The latest model from Chinese AI startup DeepSeek has dominated conversations this past week. The company claimed its R1 model performance rivaled leading American-made models at a fraction of the training cost.
The announcement sent shock waves through the AI provider space. Executives from IBM, SAP, Anthropic, OpenAI, ServiceNow, Meta, Microsoft, Nvidia and other notable vendors have already weighed in.
DeepSeek’s R1 isn’t breaking new ground in terms of capabilities or greatly exceeding the performance of existing models, according to analysts, but it does challenge conventional wisdom about what it takes — in dollars and methods — to build these systems.
“A lot of organizations have already invested in so many GPUs to improve and increase infrastructure capacity to support these models … so that is causing panic,” said Haritha Khandabattu, senior director analyst at Gartner. “For CIOs, what they need to actually do is not overreact.”
Enterprises poured countless resources into their AI initiatives, so the opportunity to significantly lower costs could be quite alluring to other C-suite members and the board. CIOs will have to work to set expectations.
Despite the promise of lower cost, security and privacy concerns are coming into focus as businesses evaluate the model provider.
Enterprises can start assessing how to build model-agnostic platforms if they haven’t yet. Overreliance on a single model provider will make it harder for organizations to capitalize on innovations.
CIOs interested in trying out DeepSeek models should also assess whether existing guardrails and governance measures will protect the business or if new controls are needed, Khandabattu said.
Technology leaders can leverage the interest in DeepSeek to explore more affordable AI models and tools. Costs have already been on a downward trajectory since the race to deploy the technology kicked off. Accenture tracked a 74% annual cost decline between OpenAI’s GPT-3 and GPT-3.5 Turbo in three years. GPT-4’s cost dropped 58% between March 2023 and the end of 2024, the firm found.
DeepSeek models are available in Microsoft’s Azure AI Foundry and GitHub. Snowflake added the model to its Cortex AI platform as part of private preview Thursday. Enterprise customers can also access the models on Amazon’s Bedrock.
Staying risk-aware
The hype around DeepSeek has made its way to dinner tables and water coolers, which can have its downsides for business leaders.
CIOs should be aware of employee curiosity turning into instances of shadow AI. Enterprises should prioritize communicating with workers about the dangers of using the models without approval.
The privacy and data security concerns are not unfounded. Threat actors exposed a “significant volume” of chat history, sensitive information and operational details tied to DeepSeek’s database earlier this week, according to Wiz Research. Cybercrime monitoring and analysis company Kela also highlighted several security flaws in a report published Monday.
DeepSeek’s privacy policy gives cause for pause, too. The document states the company uses customer data for training its technology, allowing the startup access to user keystroke patterns, audio, text and more.
Italy’s data privacy watchdog sent a request for information to DeepSeek earlier this week, citing the potentially high risk for millions of people’s data in the country. DeepSeek’s app could not be accessed in app stores shortly after, according to Reuters. DeepSeek has 20 days to respond to Italy’s request. The country was one of the first nations to take action against ChatGPT-maker OpenAI during its meteoric rise, too.
There is also little information about DeepSeek’s responsible AI practices.
“Their research is all alleged, and yes, there are strengths … but in a real business setting, we don’t know how that’s going to play out,” Khandabattu said. “The total cost of deploying any kind of generative AI application is not just the cost of the models.”
CIOs learned the hard way last year that there are a lot of steps in between choosing an AI model and realizing value. Skepticism is a valuable tool for business leaders to deploy as new claims arise.
“The newer, cheaper models are great as they unlock more use faster, but they still don’t know anything about you — your customers, products or the problems you solve,” Collibra’s Chief Data Citizen and Co-Founder Stijn Christiaens told CIO Dive via email. “When the models are all the same, your organization’s moat is your data, and you need to combine the foundational models with relevant data, especially high-quality, governed data, to ensure data confidence.”
