Investors went wild this week over DeepSeek, a startup based in Hangzhou, China, which has an AI assistant that looks and acts like OpenAI's ChatGPT but costs far less.
The release of the latest version of DeepSeek's R1 model triggered a sharp sell-off in AI-related stocks, particularly Microsoft and Nvidia. DeepSeek surpassed ChatGPT to become the most popular free app in the Apple App Store.
Banks considering using DeepSeek will need to sit this one out, at least for now, amid security, data privacy and Chinese propaganda concerns.
On Tuesday, DeepSeek reported that "large-scale malicious attacks on DeepSeek's services" prevented it from registering new users. According to the
Data privacy and data security are big question marks for DeepSeek. According to
"The geopolitical and potential security risks of working with DeepSeek will keep bankers away," said Alenka Grealish, principal analyst at Celent.
But DeepSeek's innovations, which have been questioned by some, could help bring down the cost and energy consumption of generative AI models going forward.
"It's not often that a model is released and the Nasdaq falls 5%," said Krishnan Swamy, chief data and analytics officer at Citizens Bank. "I think this tells you what this could mean for the AI landscape, and I think there's a positive from a user standpoint, because it lowers the cost barrier to entry."
The risks
Gilles Ubaghs, strategic advisor, commercial banking and payments at Datos Insights, said there are a lot of unknowns behind DeepSeek's developers, "including how accurate their reports of training really are, as well as deeper questions of who are they and what limits do they have on absorbing enterprise data?"
"Even if a bank did want to work with DeepSeek, the compliance risks are at best unclear, and at worst likely to be severe," Ubaghs said.
Swamy said before a model like DeepSeek's could find a home in a bank like Citizens, it would have to go through all of the security scans any open source software goes through, including tests of data privacy and security.
"We have scans around the presence of malware, suspicious code, license checks," he said. "DeepSeek would need to go through those checks, no question."
Banks are expected to look at this type of innovation and try to understand it, Swamy said. At the same time, banks must comply with regulations governing data privacy and security.
"I think data privacy and data security concerns are heightened when the model is hosted by DeepSeek, OpenAI, Microsoft or whoever it might be," he said. "I think they are significantly diminished when these models are brought into parameters of our cloud environment or on-prem environment."
This will be easier for large banks than for small ones, Swamy said.
Most banks lack the model validation frameworks to properly test models like DeepSeek's, according to Ryan Cox, global head of artificial intelligence at Synechron.
"These are non-deterministic models," Cox said. "You put some information in and you get some information out, but it's effectively an algorithm with a weighted probability of an answer, and that answer could be right, it could be wrong.
"That's why it's really important to have that model validation in place, because then you can do things like testing, you can run your normal software checks, you can qualitatively say, do the answers I'm getting back make sense?"
Such checks also include looking for guardrails that prevent a model from generating answers that are violent, harassing, ethically wrong or politically skewed.
Lower barrier to entry
The idea of lower-cost artificial intelligence holds great appeal to companies of all kinds.
"What companies have been discovering for the past year and a half is that while the big commercial models — OpenAI, Anthropic and so on — are very, very powerful, they are also expensive," Citizens' Swamy said. "All of these companies made big investments into these models, and there's got to be a way in which they recoup those investments."
To counter the high cost of AI, some companies have started using open-source models like Llama, he said. Others have turned to smaller models and fine-tuned them for specific applications. Smaller models use fewer tokens and consume less computing resources, so they end up being more economical, Swamy said.
Companies like DeepSeek are changing the whole AI landscape, according to Beerud Sheth, CEO of Gupshup, a conversational AI agent provider that lists Citi as a customer.
"As AI is developing, especially in the open-source ecosystem, people are finding newer techniques, newer algorithms, newer methods to make it faster, better and cheaper," Sheth said. "That's been the quest from forever." He cited the transformer architecture Google came out with in 2017 as an example.
"Once that technique came out, well, guess what? The whole world knows about it, and OpenAI and a few other engineers were among the first ones who used that technique," Sheth said. "After that, there were many other developments."
The techniques DeepSeek's researchers came up with that lower the cost of training generative AI models are out in the public domain for other companies to copy and get similar results, Sheth said. DeepSeek has published a research paper about its technology, which is open source.
"Nobody ever says, 'Oh, I'm not going to adopt a technique because it came from somebody I don't trust,' because a technique is a technique, and you just re-implement it," Sheth said.
"This is such a significant paradigm shift that there is no question that everybody has to be testing, tinkering, iterating with it right now, in general," he said.
