Chinese language firms hope to deploy about 115,000 Nvidia H100 and H200 GPUs—chips which have been restricted from sale to China by the U.S. authorities since 2022 over safety issues
In sum – what to know:
Unlicensed AI chip plans – Chinese language corporations goal to construct knowledge facilities in Xinjiang utilizing 115,000 Nvidia chips, regardless of strict U.S. export restrictions on such {hardware}.
Power-rich desert technique – Xinjiang’s entry to photo voltaic, wind and coal energy makes it a strategic location for energy-hungry AI compute, with most capability centered in a state-backed compound.
Feasibility in query – U.S. officers are skeptical about China’s potential to safe so many banned chips, with Nvidia warning that smuggled merchandise don’t help scalable or environment friendly knowledge heart deployment.
Chinese language corporations are reportedly planning to construct dozens of information facilities in Xinjiang, probably housing almost 115,000 restricted Nvidia AI chips, in response to a Bloomberg investigation.
The report relies on a overview of funding approvals, tenders and company filings. The one firm named within the report is Nyocor, an vitality agency based mostly in Tianjin.
In its investigation, Bloomberg states that within the fourth quarter of 2024 alone, native governments in Xinjiang and close by Qinghai greenlit 39 knowledge heart tasks. As of June 2025, at the very least seven of these tasks have been both beneath building or had secured tenders for AI compute companies. Round 70% of the computing energy linked to those knowledge heart developments is reportedly centered in a single compound backed by the Xinjiang regional authorities.
Whereas the compound wasn’t particularly named in Bloomberg’s reporting, the Xinjiang authorities had beforehand introduced a plan in April 2024 to construct a large-scale “converged computing heart” in Urumqi. The power would supply supercomputing, AI computing and knowledge storage and was described as having national-level computing capability.
Based mostly on Bloomberg’s overview of funding plans, the concerned Chinese language firms hope to deploy about 115,000 Nvidia H100 and H200 GPUs—chips which have been restricted from sale to China by the U.S. authorities since 2022 over safety issues. In April 2025, the U.S. expanded restrictions to incorporate the H20 chip, a China-specific model designed to adjust to earlier export guidelines. “On account of these new necessities, Nvidia incurred a $4.5 billion cost within the first quarter of fiscal 2026 related to H20 extra stock and buy obligations because the demand for H20 diminished,” Nvidia beforehand acknowledged.
Regardless of the ban, some Chinese language corporations are reportedly nonetheless trying to acquire the restricted {hardware}. Nvidia started growing a brand new line of chips for China in Could 2025 that will align with present U.S. export controls.
Bloomberg consulted people accustomed to ongoing U.S. investigations into chip exports, in addition to sources educated about China’s black-market {hardware} channels. These sources stated they have been unaware of the reported scale of the Chinese language knowledge heart plans and doubted the existence of a community able to smuggling over 100,000 high-end processors. U.S. officers estimate that solely round 25,000 restricted Nvidia chips have made it to China.
Nvidia responded to Bloomberg saying that an organization merely posting on-line inquiries for restricted GPUs doesn’t equate to constructing a functioning knowledge heart. The corporate additionally confused that establishing a contemporary knowledge heart requires intensive help and engineering, and added that the agency doesn’t supply companies or repairs for restricted chipss. It added that counting on outdated or smuggled chips to energy knowledge facilities is neither environment friendly nor cost-effective, particularly given the native availability of Huawei chips.
The investigation additionally famous that the Xinjiang authorities and China’s Ministry of Business and Info Know-how didn’t reply to requests for clarification or remark about these tasks.
