How safe is DeepSeek R1 on Microsoft's Copilot+ PCs, and is your data being sent to China?

DeepSeek’s arrival in the AI race shocked everyone, and now, Microsoft is (literally) taking a page of its book and bringing the open-source model to Copilot+ PCs.

But, just how safe will it be?

The Redmond tech giant said that it’s bringing NPU-optimized DeepSeek R1 models to Copilot+ PCs, starting with Qualcomm Snapdragon X and later expanding to Intel Core Ultra 200V. These models, including DeepSeek-R1-Distill-Qwen-1.5B, will use the AI Toolkit and Windows Copilot Runtime for efficient on-device AI execution.

The company said that it’s making changes when DeepSeek gets to your Copilot+ PC. In their own words, they will use a sliding window design for 130ms first token latency and 16 tokens/s throughput, along with 4-bit QuaRot quantization to optimize memory and accuracy for large-model performance in a compact footprint.

It’s also worth mentioning that the models will run locally, which means that your data won’t even be uploaded to Microsoft’s cloud, let alone sent to China—an anxiety that most users are worried about, especially with DeepSeek’s heavy censorship of sensitive queries about the Chinese government and the models’ potential link to the Chinese Communist Party (CCP).

Perplexity, a popular AI-power search engine tool, has also hosted DeepSeek in their product, but it’s US-hosted and won’t be censored like its original.

DeepSeek, a rival to OpenAI’s ChatGPT, claimed that it’s built its models using just $6 million and older Nvidia H100 GPUs, a low-cost solution especially as making and building an AI is so much more expensive.

DeepSeek’s surge in popularity caused so much panic and sense of urgency in the US, so much so that the market lost over $1 trillion after its arrival. Both Nvidia and Microsoft were among the tech firms that were affected by it, and even OpenAI launched ChatGPT Gov, a special ChatGPT for US government agencies to maintain “America’s global leadership in AI.”