DeepSeek says it used Nvidia H800 chips, available in China until October 2023, to train R1, suggesting US export controls could be a problem for future models
- Nvidia calls R1 an ‘excellent’ advance that meets US limits — Trump says release should be a ‘wake-up call’ for US companies
Context & Ripple Effects
Earlier coverage described export controls as both a drag on China’s AI progress and a spur to DeepSeek’s efficiency work, including its release of DeepSeek-V3 without the newest chips. This report ties that broader debate to the hardware basis of R1.
The significance is less that R1 used a China-available Nvidia product than that a capable model emerged from hardware that was already subject to a tightening policy boundary. It makes access to compliant compute a central constraint on DeepSeek’s next training cycle.
First-order effects
- DeepSeek can point to H800-based training as evidence that R1 was built with chips available in China before the October 2023 cutoff, while Nvidia maintains that the work complied with US limits.
- For DeepSeek, the usable installed base of older Nvidia hardware becomes strategically important because future models may not have the same supply path.
Second-order effects
- US policymakers and Nvidia must contend with the possibility that restricting the latest accelerators changes the pace and cost of training more than it prevents competitive model advances.
- Later reporting that R2 faced delays tied to a shortage of Nvidia server chips in China illustrates the practical follow-on: chip availability can become a product-timing constraint even when engineers can extract strong results from older hardware.
Third-order effects
- If this pattern persists, export controls are likely to function as a constraint on the scale, speed and economics of Chinese frontier-model development rather than a simple on/off switch for model capability.
- The resulting pressure favors AI labs that can optimize models for constrained compute and deepens the strategic value of alternative domestic chip supply, though the durability of that shift depends on available hardware and policy enforcement.
The trend: AI export controls are increasingly becoming a contest over whether compute restrictions can slow frontier-model progress faster than labs can adapt their software and hardware strategies.