DeepSeek Is Abandoning Nvidia. Jensen Huang Says That's Catastrophic. He's Right for the Wrong Reasons.
What happened
Nvidia CEO Jensen Huang warned on the Dwarkesh Podcast on April 15 that DeepSeek optimizing its upcoming V4 foundation model to run on Huawei's Ascend 950PR chips instead of Nvidia hardware would be 'a horrible outcome' for the United States. DeepSeek, which is expected to release V4 this month, is reportedly building native support for Huawei's CANN software framework, which would break its dependency on Nvidia's CUDA ecosystem. Separately, China's customs officials blocked Nvidia H200 chip shipments, causing component suppliers to pause H200 production. US lawmakers are simultaneously pushing to add DeepSeek to the entity list.
Export controls assumed China would stay trapped inside the CUDA ecosystem. DeepSeek is proving that assumption wrong, and once the software stack separates, the hardware restrictions lose most of their leverage.
Prediction Markets
Prices as of 2026-04-19 — the analysis was written against these odds
The Hidden Bet
Chip export controls prevent China from building competitive AI
The controls degrade hardware quality but they also incentivize China to optimize aggressively for lower-end chips. DeepSeek R1 already beat US models at a fraction of the compute. The restriction is teaching China to do more with less, which is the more durable capability.
CUDA lock-in is permanent because switching costs are too high
CUDA's lock-in depends on developers needing CUDA to access the best models. If the best models ship natively on Huawei's CANN framework, the dependency reverses. Chinese developers write for CANN, global developers adopt CANN to access Chinese models, and Nvidia's software moat dissolves from the edges.
Jensen Huang is warning about a threat to America
Huang's company loses its largest growth market if China adopts Huawei chips. He is warning about a threat to Nvidia's revenue. These are related but not identical. The US could theoretically benefit from Chinese AI built on Chinese hardware, since it would reduce the strategic interdependence. Huang is eliding this.
The Real Disagreement
The fork is between treating AI dominance as a hardware problem (control the chips, control the future) versus treating it as a software problem (control the ecosystem people build on). The US government bet on hardware. Huang is now telling them the software layer is what actually matters, and that you can't control it with export controls. He's right, but his proposed solution, direct diplomacy with China, would require admitting the chip war strategy failed. The harder position to take: the export controls accelerated exactly the outcome they were designed to prevent by forcing China to build an independent stack. If nothing could change this, the US should have focused on making CUDA the global standard by making it open and cheap, not by restricting access and creating the demand for alternatives.
What No One Is Saying
If DeepSeek V4 on Huawei chips matches GPT-4 class performance, every non-US company in the world has a reason to evaluate Huawei infrastructure. The chip war would have unified the non-US world around a Chinese stack faster than any Chinese industrial policy could have done on its own.
Who Pays
Nvidia shareholders and employees
Medium-term: next 12-18 months as DeepSeek V4 adoption spreads
China was Nvidia's fastest-growing market before the H100 ban. H20 was a workaround. H200 is now blocked at customs. A successful DeepSeek-on-Huawei migration closes the Chinese market to Nvidia hardware permanently.
US AI companies dependent on CUDA lock-in
Slow-burn: 2-4 years as the developer ecosystem bifurcates
If the global developer community splits between CUDA and CANN, US companies have to choose between being competitive in China or investing in a single platform that doesn't reach half the world's AI developers.
US policymakers who designed the chip export control regime
Immediate reputational, medium-term strategic
If DeepSeek V4 on Ascend 950PR performs at frontier-model quality, it falsifies the strategic premise of three years of export control policy. The political cost of that admission is high enough that the policy will probably not be updated to reflect reality.
Scenarios
Stack Split
DeepSeek V4 launches on Huawei Ascend and performs at or near GPT-4 class. Non-US companies globally begin adopting CANN-compatible infrastructure to access cheaper inference. The AI supply chain bifurcates into US-aligned CUDA and China-aligned CANN blocs.
Signal Major cloud providers outside the US, particularly in Southeast Asia and the Middle East, begin offering CANN-based inference endpoints alongside or instead of CUDA-based ones.
CUDA Holds
DeepSeek V4 on Huawei chips underperforms due to memory bandwidth and interconnect limitations that can't be optimized away. The migration stalls, Huawei chips remain a second-tier option, and CUDA lock-in is preserved for another product cycle.
Signal DeepSeek V4 benchmarks show a 30% or more performance gap versus models running on Nvidia H100s, which DeepSeek researchers acknowledge publicly.
Diplomatic Opening
Huang's call for US-China AI talks gains traction. A bilateral AI framework emerges that gives Chinese labs access to some US chips in exchange for transparency commitments, pausing the stack divergence.
Signal A US-China meeting at secretary-level or above explicitly includes AI supply chain as an agenda item within the next six months.
What Would Change This
If DeepSeek V4 on Huawei Ascend 950PR benchmarks significantly worse than its Nvidia-based equivalents, the CUDA lock-in holds and the export control logic survives. The bottom line changes entirely if the hardware gap is too large to optimize around.