China’s DeepSeek Trained Its Breakthrough AI Model on Restricted Nvidia Chips, Report Reveals, Raising Fresh Concerns Over U.S. Tech Controls

George Ellis
9 Min Read

China’s rapidly rising AI firm DeepSeek used Nvidia semiconductors that were restricted under U.S. export controls to train its latest large language model, according to a new investigative report that is sending ripples across Washington, Silicon Valley, and global semiconductor markets.

The revelation comes just months after DeepSeek stunned the global AI community with the release of DeepSeek-R1, a highly capable reasoning model that rivals frontier systems developed by the world’s top firms. Its sudden leap in performance—and its exceptionally low operational cost—prompted speculation about how the company was able to train such a system under stringent U.S. chip restrictions.

Now, newly surfaced procurement data and insider accounts suggest that DeepSeek relied on banned Nvidia GPUs, either stockpiled before restrictions took effect or diverted through intermediary channels—highlighting the growing difficulty the U.S. faces in controlling access to advanced AI hardware.


A Major Breach in the Global Tech Wall

The U.S. has spent the past two years erecting the world’s most complex export-control architecture, designed specifically to prevent China from accessing high-performance chips that can accelerate the development of advanced AI models and military technologies.

Under these rules, Nvidia’s most advanced accelerators—including the A100, H100, and various custom variants—cannot be sold to China in their unrestricted versions.

Yet the report concludes that:

  • DeepSeek used thousands of restricted A100-class GPUs
  • Some chips were shipped before the export ban
  • Others were acquired through third-party distributors in Southeast Asia and the Middle East
  • Additional units may have been purchased on secondary markets or rerouted through shell entities

If true, these findings highlight the porous nature of global semiconductor supply chains.

One senior U.S. official told reporters:

“This is exactly the scenario we feared—high-end hardware leaking through indirect channels despite formal restrictions.”


How DeepSeek Achieved Breakthrough Training Efficiency

DeepSeek shocked the global AI landscape in early 2025 by delivering a cutting-edge model trained at a fraction of the cost reported by U.S. and European lab competitors.
The company claimed that its innovations in:

  • Sparse training
  • Low-precision compute
  • Advanced data selection
  • Distributed training techniques

…allowed it to train DeepSeek-R1 for just $6 million—an order of magnitude cheaper than Western labs expected.

However, the new report suggests that a key ingredient may have been access to powerful Nvidia chips that were never supposed to reach the Chinese AI ecosystem.

If DeepSeek indeed trained on thousands of A100-level GPUs, its cost advantage becomes clearer: these chips remain extraordinarily efficient at large-scale training and are well understood by engineers, enabling maximum hardware utilization.


The Procurement Trail: How Restricted Chips May Have Reached DeepSeek

The report outlines several potential supply routes:

1. Pre-Ban Inventory Accumulated by Trading Firms

Before the U.S. tightened controls in late 2022 and again in 2023, Chinese companies legally purchased large quantities of high-end Nvidia chips.
Some of these inventories may have been later redirected to DeepSeek.

2. Third-Party Resellers in Singapore, UAE, and Vietnam

Hardware brokers in these markets have become hotspots for chip resales.
Export control circumvention often involves:

  • Mislabelled shipments
  • Shell-company intermediaries
  • Disguised end-users

DeepSeek may have benefited from these global grey markets.

3. The Massive Chinese Underground Chip Market

China has a vast secondary market for used or “recycled” GPUs from crypto mining farms, data centers, and cloud providers.
Some restricted accelerators are believed to circulate quietly in this ecosystem.

4. State-Linked Supply Channels

Although unconfirmed, several analysts speculate that certain entities with government connections may still obtain restricted chips through unofficial channels.

The report does not claim definitive proof for any single pathway—only that restricted chips were used and likely obtained through a combination of the above.


Why the Revelation Matters: Policy, Geopolitics, and AI Competition

The implications of this report are far-reaching.

1. U.S. Export Controls Face Their Biggest Test Yet

Washington’s efforts to slow China’s progress in frontier AI now appear less airtight than expected.
If DeepSeek trained a frontier model on banned chips, it raises questions about:

  • Enforcement gaps
  • Third-country compliance
  • Traceability of hardware
  • The viability of long-term semiconductor restrictions

2. China’s AI Ecosystem Is More Resilient Than Anticipated

The success of DeepSeek-R1 has already positioned China as a formidable competitor to U.S. AI labs.
If Chinese firms can:

  • Access restricted chips through indirect means
  • Combine them with algorithmic efficiency breakthroughs

…Beijing’s technological position strengthens considerably.

3. Global AI Competition Is Accelerating

DeepSeek’s breakthrough model is already being integrated into:

  • Domestic cloud platforms
  • Chinese tech conglomerates
  • Robotics and manufacturing systems
  • AI-enabled government services

Its success narrows the perceived gap between the U.S. and China in frontier AI capabilities.

3. Nvidia Is Caught in a Geopolitical Crossfire

Although Nvidia complies with U.S. export laws, the circulation of its chips in grey markets puts it under renewed scrutiny.
Analysts fear further restrictions may hit the company.


DeepSeek Denies Violating Export Rules

In a statement, DeepSeek said:

“All training operations were conducted in full compliance with Chinese law.
We do not comment on the sourcing of hardware beyond confirming that no illegal activities were undertaken by our company.”

The wording leaves unanswered how the company obtained chips that U.S. regulators classify as restricted.


Washington Prepares a Response

U.S. lawmakers are already calling for investigations into:

  • The semiconductor supply chain leak
  • Third-party distributors
  • The role of intermediary countries
  • Enforcement mechanisms for AI hardware

Some policymakers view the situation as a wake-up call:

“We must assume China will eventually gain access to restricted hardware.
Our export controls need to evolve accordingly.”

Future measures may include:

  • Global tracking systems for advanced chips
  • Sanctions on intermediary resellers
  • Stricter export licenses
  • Expanded AI governance frameworks

The Bigger Story: The End of Containment?

DeepSeek’s achievement—and its possible use of restricted hardware—suggests that controlling China’s access to advanced chips may be more difficult than Washington anticipated.

If China can:

  • Buy restricted chips through global grey channels
  • Develop its own AI-optimized architectures
  • Advance training-efficiency breakthroughs faster than competitors

…then the era of the U.S. maintaining a decisive hardware advantage could narrow rapidly.

Some analysts argue that restricting hardware may simply push China to accelerate domestic development—an outcome already unfolding.


Conclusion: A New Phase in the U.S.–China AI Technology Race

The revelation that DeepSeek used banned Nvidia chips to train its frontier model marks a pivotal moment in the global AI competition. It exposes weaknesses in enforcement, highlights China’s ability to adapt and innovate under pressure, and raises profound questions about the future of export controls, semiconductor pathways, and the geopolitics of computation.

Whether this becomes a singular episode or the first in a pattern of circumvention will depend on how aggressively the U.S. moves to reinforce the global supply chain—and how rapidly China advances towards semiconductor self-sufficiency.

One thing is certain: DeepSeek’s rise has fundamentally reshaped the global AI landscape, and the scramble to control advanced compute just entered a more volatile chapter.

author avatar
George Ellis
TAGGED:
Share This Article