out if it makes sense for a Microsoft to spend $80 billion a year on capex to build out data centers. The short answer is that Microsoft is has to do it. It's really about the payback period that that's going to actually hit them. The second question is whether we need to pay this much for token economics.

"We're living in a world we call exponential efficiency. If you're not 10 times better one tenth the cost, nobody cares. And we're at this point where our existing software vendors have made life so expensive to hold their stock price. So this reset is a good thing in general, because it's going to lower the cost of technology for customers. It's a bad thing for stock investors, because we're going to see valuations at the top plummet if you're not one of the winners."

More:

What are the security concerns with DeepSeek?

Mehta said there are concerns about prompt injection and jailbreaking DeepSeek. "AI security is one of the biggest topics for CxOs," said Mehta. "If you don't know how the model has been trained, what data has been used, and how easy or difficult it's going to be to actually break it, do you really want to use that model for your most sensitive data and use cases? Are you really going to do that?"

Business Research Themes

Data to Decisions

Roles

Chief Information Officer, Chief Data Officer, Chief Executive Officer, Chief Financial Officer, Chief Information Security Officer, Chief Technology Officer

Tags

#AI #artificialintelligence #machinelearning #ML #generativeAI