Technology & AI Research Desk  |  April 2025  |  5 min read

 

⚡ Quick Takeaways •       Nvidia CEO Jensen Huang warns that a Huawei-powered DeepSeek V4 would be a negative signal for US AI dominance. •       DeepSeek V4 may run on Huawei Ascend chips — though Nvidia Blackwell support has also been reported. •       Chinese AI firms are compensating for hardware limits through software optimisation and Mixture-of-Experts scaling. •       China’s AI talent pool and lower energy costs sustain large compute clusters despite chip restrictions. •       The AI race is shifting from raw chip performance to software efficiency, scale strategy and ecosystem control.

 

Nvidia CEO Jensen Huang has raised an alarm over the US-China AI competition, warning that a successful deployment of DeepSeek’s next model on Huawei’s domestic chips — rather than American processors — would represent a meaningful shift in the global AI landscape. The statement arrives as China’s DeepSeek lab prepares to release V4, already sparking debate over whether it marks the first major defection from the US chip ecosystem.

 

📌 Key Factor	📌 Key Factor	📌 Key Factor
DeepSeek V4 Next major Chinese AI model expected to run on Huawei Ascend chips	Huawei Ascend China’s primary domestic alternative to Nvidia GPUs in AI training	MoE Scaling Mixture-of-Experts — China’s key strategy to offset hardware gaps

 

Jensen Huang’s Warning and What Triggered It

Nvidia’s chief executive described a Huawei-powered DeepSeek V4 as a negative outcome for the United States — commercially and strategically. Huang’s concern is that such a move would signal that frontier AI development is no longer exclusively dependent on American chip infrastructure, weakening one of the US’s most significant points of leverage in its technology rivalry with China.

DeepSeek V4 is expected imminently and has drawn significant industry attention. Reports indicate the model could operate on Huawei’s Ascend processors, though some sources suggest it may retain partial support for Nvidia’s Blackwell architecture. The ambiguity underscores how closely the industry is watching which hardware ecosystem a leading Chinese model will anchor itself to.

 

⚠️ Unconfirmed Detail Reports that DeepSeek V4 will run on Huawei Ascend chips have not been officially confirmed by DeepSeek. Separately sourced indications also suggest partial Nvidia Blackwell compatibility may remain. Both claims should be treated as reported rather than verified at time of publication.

 

US vs China: How the AI Race Actually Stacks Up

Raw chip performance is not the only variable in the AI competition. The two countries approach the challenge from fundamentally different positions — and China has been systematically working around its hardware disadvantage rather than waiting to close it.

 

Dimension	US Position	China Position
Leading AI Chip	Nvidia Blackwell (H100/B200)	Huawei Ascend 910B/910C
Raw Compute	Ahead — higher performance per chip	Behind — offset by volume and scale
Software Strategy	Ecosystem-dominant (CUDA)	Optimisation-first; MoE scaling
Energy Cost	Higher — major data centre expense	Lower — state-backed energy advantage
AI Talent Pool	Deep — global recruitment	Large domestic base; rising rapidly
Export Controls	Restricting chip access to China	Developing domestic alternatives
Key Model	GPT-4o, Claude, Gemini Ultra	DeepSeek V3 / V4 (upcoming)

 

How China Is Closing the Gap Without Better Chips

Huawei’s Ascend processors trail Nvidia’s leading GPUs in raw compute per chip. However, Chinese AI developers have turned to architectural and software strategies to compensate. Mixture-of-Experts — a model design approach that activates only the most relevant sub-networks for a given task — allows significantly more efficient use of available hardware. DeepSeek’s earlier models have already demonstrated competitive benchmark performance using this approach.

Beyond software, China benefits from structural advantages difficult for the United States to counter through export controls alone. A large domestic pool of AI engineers, state-directed investment in compute infrastructure and lower-cost energy all support large-scale parallel computing clusters needed to train frontier models — even when individual processors are less capable than their American counterparts.

 

ℹ️ What Is Mixture-of-Experts? MoE is an AI architecture where a model is split into specialist sub-networks. Only the most relevant sub-networks activate for any given input — dramatically reducing compute per inference step and enabling large model capability on less powerful hardware.

 

Policy Pressure and the Broader Stakes

The DeepSeek V4 debate is unfolding against escalating US-China technology tensions. American lawmakers have pushed for tighter controls on Chinese AI firms’ chip access, and Nvidia export restrictions have been in place for years. Yet their effectiveness is increasingly questioned as Chinese developers demonstrate the ability to work around hardware limits through software and scale.

Huang’s warning points to a competition evolving beyond the chip layer. Software efficiency, model architecture, ecosystem control and scaling methodology are becoming the defining battlegrounds. If DeepSeek V4 successfully runs on domestic hardware, it would shift the debate from who has the best chips to who can build the most capable systems regardless of what they use.

 

✅ Analyst Perspective Export controls on hardware may slow but not stop Chinese AI progress. Long-term US competitiveness may depend as much on software ecosystem strength, research openness and talent retention as on chip export policy.

 

DISCLAIMER: All information in this article is based on publicly available data and official statements. This content is intended for general informational purposes only.