US export controls may have slowed China's AI progress but they also spurred DeepSeek to innovate and release DeepSeek-V3, a top LLM, without the latest chips
A promising new model shows that innovations in artificial intelligence don't necessarily depend on the latest chips.
Context & Ripple Effects
DeepSeek-V3 is an early test of whether restricted access to leading hardware changes the pace of AI development or the methods used to achieve it. Subsequent coverage describes Chinese startups responding to sanctions by emphasizing efficiency and collaboration, while DeepSeek's leadership still characterized advanced-chip access as a bottleneck in discussions with Chinese officials.
The story also foreshadows a hardware-software co-design path: DeepSeek later described a V3.1 version tailored for next-generation Chinese-made AI chips. That makes V3 less a simple claim that chips do not matter than evidence that model builders can adjust their technical approach under supply constraints.
First-order effects
- DeepSeek gains a prominent proof point that a high-performing LLM can be developed without the newest chips, strengthening its position among Chinese AI model developers.
- US controls impose a real constraint on access to advanced hardware, but DeepSeek's release shows that the constraint does not stop model development outright.
Second-order effects
- Other Chinese AI teams have greater incentive to compete on training and inference efficiency, collaboration, and fit with available hardware rather than assuming access to frontier chips.
- The result increases the strategic value of domestic chip compatibility; DeepSeek's later work customized for Chinese-made chips points to tighter coordination between model developers and local hardware suppliers.
Third-order effects
- If this pattern persists, export restrictions may redirect AI competition from a race for the newest accelerators toward a split between frontier-hardware access and software-plus-hardware efficiency.
- Controls can remain an infrastructure bottleneck while also accelerating regional AI stacks; later reports of DeepSeek limiting early V4 access to Chinese companies suggest model access itself may become part of that strategic boundary.
The trend: AI export controls are pushing Chinese model developers toward efficiency-focused, domestically integrated AI stacks rather than eliminating their ability to advance models.