PLA Has Been Acquiring America's Most Advanced AI Chips — Including Top Nvidia GPUs — Openly Through Its Own Procurement Channels for Years, WireScreen Report Finds

New York, June 1, 2026 — The People's Liberation Army has been acquiring America's most advanced computing technology, including Nvidia's most powerful artificial intelligence chips, openly throughout the several years that the United States has been trying to stop it, according to a new special report released today by WireScreen.

‍

The report, Processing Power: Export-Controlled GPU Procurement Across Chinese Military Services, analyzes more than 3,800 Chinese military and government procurement records from 2019 through 2025. More than 540 of those records name specific accelerators by model. Twenty procurement actions are developed as detailed cases, mapping a 374-firm commercial ecosystem behind the transactions. The buyers span every major PLA service branch — the Cyberspace Force, the Rocket Force, the Aerospace Force, the Air Force, the Joint Logistics Support Force, and the Central Military Commission's own equipment development apparatus — as well as Entity-Listed institutions including the Jiangnan Institute of Computing Technology (JNICT) and the National University of Defense Technology. The report does not assert that Nvidia or other American technology firms were aware of the transactions.

‍

"Export controls introduced friction — they raised the cost of acquisition at every stage,” said John Costello, author of the report and Director of Strategic Affairs at WireScreen. "The story of this report is what that friction looks like in practice: more shell companies, more rebrands, more compute-as-a-service, more procurement that runs through approved domestic OEMs with American silicon inside. These adaptations are the intelligence value."

‍

WireScreen's analysis shows that export controls have introduced real friction — failed tenders, supplier shortages, and procurement obfuscation — but have not severed Chinese military access to advanced U.S.-origin compute. Rather than isolated violations, the report documents an ecosystem of PLA units, shell companies, state-owned enterprises, and commercial intermediaries that continued sourcing restricted hardware after each successive wave of U.S. controls.

‍

Other key findings include:

• Every Nvidia generation persists through every wave of controls. The A100 — banned to China in October 2022 — appears in 177 procurement records, the most of any model in the dataset. Its compliance variant, the A800, accumulated 70 records during its roughly twelve-month legal window. The H100 (58 records) and H800 (17 records) continue to appear in records after the October 2023 expansion that was supposed to close the A800 loophole.
• The largest single contract documented is a ¥395.99 million (roughly $55 million) 474-node compute cluster awarded by NUDT — on the U.S. Entity List since 2015 — to Shenzhen-listed IT distributor Digital China.
• A structural enforcement gap is emerging through compute-as-a-service. The Joint Logistics Support Force awarded a 20-million-yuan (roughly $2.8 million) GPU machine-hour framework to listed commercial compute aggregator Paratera, allowing controlled hardware to remain in a domestic data center while military workloads arrive as job submissions — leaving no shipment to interdict at the border.
• Operational, not academic, units are the buyers. A Cyberspace Force unit specified hashcat — the world's most widely used GPU-accelerated password-cracking tool — as a required workload in a January 2024 procurement for A100-class GPU servers. A separate cyber unit procured Nvidia A100 and A800 accelerators alongside Cisco, Palo Alto Networks, Fortinet, and Juniper security appliances in a single 2023 run — consistent with a unified offensive cyber test range build.
• Military units frequently sought to combine Nvidia accelerators with PLA-approved "domestic brand" chassis from Sugon or Inspur, so the transaction reads as a domestic-brand server purchase while the enabling silicon remains American.
• The 374-firm supplier ecosystem ranges from publicly listed national integrators with decades of operating history down to single-employee shell vehicles whose sole commercial function is bidding for one Entity-Listed buyer. Recurring patterns include sixty-four-fold capital pumps, same-day rebrands from food trading to GPU distribution, and ownership rotations synchronized to procurement cycles.

"Export controls have introduced real costs and real friction. But the procurement record shows that military demand persisted, and the networks supplying it adapted," Costello added. "This isn't a story about a few bad actors slipping through the cracks — it's a story about a system that was built to keep moving."

‍

The full report is available at https://www.wirescreen.ai/research/gpu-procurement-report.

‍

About WireScreen 

WireScreen is a corporate intelligence platform with proprietary data and analytics on more than 20 million Chinese business entities. The platform integrates Chinese corporate registry data from the State Administration for Market Regulation, beneficial ownership tracing through the Unified Social Credit Code system, risk-flag classifications covering Entity List, CMIC, FDPR, and military end-use designations, and cross-entity relationship mapping to surface commercial relationships not visible from any single conventional data source.

‍

Media Contact

Krassi Genov
Head of Marketing, WireScreen
krassi.genov@wirescreen.ai
www.wirescreen.ai

‍