← April 30, 2026
tech power

DeepSeek Bets on Huawei and China's Chip Independence Finally Gets a Proof of Concept

DeepSeek Bets on Huawei and China's Chip Independence Finally Gets a Proof of Concept
Reuters / AsiaOne

What happened

Chinese AI startup DeepSeek released its V4 model on April 24, optimized specifically for Huawei's Ascend 950 chip series rather than Nvidia hardware. Within days, ByteDance, Tencent and Alibaba were rushing to place new orders for Huawei's Ascend 950 processors. Simultaneously, Nvidia B300 servers are selling in China for roughly $1 million per unit, nearly double the US price, as export controls and Beijing's import block on the H20 have created severe scarcity. DeepSeek founder Liang Wenfeng told Reuters the company was deliberately pivoting to domestic silicon to reduce reliance on Nvidia.

Export controls designed to keep China from building frontier AI have instead given Huawei its first real commercial validation: the US didn't deny China AI capability, it just handed Huawei a captive market.

The Hidden Bet

1

US chip export controls are slowing China's AI development

DeepSeek V4 trained on Nvidia chips through circuitous channels and then optimized for domestic hardware. China now has a frontier model that doesn't need Nvidia. The controls delayed but did not prevent the outcome they were designed to stop, and they've accelerated Beijing's domestic chip industrialization by removing the easier alternative.

2

Huawei's Ascend chips are technically inferior and companies adopt them only under duress

The Ascend 950PR now outperforms Nvidia's H20, the last chip Washington permitted for sale in China. Huawei isn't a fallback; it is now the performance leader for the chips China can actually get. The gap to Nvidia's H200 is real, but the H200 isn't available in China at any price.

3

China's tech giants prefer Nvidia and will return to it once controls ease

ByteDance and Alibaba are now building procurement relationships, training pipelines, and software stacks around Huawei. Switching costs grow with every model cycle. If the ecosystem locks in, even a future easing of US controls might not reverse the shift.

The Real Disagreement

The core tension is whether the US should continue tightening chip export controls or whether the policy has already failed and tightening further only accelerates Chinese independence. The administration's position is that controls buy time and deny capability. The empirical case is that DeepSeek trained on smuggled Nvidia chips, pivoted to Huawei, and now has a domestically validated frontier model, all within the control regime. Both arguments have real force: controls may still matter at the bleeding edge, but they have clearly not stopped China from fielding a capable AI ecosystem. The honest answer is that controls are now a delay mechanism with a closing window, not a denial strategy. Treating them as the latter is the more dangerous mistake.

What No One Is Saying

DeepSeek's decision to optimize V4 for Huawei hardware is as much a business move as a patriotic one. Nvidia chips in China now cost twice what they cost in the US and arrive through gray channels with no warranty or support. Huawei can offer competitive hardware at scale with a domestic supply chain. The 'independence' framing obscures that this is also just the cheaper, more reliable option given the sanctions environment.

Who Pays

Nvidia

Medium-term: 2-3 chip generations

Loss of China's AI procurement market, which was still substantial through gray channels. As Chinese companies build Ascend-native stacks, the switching cost rises every quarter. The addressable market for future Nvidia sales in China shrinks with each Huawei design win.

US AI companies competing with Chinese models

Emerging now, significant within 18 months

If Chinese frontier models run on cheap, domestically abundant hardware while US models depend on expensive Nvidia clusters, Chinese companies can offer inference at lower cost. The competitive gap in deployment economics could widen even if US models retain a capability lead.

Smaller Chinese AI startups

Immediate

Huawei's Ascend supply is constrained: ByteDance, Tencent and Alibaba are crowding out smaller buyers. Companies without leverage to secure allocation will be stuck paying black-market Nvidia prices or falling behind on compute.

Scenarios

Domestic stack locks in

Huawei scales Ascend 950 production and Chinese tech giants build deep integration with its ecosystem. By 2027, China has a self-sufficient AI supply chain running frontier models at lower cost than US firms. Export controls become irrelevant to Chinese AI development.

Signal Huawei announces Ascend production capacity doubling or delivers on a major multi-year procurement contract with two or more of the big three Chinese platforms

Supply bottleneck stalls the pivot

Huawei cannot produce enough Ascend 950 chips to satisfy demand from ByteDance, Tencent and Alibaba simultaneously. Companies revert to black-market Nvidia procurement or slow their AI deployment cycles. The pivot becomes aspirational rather than operational.

Signal Reports of order backlogs exceeding 12 months, or Chinese firms publicly disclosing deployment delays tied to chip availability

Trump-Xi deal reopens chip trade

The May 14-15 summit produces a technology trade arrangement that eases restrictions on H200 or future Nvidia exports in exchange for Chinese concessions. Chinese firms abandon Huawei procurement, returning to Nvidia. Beijing's domestic chip strategy loses its commercial anchor.

Signal Bessent or USTR Greer publicly mentions semiconductor access as a summit deliverable

What Would Change This

If Huawei's Ascend 950 production cannot scale to meet current demand from China's big three platforms within 18 months, the chip independence thesis collapses back into a political aspiration. The signal to watch is not performance benchmarks but production volume: can Huawei actually manufacture enough chips to replace the Nvidia supply that no longer reaches China?

Sources

Reuters — DeepSeek founder Liang Wenfeng says domestic chipmakers must contribute to building a Nvidia alternative; frames the move as strategic, not just economic
Council on Foreign Relations — V4 represents a qualitative shift in the US-China AI rivalry: China now has a frontier model optimized for domestic hardware, breaking the assumption that Chinese AI depends on US chips
The Next Web — Nvidia B300 servers are selling for $1 million in China, nearly double US prices, driven by scarcity after export controls and Beijing's import block on the H20
Capacity — ByteDance, Tencent and Alibaba are all scrambling for Huawei Ascend 950 orders after DeepSeek V4 showed the chip can run a frontier model, but supply constraints are already binding

Related