China spent many years launching cyber attacks and stealing our IP, now our employees can cut out the middleman and upload it to them for free.
Last week, we wrote about DeepSeek and the increasing AI arms race between China and the United States. Over the past few days, DeepSeek has jumped to the top of the app charts – prompting alarm in the United States. Nvidia lost a staggering $593 billion yesterday alone, leading President Trump to describe the rise of DeepSeek as a “wake up” call for US tech.
But what do we know so far? Are employees using it? What data privacy implications does this have for US enterprises? What implications does this have for who “wins” the arms race?
DeepSeek versus ChatGPT
Last week, a Chinese AI company named DeepSeek released a new model – DeepSeek-R1. Immediately, it was clear that this model was set to give OpenAI a run for their money in performance at a fraction of the cost. One blogger said that it had “the Chinese OBLITERATED OpenAI”.
ChatGPT might still edge DeepSeek in areas like vision and image generation, but the concern is around the cost savings associated with DeepSeek - reportedly 95% less than OpenAI. There are still problems with bias, of course. Numerous users have tried to ask questions related to China’s Three T’s (Taiwan, Tibet, and Tiananmen) and had responses cut short.
There is also a reported lack of controls against prompt injection, as reported by Donato Capitella on LinkedIn. This data is based on LLM benchmarks on https://spikee.ai/#leaderboard .
Before going any further, however, we should clarify what we mean by “DeepSeek”.
DeepSeek: Open Source Models and a Service
DeepSeek delivers these models in two main ways: open source models and a service. Its open-source models which are available for download on sites like HuggingFace. The DeepSeek-R1 model is currently top of the “trending” list on HuggingFace, but it also has three other models in the top five.
The DeepSeek service is available on the web or mobile app; both of which have a similar visual style to ChatGPT.
DeepSeek Data Privacy Policy Concerns
DeepSeek gets a little dicey when we start to look at the data collection and data privacy policies associated with the DeepSeek service (not the open source models available for download). In fact, DeepSeek’s privacy policy clearly states that customer data is used to train models and that the data resides in China. This was well documented yesterday in an article in Wired ‘DeepSeek’s Popular AI App Is Explicitly Sending US Data to China’.
This is a real problem and one that exacerbates existing problems with AI adoption. Given security leaders are already concerned about sensitive data leaking into these tools, the idea of sensitive data flowing into China is not appealing. The Chinese government can likely just request access to this data, and data shared with DeepSeek should be considered property of the Chinese Communist Party.
Much hullabaloo was made of TiKTok’s data collection and privacy concerns. However, if American employees end up flocking to DeepSeek this would be orders of magnitude worse.
Will employees really use DeepSeek?
There’s a fairly obvious step for enterprises grappling with the use of the DeepSeek service: Block access to deepseek.com.
However, if there's one thing we know about AI adoption in the enterprise, it’s that employees will use whatever tool helps them in their job the most. No amount of blocking will stamp out it’s use. Just as employees are bypassing controls with ChatGPT, so too will they do it with DeepSeek.
According to our own analysis, there is certainly experimentation in the enterprise but, as of now, it is at a low level. On average, only 0.21% of employees have played around with DeepSeek as of today.
The AI Arms Race
Finally, what does this mean for who will win the US-Chinese AI arms race?
First of all, this is not necessarily bad for the cloud providers like Amazon, Microsoft, and Google. Afterall, very few companies are likely to opt-in to using Chinese-hosted versions for their corporate use.
Of course, this will make it harder for foundational models like OpenAI who will have to quickly close the difference in cost between themselves and DeepSeek. There’s a silver-lining here, of course. More competition among model providers can be a good thing for other startups in the application layer, who may well benefit from lower costs and higher-performing models in the long-run.
(Much) More to Come: Kimi, Qwen, and More
So much has changed in one week, that we can expect this space to evolve rapidly. China has already announced one trillion yuan investment as part of an AI plan to rival Project Stargate.
This will mean that there are going to be more models and services that emerge from China, such as Kimi. Qwen is another example of an AI company with both models on HuggingFace and a chat interface. If you care about data privacy implications, you should probably care about Kimi and Qwen, too.
.png)