The great AI Irony: China cracks down on Western models while US companies flock to DeepSeek
- China continues to purge both demand for AI chips from its ecosystem and foreign AI models, citing ‘security risks’ and privacy concerns
- The starkest warnings come from China’s Ministry of State Security (MSS)
- US firms continue to flock to cheaper alternatives for their AI needs
In what many may consider a tit-for-tat response to increasing US hardening of controls and sanctions over both AI hardware and software service providers, and procurement across the board, Chinese authorities seem to be increasingly pointing to further disentanglement from US-based providers, citing security concerns for users leveraging grey-market access to frontier models such as Anthropic’s Fable.
The Chinese Ministry of State Security (MSS), responsible for both domestic and foreign intelligence operations and surveillance, has warned that users who leverage third-party tools and marketplaces to access highly sought-after compute resources from US-based AI models may be exposing themselves to security risks and potential backdoors for cyber espionage.
It pointed to inadequate encryption, bait-and-switch models, and even the potential retention of data as key concerns, aligning with the narrative of many US-based security agencies that sprang into action once models such as DeepSeek began gaining traction in domestic markets.
AI abuse concerns mount on both fronts
While both the Chinese and the US governments are increasingly hostile to foreign AI, citing them as security threats, US consumers continue to use models such as Alibaba’s advanced Qwen 3.6, DeepSeek V4 Pro, and GLM 5.1, all of which offer open-source models that allow for localized AI as well as cheaper hosted inference options than what OpenAI and Anthropic currently charge for their services.
The reason boils down to a simple factor: cost. Not only do distilled models like those offered by DeepSeek cost a fraction of their peers, but they can also be deployed without any licensing cost on existing hardware, even as models grow increasingly larger and complex in terms of compute power and memory requirements.
This has not stopped Chinese developers and AI enthusiasts from doing the exact opposite, hunting for the cheapest way to access US-based AI models and their compute at a fraction of the cost, regardless of the trade-offs involved.
Research published in May 2026 by Zilan Qian of the Oxford China Policy Lab documented a thriving ecosystem of API “transfer stations” — proxy services operating openly on Taobao, GitHub, and Telegram that resell access to Anthropic’s Claude models at as little as a tenth of the official price.
Listings advertise unlimited Claude Code subscriptions, full-fat Claude Opus access, million-token context windows (no VPN required), and payment in RMB via WeChat or Alipay.
The economics only work because the product is rotten underneath: bulk-registered accounts farming free credits, subscriptions split across dozens of users, credentials bought with stolen credit card data, which does indicate that Chinese authorities are warning of a trend that is becoming both a security and a financial risk to all parties concerned.
This, coupled with the fact that the White House continues to accuse Chinese developers of ‘jailbreaking’ or stealing data from US AI models, indicates that the issue is a tenuous one, as both countries vie for the coveted position of being the first to chart a path to AGI (Artificial General Intelligence).
A message from Chinese security services, such as the MSS, in this context implies that while US consumers continue to embrace technology from the region, Chinese consumers may face future restrictions, with advisories like the one provided, often backed by covert and overt actions by state agencies to ensure compliance.
An example of how this could work out can be found in the domestic chip market: while there is no blanket ban on importing Nvidia GPUs for AI training and inference workloads from the Chinese state, the country widely discourages such imports, which has led to some interesting observations by Huawei’s chairman after the initial heavy-handedness by US authorities when it came to export permissions for Nvidia and AMD’s highest-end AI chips.
China now continues to build, upgrade, and redesign both its chipmaking and memory sectors on a war footing, much to the chagrin of Nvidia’s CEO, Jensen Huang.
Both countries may see users from their counterparts as a victory for their AI models for now; additional users, and therefore, queries, allow for more training and information to be imparted, allowing for the perfecting of both security issues, such as jailbreaking, and also access to a trove of information from end users, which strengthens the cognitive abilities of frontier AI models.
As things stand, while US consumers continue to enjoy access to Chinese AI models, the US government might also, in the near future, push further than it already has when it comes to enforcing enhanced restrictions on users aiming to use ‘foreign AI’ as battle lines continue to be drawn, even as China seems to be pre-empting such a move using the same narrative.
Follow TechRadar on Google News and add us as a preferred source to get our expert news, reviews, and opinion in your feeds.
