The DeepSeek V4 Test: Can China's AI Ambitions Survive Without Nvidia?

The report that DeepSeek is training its V4 model on Huawei Ascend chips rather than Nvidia hardware carries implications that extend far beyond one company's technical choices. If V4 performs at or near the frontier on domestic Chinese silicon, it would challenge the foundational assumption underpinning U.S. export controls: that cutting off Nvidia access can meaningfully constrain China's AI capabilities.

The Strategic Calculation

DeepSeek's decision to build on Huawei's Ascend 910C processors — alongside chips from Cambricon and other domestic suppliers — is partly pragmatic and partly political. U.S. export restrictions have made acquiring Nvidia's latest GPUs difficult, expensive, and legally risky for Chinese AI labs. The allegations that Chinese companies have acquired Nvidia Blackwell chips through gray-market channels underscore the impracticality of relying on a supply chain that Washington can disrupt at will.

But the shift is also a calculated bet on self-sufficiency. By training a flagship model on domestic hardware, DeepSeek is building institutional knowledge — understanding how to optimize for Huawei's chip architecture, working around its limitations, and developing software tooling that does not depend on Nvidia's CUDA ecosystem. That knowledge is arguably more strategically valuable than any single model.

The Performance Question

The critical unknown is whether V4 can match frontier performance. Huawei's Ascend chips are widely acknowledged to be one to two generations behind Nvidia's best offerings in raw computational throughput and energy efficiency. The software ecosystem around Ascend — Huawei's MindSpore framework and CANN computing architecture — remains less mature than Nvidia's deeply entrenched CUDA stack.

DeepSeek's earlier models demonstrated that algorithmic innovation can partially compensate for hardware disadvantage. The company's mixture-of-experts architectures and training efficiency techniques have consistently delivered performance that punches above its compute budget. If V4 extends that pattern, it would demonstrate that the hardware gap, while real, is not insurmountable.

The Broader Ecosystem Shift

DeepSeek is not operating in isolation. Alibaba, ByteDance, and Tencent have all reportedly placed significant orders for domestic AI chips, diversifying away from Nvidia dependency. This collective shift creates a virtuous cycle for Chinese chip makers: more demand drives more revenue, which funds more R&D, which produces better chips, which attracts more customers.

Huawei's semiconductor ambitions benefit enormously from this dynamic. The company's chip division, HiSilicon, has been under U.S. sanctions since 2020 but has continued developing AI accelerators by leveraging domestic fabrication capabilities — primarily through SMIC's advanced nodes. Each major Chinese AI lab that adopts Ascend chips strengthens the economic case for continued investment in domestic chip development.

What This Means for Export Controls

If DeepSeek V4 succeeds on Huawei silicon, U.S. policymakers face an uncomfortable reality: export controls may be accelerating Chinese chip independence rather than preventing Chinese AI advancement. The restrictions were designed to create a capability gap that would slow China's AI progress. Instead, they may be creating a parallel ecosystem that, once mature, operates entirely outside Washington's influence.

The counterargument is that export controls have bought time — forcing China to spend years rebuilding capabilities that it could have acquired commercially. But time is only valuable if the country imposing restrictions uses it to extend its lead. Whether the U.S. and its allies have leveraged that window effectively enough remains an open question.

Nvidia's Long-Term Exposure

For Nvidia, the stakes are existential in terms of market access. China represented roughly 25% of Nvidia's data center revenue before export restrictions took effect. If Chinese AI labs successfully transition to domestic hardware, that revenue does not return even if restrictions are eventually loosened — because the switching costs will have already been paid.

The scenario most damaging to Nvidia is not that Chinese chips match its current products, but that Chinese AI labs learn to build frontier models on good-enough hardware. If the industry discovers that 80% of Nvidia's performance at 60% of the cost is sufficient for most production AI workloads, the premium pricing that drives Nvidia's margins comes under structural pressure — not just in China, but globally.

DeepSeek V4 is, in this sense, a test case with ramifications well beyond China's borders. Its success or failure will shape the semiconductor industry's structure for the next decade.