U.S. chip export controls continue to impose real costs on China’s access to leading Nvidia accelerators. The question being pressed with increasing urgency in policy circles, financial markets, and by companies on both sides is whether those costs translate into durable strategic constraints or whether China’s adaptation is gradually eroding them.
The clearest recent evidence came in late April 2026. Reuters reported that DeepSeek had returned with a new model, DeepSeek V4, adapted or optimized for Huawei’s Ascend AI chips, hardware produced entirely within China and without any U.S. technology. A follow-up Reuters report on April 29 described DeepSeek as “betting on Huawei” as China pushes to end reliance on Nvidia. The framing matters: DeepSeek’s previous international prominence was built partly on its ability to achieve near-frontier performance with fewer and older Nvidia chips. That it is now oriented around Huawei Ascend chips signals a further step toward stack independence.
A Hardware Alternative Taking Shape
Digital in Asia has reported that Huawei is building Ascend-based cluster architectures , including the Atlas 950 SuperPoD, which links 8,192 Ascend chips, alongside CANN, the domestic software layer that Huawei positions against Nvidia’s CUDA ecosystem. Approximately 600,000 Ascend 910C chips are reportedly planned for 2026. On the memory side, CXMT’s move into DDR5 commercialization, as reported in May 2026, adds another component to what is becoming a more complete domestic hardware stack.
The software gap remains the harder problem. CANN is not CUDA: developer familiarity, ecosystem tooling, and software library coverage built around Nvidia’s platform over more than a decade are not easily replicated. Digital in Asia’s Tom Simpson noted plainly that “the software stack is the harder problem.” Ascend chips appear more mature for inference deployment than for frontier-model training, based on reporting about DeepSeek’s optimization priorities and the timeline of high-end Ascend availability.
This distinction is analytically important. Export controls are designed primarily around constraining access to advanced training chips , the compute used to build frontier models from scratch. If China’s domestic stack reaches sufficient capability for inference at scale while U.S. firms retain an advantage in frontier training, the controls would be partially effective but would not prevent China from deploying capable AI systems broadly.
The Policy Framework Gap
USCC’s Two Loops report made a related structural argument: U.S. chip controls address the digital loop of AI competition , training compute, model capability, benchmark performance , but are less suited to the physical loop, where industrial deployment generates proprietary operational data that does not depend on access to the most advanced chips. Small, task-specific models fine-tuned for factory quality control, logistics routing, or robotics perception require far less compute to deploy than frontier training runs. That compute is available domestically.
CNBC reported on May 14 that Tencent anticipated increased availability of China-designed AI chips and planned to raise AI infrastructure spending accordingly. Investors in Chinese chip stocks have read the same signals: as reported during the Trump-Xi summit week, domestic chip equities surged on the expectation that sustained controls would accelerate China’s self-sufficiency drive regardless of summit outcomes.
What Controls Are Still Preventing
The honest accounting of export controls’ current effectiveness distinguishes between domains. For frontier model training at the largest scales, controls on advanced Nvidia chips remain materially relevant. Building a model comparable to GPT-4 class systems requires compute that is difficult to substitute domestically, and Huawei’s Ascend chips have not been demonstrated at scale for frontier training runs. In that domain, the controls are imposing real costs and real delays.
For inference at scale, the picture is already more complex. Inference workloads are generally less compute-demanding per unit than training runs, and the Ascend ecosystem appears better suited to inference than to training based on available evidence. For industrial AI deployment using small specialized models, the chip constraints are least binding of all.
The longer-run policy question is whether export controls need to expand beyond accelerator chips to address software ecosystems, cloud access, chipmaking equipment, and deployment channels. That would be a substantially more complex and economically costly intervention than hardware restrictions alone. For now, China’s AI chip independence remains a developing reality rather than a completed one, but the trajectory is clear enough that framing the debate as “do controls work or not” is becoming less useful than asking what exactly they are still preventing.