SUBSCRIBE TO OUR NEWSLETTER
Stay informed! Sign up to get expert advice and insight delivered direct to your inbox
DeepSeek Shifts Network Operators' View of AIDeepSeek Shifts Network Operators' View of AI
While DeepSeek might not have much impact in the U.S., its emergence has ramifications for the AI arms race as well as network architecture.
Brian T. Horowitz, Contributing Reporter
April 14, 2025
5 Min Read
CREDIT: GK IMAGES / ALAMY STOCK PHOTO
The Chinese startup, DeepSeek, could reshuffle how operators plan to use AI in their networks.
DeepSeek made headlines earlier this year when it released an open source large language AI model it claimed is more efficient and easier to train than U.S. platforms from OpenAI and others. In March, the company rolled out DeepSeek-V3, promising improved executability of code and a boost in benchmark performance over an earlier V3 model that came out in December. The new model requires less than $6 million worth of computing power from Nvidia H800 chips, Reuters reported.
DeepSeek could reduce the hardware and costs for training, but what does that mean for network operators?
A Move to the Edge
One change is that organizations could use DeepSeek to fortify their edge computing capabilities. The edge, along with the LAN, is where AI models learn, according to Ed Fox, CTO at MetTel, a global provider of integrated digital communications products for businesses and government agencies.
Usman Javaid, chief products and marketing officer at Orange Business, also sees DeepSeek-R1 models running on an edge node. The models are efficient and can run wherever users want, and he said he foresees a day when mobile phones that incorporate efficient chipsets will be able to run small AI models.
"DeepSeek has helped to make this concept of immersive AI real, which means you can take AI and put it as close as possible to your workload or to your data," Javaid said.
More Mathematically Efficient Models
DeepSeek can also enable organizations to create more mathematically efficient AI models that require less computation, Javaid said.
"Instead of having a brute-force approach to compute, you can take a model-efficiency approach," Javaid said. "Then eventually, with less amount of compute, you can get to similar or better outcomes."
This, however, comes with a caveat: Models with more reasoning require less compute to train them, but compute will still be required for inference.
"You need a lot of compute to infer from the models, because a lot of processing is happening at the inference level, at the reasoning level," he said. "So how you look at your total cost of ownership, not just in the cost of training, but actually your course of inference, is way higher."
DeepSeek for Research
Today, DeepSeek is mainly used for research purposes. Companies experiment by using machine learning and AI to connect to multiple AI engines and gain a better understanding of different learning models, according to Jim Coyle, U.S. public sector CTO at Lookout, a mobile endpoint and cloud security company.
"Especially when it comes to the U.S. -- because of U.S.- China relations -- I just don't see [DeepSeek] being utilized in an everyday business world aspect beyond research," Coyle said.
Indeed, the U.S. Congress introduced a bipartisan bill to ban DeepSeek. Some U.S. agencies, such as the Department of Defense, NASA and the Department of Commerce already instituted a ban. Meanwhile, states that don't allow DeepSeek include Alabama, Iowa, New York, Oklahoma, Texas and Virginia.
Andrew Athan, technical solutions architect III at World Wide Technology, an IT systems integrator, echoed Coyle's contentions.
"DeepSeek in and of itself will have limited impact, primarily because both the company and its models are not U.S.-based, and therefore have various security and national security concerns attached," he said.
Instead, he said, "resource-constrained" DeepSeek might be better suited for use by universities and other similar institutions.
"DeepSeek's approaches are helpful to many ecosystem participants, including in academia, which has similar resource constraints, and where model sizes used for research are dwarfed by industrial-scale model developers," Athan said.
Security Implications
Organizations that decide to roll out DeepSeek must tread carefully to ensure information is secure. It's crucial to implement zero trust and a uniform policy dictating how material is encrypted while in transit, Javaid said.
"There is a need for perhaps a more common and consistent underlying layer of trust, which includes security responsibility -- not just per model, but an underlying layer that assures compliance," he said.
MetTel's Fox said it's also important that companies accurately identify, track and report on traffic coming from an app such as DeepSeek. Coyle advised companies to use caution when blocking AI tools, as that could result in a denial-of-service attack. Additionally, tools like DeepSeek might further drive organizations to adopt zero trust and the concept of data sovereignty, in which rules and regulations govern data depending on the operating region.
A New Type of Network?
Ultimately, DeepSeek and models like it could usher in a new type of network infrastructure, Javaid said, one that is more distributed and less centralized.
"You can even run your models in your network, routers and switches," Javaid said. "And then you can also create intelligence at the prompt and application level to be able to have your network have the same reasoning capability as you have in your generative AI models."
A new generation of efficient open source AI distributed models will eventually spur network operators to consider new, emerging types of architecture, he said.
DeepSeek's evolution is just the most recent iteration in the larger AI arms race. As organizations develop their own AI tools, they might want the smaller footprints DeepSeek offers.
"The most capable systems consist of collections of smaller models, each tuned to perform optimally for a specific end task," WWT's Athan said.
Going forward, AI platforms will be propelled to distinguish their offerings from DeepSeek and incorporate features not found in the Chinese platform, said Nic Benders, chief technical strategist at New Relic.
"In the coming months, as we see other AI companies copying what DeepSeek has done, we should see more capable models and lower usage prices across the board," Benders said.
About the Author
You May Also Like
