How would a potential ban on DeepSeek impact enterprises?

Chinese AI startup, DeepSeek, has been facing scrutiny from governments and private entities worldwide but that hasn’t stopped enterprises from investing in this OpenAI competitor.

European regulators joined Microsoft, OpenAI, and the US government last week in independent efforts to determine if DeepSeek infringed on any copyrighted data from any US technology vendor. The investigations could potentially lead to a ban on DeepSeek in the US and EU, impacting millions of dollars that enterprises are already pouring into deploying DeepSeek AI models.

What happens when a ban is enacted on DeepSeek’s models?

Although the probe is still ongoing and the nature or extent of the ban is yet to be decided, experts believe that the ban may impact enterprises or any user in multiple ways, including loss of access, compliance risks, security concerns, data continuity issues, and migration.

“If the ban is enacted, cloud-based deployments on Azure, AWS, and Nvidia could be discontinued, potentially requiring urgent migration to alternative models,” said Anil Clifford, founder of UK-based IT consulting firm Eden Consulting.

“Enterprises using these models in sensitive applications could face legal and regulatory challenges,” Clifford said, adding that a ban could halt future updates, security patches, and model improvements, making DeepSeek R1 increasingly outdated and vulnerable over time.

Enterprises that had invested time, effort, and money into configuring the models might have to spend more time switching to alternative models requiring significant time and reconfiguration costs, Clifford further explained.

Some US organizations, such as the US Navy, Congress, and the state of Texas have already banned the use of DeepSeek’s models. 

Experts are divided in their stand on the extent of the ban and thus how to deploy the models without risking their investments.

Amalgam’s Park believes that any ban on DeepSeek’s models will “probably” be just limited to the mobile application, which is based on Chinese servers.

“It would be unlikely that the US would take any action on using the open-source R1 or V3 models as long as they were hosted on US-based servers. So far, America’s issues with Chinese technology have mainly been based around storing American-based data on overseas servers,” Park explained.

In contrast, other experts believe that the API being offered by DeepSeek may be affected by the ban and they precisely advise against using the API.

Pricing drives popularity despite ongoing probes

But despite the possibility of a ban and ongoing investigations that risk their investments, several cloud service providers and enterprise users are willing to at least start pilots on DeepSeek’s models.

The rise in popularity and availability of these models and the DeepSeek app itself can be attributed to three key themes — market demand, open-source nature, and the desire to be agile.

“Enterprises are looking for cost-effective, open-weight AI alternatives as proprietary AI models remain costly and restricted. This is why the R1 model at least became popular and is now a part of several offerings,” said Eden Consulting’s Clifford. 

On an >API level, which is to say that when enterprises are accessing the model directly from DeepSeek or OpenAI, the former offers a nearly 95% discount, according to Amalgam Insights’ chief analyst Hyoun Park, making DeepSeek extremely lucrative from a cost-effectiveness perspective.

While the pricing page for OpenAI details that their frontier o1 model costs $15 per one million input tokens and $60 per one million output tokens, the pricing page for DeepSeek says that it charges $0.014 for one million input tokens and $2.19 per one million output tokens for its R1 reasoning model.

Cloud providers and technology firms including Nvidia, AWS, Azure, and Snowflake are rapidly trying to include DeepSeek within their offerings despite the heightened scrutiny against the startup. The companies say their offerings are a result of massive demand for DeepSeek from enterprises that want to experiment with the model firsthand.

“Given the curiosity around DeepSeek-R1 and appetite to engage with it from our customers, we are supporting this model to give our customers more optionality in the models they use within Snowflake,” said Baris Gultekin, head of AI at Snowflake. 

In the same vein, Amalgam’s Park said that some testers have found that DeepSeek, for instance, has a different type of default writing style compared to ChatGPT or Anthropic Claude.  

Other experts, such as agentic AI-providing Doozer.AI founder Paul Chada said his company was actively testing a private instance in Azure and it noticed that the R1 model is easily able to get the same results for complex unstructured data extraction as OpenAI’s o1 or Claude-Sonnet for instance at a fraction of the cost.

So, how to deploy DeepSeek’s models?

Enterprises and other users have the option of deploying or using DeepSeek’s models in four ways — via the application, through their API, via model or cloud service providers and hosting it locally on their servers.

“If I was an enterprise CIO, I would not use the hosted version of DeepSeek, from DeepSeek via the API. Their terms of service (ToS) explicitly states that they [DeepSeek] will log all queries, metadata, etc and that will be domiciled in China,” said Greg Ceccarelli, advisor at US-based venture capital firm Tola Capital.

While the use of DeepSeek’s models via cloud service providers and other platforms, such as Azure, AWS, Hugging Face, and Snowflake, is widely considered safe given that they don’t send back telemetry to DeepSeek, Eden Consulting’s Clifford said that the ban may encompass these companies as well and they may be forced to stop offering the model.

However, Bradley Shimmin, chief analyst at Omdia pointed out that model users, be those consumers or enterprises, need to be very careful about where they access this model, whether that’s stateside or overseas.

“The moment a user sends in a query or an API call that hits a backend service, the data associated with that call or usage becomes the responsibility of that hosting provider,” Shimmin explained.

When asked about the impact of the ban on these models, AWS and Nvidia did not comment.

However, Snowflake’s Gultekin pointed out that the ban, if enacted, would be followed by his company and would act as a deterrent for the open source community and innovation.

“As the landscape continues to evolve, we will ensure we remain in lockstep with government regulation and will take action to remain compliant and do what’s best for our customers,” Guletkin said.

Local DeepSeek instances are safer but expensive

The only other viable option for enterprises is to self-host the models locally on their servers and many analysts say that it is the safest option.

“It’s considerably unlikely they can ban the open-weight version of Deepseek R1 since it’s available on the internet for anyone to download, host, and run in multiple distilled versions,” Tola Capital’s Ceccarelli said.

Separately, Omdia’s Shimmin pointed out that since the models are open source via an MIT license, it means that enterprises can easily vet the models’ code (weights and other assets) to satisfy any concerns or issues. 

However, Doozer.AI’s Chada pointed out that running the model locally, especially the larger 671B parameter model, is going to be expensive.

Chada said that many enterprises might see this as a worthwhile trade-off for breaking free from cloud vendor dependency. “They’re essentially betting that having full control over a top-tier AI model is worth the regulatory uncertainty,” he explained.

Should CIOs wait and watch?

Analysts are also divided on this question but many believe that the models are worth experimenting with.

“While CIOs already implementing AI solutions should evaluate DeepSeek’s potential benefits, they should exercise caution about relying on it exclusively for production use cases at this stage – mainly due to the fact it needs more battle-tested,” Chada said.

In the same vein, Eden Consulting’s Clifford pointed out that while DeepSeek’s models might be a viable option for general AI workloads, careful due diligence is required before adoption for high-risk, mission-critical, or regulated industries.

Explaining further, Clifford pointed out that researchers have flagged R1 as more vulnerable to prompt injection attacks, adversarial exploits, and jailbreaks than its Western counterparts, making it a risk for sensitive applications. “Some users also have observed that R1 aligns with Chinese government regulations, which may lead to content restrictions and filtering that do not align with Western enterprise policies,” he added.