- A new open source model from Chinese startup DeepSeek is giving OpenAI a run for its money
- The model has sent U.S. investors into a panic
- But analysts told Fierce the model's performance is actually good news for the AI industry at large
The global artificial intelligence (AI) community was rocked by a 9.5 earthquake this weekend when Chinese startup DeepSeek unveiled a new open source model capable of effectively competing with big name competitors at a fraction of the cost. But while the news prompted dire takes about U.S. AI dominance, Nvidia’s future and the bursting of the AI bubble, analysts told Fierce the model is actually good for AI proliferation.
“I would argue that the next phase of AI growth has just started,” Gartner VP Analyst Chirag Dekate told Fierce.
Why? Well, DeepSeek solves one key challenge that has been holding AI back: the ability to scale deployments.
Bubbling below all the recent hype were lurking concerns about whether or not the industry was “running into a brick wall from a scaling perspective,” Dekate said. But DeepSeek shows one of the “pathways that you would be able to continue engaging in scaling.”
As Dekate and AvidThink Founder Roy Chua both explained, there are three ways to scale AI beyond hardware innovations. You can add more compute, more data or change the model architecture. DeepSeek does the last of these three, basically showcasing better utilization of the underlying hardware resources.
The end result? A model that is “more than 25 times cheaper” than leading competitors on a per million token basis, Chua said. And, simply put, that is good for enterprises looking to deploy AI on the cheap.
“You now have yet another technique that model developers can use to improve scaling of the training of their models,” Dekate said. And “because they are using differentiated techniques you can actually now deliver inferencing at lower cost profiles. That basically means diffusion of AI products everywhere … DeepSeek sets a new floor from a pricing perspective.”
And while the markets have panicked about what this means for the likes of Nvidia, AMD and next-generation silicon startups, Chua said he doesn’t actually see the release of DeepSeek reducing the need for AI accelerators.
“I expect full speed ahead on all axes for model scaling for both companies and governments in the West and the East,” he predicted.
Indeed, Nvidia issued a statement on Monday praising DeepSeek, calling it a "perfect example of Test Time Scaling" which "illustrates how new models can be created using that technique, leveraging widely-available models and compute that is fully export control compliant." The company also took the opportunity to reiterate that model training aside, inference applications will still require "significant numbers" of GPUs.
Open source approach
Without getting too deep into the technical details, Dekate said DeepSeek’s latest model (called DeepSeek-R1) utilizes memory more efficiency by tapping into an FP8 data format rather than FP16, which consumes more memory. It further maximizes memory use through compression and key value caching.
Additionally, the model leverages a mixture-of-experts approach (aka it taps into several smaller expert subset models rather than trying to be one giant know-it-all) and selectively activates parameters as needed.
Most notably, the model is open source. Yet it still managed to deliver returns similar to closed models, like OpenAI’s o1.
Why does that matter? Well, even other “open” models like Llama come with restrictions that they can’t be used by direct competitors, Chua said.
Dekate said there are some downsides to DeepSeek’s open approach, at least from an enterprise perspective. More specifically, those developing external facing applications will need litigation risk protection and enterprise guardrails that open models generally lack.
Then again, Dekate noted that because DeepSeek is so open, every one of the techniques it uses can be replicated (if they’re not already being used) by other models with the protections enterprises need.
AI arms race
Of course, there’s also one other HUGE factor at play in the conversations happening about DeepSeek: politics.
China and the U.S. in particular are battling it out for AI supremacy. While the U.S. has sought to curtail China’s access to top-tier AI tech like certain GPUs and lure away AI investors, this approach doesn’t really appear to be effective. (We hate to say we told you so, but … we kind of did.)
“Creating scarcity can drive innovation and efficiency – assuming the details on the training costs are accurate,” Chua said. “China is adaptable and it's like going after a tennis player with a weak backhand repeatedly: they will improve and develop a very strong backhand at some point.
“While it doesn't mean the U.S. has lost the arms race, it does mean that China is closer than expected, and that they can be more efficient,” he concluded.