- DeepSeek has advertised at least two data center engineering roles located in Inner Mongolia, China, according to a Bloomberg report published April 10, 2026.
- Bloomberg reports the company is relying on Nvidia Blackwell-architecture chips, which are subject to U.S. export restrictions barring their transfer to Chinese entities without a license.
- Inner Mongolia hosts a concentration of Chinese data centers due to low electricity costs from coal and wind generation.
- The hiring indicates DeepSeek is building dedicated on-premises compute infrastructure rather than relying solely on cloud providers.
What Happened
DeepSeek, the Chinese AI startup behind the R1 reasoning model released in January 2025, is advertising two data center positions in Inner Mongolia, Bloomberg reported on April 10, 2026. The report also states the company has been relying on Nvidia Blackwell-class chips — hardware that falls under U.S. Commerce Department export controls restricting sales to Chinese buyers without an export license.
The job listings, which Bloomberg identified as data center engineering roles, point to an expansion of physical infrastructure rather than a shift in research headcount.
Why It Matters
DeepSeek’s R1 model generated significant industry attention in early 2025 partly because of its reported efficiency: the company claimed competitive benchmark performance against leading U.S. models while using older Hopper-generation Nvidia chips — specifically H800s, a downgraded variant permitted under earlier export rules. Reports of Blackwell access would represent a different hardware profile than the one DeepSeek publicly described during R1’s release.
The U.S. Bureau of Industry and Security has expanded semiconductor export controls in successive rounds since October 2022, with Blackwell-architecture parts — including the B100 and B200 — classified above the threshold permissible for export to China under current rules.
Technical Details
Nvidia’s Blackwell architecture, launched commercially in 2024, offers substantially higher memory bandwidth and compute density than the prior Hopper generation. B200 GPUs deliver up to 20 petaflops of FP4 throughput and 192 GB of HBM3e memory per card — roughly four times the memory capacity of an H100. Inner Mongolia has become a preferred site for large-scale Chinese AI data centers because provincial electricity prices are among the lowest in China, driven by surplus coal and wind capacity. Operating data centers there rather than in coastal cities lowers per-kilowatt-hour costs meaningfully at scale.
Bloomberg did not specify the scale of the reported Inner Mongolia deployment or the number of Blackwell units involved. The two advertised positions are described as engineering roles focused on data center operations, not model development.
Who’s Affected
If Bloomberg’s reporting on Blackwell chip access is confirmed, it raises compliance questions for any distributors or intermediaries in the supply chain. Nvidia has stated publicly that it cannot control the disposition of chips after they leave authorized channels. U.S. policymakers and Commerce Department enforcement staff — who have been tracking Chinese AI compute acquisition as a national security matter under both the Biden and Trump administrations — would view such reports as relevant to ongoing export control enforcement reviews.
For competitors and researchers outside China, the reports add to an incomplete but growing picture of the hardware base underlying Chinese frontier AI development in 2026.
What’s Next
Bloomberg did not report responses from DeepSeek or Nvidia regarding the Blackwell chip claims as of the April 10 publication date. DeepSeek has not publicly disclosed its chip procurement practices. Continued data center hiring in Inner Mongolia suggests the company plans to expand dedicated inference and training capacity through 2026, independent of external cloud infrastructure.
