- China’s national strategic plan to achieve AI chip self-sufficiency is hitting a brutal reality: the country’s tech giants bought more than 1 million Nvidia gaming and workstation chips this year as a desperate workaround to a worsening supply shortage. The irony is visceral. Four years after the US began restricting advanced AI processor sales to China, and two years after Beijing told Chinese tech companies to reduce their reliance on Nvidia, the country is buying Nvidia chips—just older models (RTX Pro 6000D, 5000, 5090) that technically comply with export controls but aren’t designed for AI. The chips are inferior to Huawei’s Ascend 910C for most workloads, but they’re good enough for simpler tasks like the prefilling phase of inference. Chinese companies are using them because they have no other choice: domestic supply can’t match demand. This year, China will produce about 3.5 million AI chips. Total demand exceeds 5 million. By 2027, demand could double to 10 million+, while domestic production rises to maybe 5-6 million. That’s a structural deficit that will persist for years.
- The supply shortage exposes a fundamental tension in China’s strategy: the country is trying to build AI infrastructure at a pace that domestic chip production simply cannot support. Alibaba, ByteDance, and Tencent combined are spending upwards of 700 billion yuan ($104 billion) on AI capex this year alone. China’s total AI compute spending is $193 billion—about 21% of US spending. That’s massive. And the intensity keeps accelerating. These companies are racing to build foundation models, deploy agentic systems, and secure computing infrastructure. They need chips now, not in 2028 when Huawei production scales. Alibaba CEO Eddie Wu said on an earnings call that industry consensus is the shortage won’t resolve until 2030. That statement is a strategic admission: China is going to be compute-constrained for years. And the constraint is becoming a competitive disadvantage because the US isn’t facing it. US companies have access to Nvidia’s best chips. Chinese companies have access to workarounds and the hope that Huawei chips improve.
- The CUDA lock-in is real and underestimated. Nvidia’s CUDA software system is the standard programming interface for GPU computing. Most Chinese tech companies cut their teeth on CUDA before being forced to transition to Huawei’s alternative, CANN. But CANN is “not as easy to work with,” which means training costs are higher, debugging is harder, and time-to-market for new models is longer. DeepSeek is working closely with Huawei to accelerate the CUDA-to-CANN transition using AI tools. But that’s an extra engineering cost and timeline stretch that US companies (using CUDA natively) don’t face. The fact that 2 million+ pre-orders already exist for Nvidia’s new RTX Pro 5500 workstation chip tells you how desperate Chinese tech companies are to avoid the CANN integration cost. They want to keep using CUDA-based development, even if it means buying Nvidia workaround products that comply with export controls. That’s a strategic win for Nvidia and a strategic loss for China’s AI independence goal.
- The overseas leasing strategy reveals strategic vulnerability. Unable to source sufficient compute domestically, Chinese tech companies are leasing GPU capacity from overseas data centers. Tencent signed an Oracle deal to lease 100,000 GPUs across Southeast Asia data centers. ByteDance and Alibaba are pursuing similar massive leases. Most of these overseas data centers are filled with Nvidia’s cutting-edge B300 chips—the kind of advanced compute that Chinese companies can’t access domestically due to export controls but can access through offshore leasing. This workaround is clever but strategically pathetic: China is spending billions on domestic chip production to build self-sufficiency, then turning around and renting advanced Nvidia chips overseas to actually train its foundation models. That means the frontier models being built by Chinese companies (Alibaba, ByteDance, Tencent, DeepSeek, Zhipu) are running on Nvidia infrastructure, not Chinese infrastructure. That’s not self-sufficiency. That’s strategic dependence hiding behind geographic arbitrage. If the US wanted to tighten controls, it could pressure cloud providers to restrict access. That would instantly cripple Chinese AI model development.
What Happened?
Chinese technology companies purchased more than 1 million Nvidia gaming and workstation chips in 2026, revealing a worsening supply deficit in China’s domestic AI chip production and exposing the limits of Beijing’s strategy to achieve semiconductor self-sufficiency. The workaround purchases of Nvidia RTX Pro products (6000D, 5000, 5090 models) occur despite four years of US export controls restricting advanced AI processor sales to China and a two-year-old Chinese government directive instructing domestic tech companies to reduce reliance on Nvidia. China’s domestic AI chip production is forecast to reach approximately 3.5 million units this year, including about 1 million chips manufactured overseas, while total demand exceeds 5 million units. By 2027, demand could double to exceed 10 million units while domestic production rises to only 5-6 million, according to industry insiders. Domestic chipmakers including Huawei (targeting 1.6 million chips), Alibaba and Baidu combined (1 million), and Cambricon (300,000) are expanding production, but the growth rate lags demand acceleration. Chinese tech companies’ combined artificial intelligence capital expenditure from Alibaba, ByteDance, and Tencent could exceed 700 billion yuan ($104 billion) this year, driving soaring computing demand. Alibaba CEO Eddie Wu stated on an earnings call that industry consensus holds the supply shortage will not resolve until 2030. Chinese tech companies are compensating by leasing compute capacity from overseas data centers, with Tencent securing 100,000 GPUs through an Oracle deal in Southeast Asia, while ByteDance and Alibaba pursue similar large-scale overseas lease arrangements, predominantly using Nvidia’s cutting-edge B300 chips. Nvidia is introducing a new RTX Pro 5500 workstation processor with more than 2 million pre-orders from Chinese companies already placed, expected to ship late 2026.
Why It Matters?
China’s forced reliance on Nvidia workaround chips and overseas GPU leasing reveals a critical vulnerability in Beijing’s AI strategy: the country is attempting to build computing infrastructure faster than it can develop domestic semiconductor alternatives. This constraint has three major strategic implications. First, it undermines China’s self-sufficiency objective. Beijing has made AI development a national strategic priority and instructed tech companies to reduce Nvidia reliance. Instead, companies are buying 1 million+ Nvidia chips annually and leasing cutting-edge Nvidia compute overseas. That’s not self-sufficiency; it’s strategic dependence masked by geographic arbitrage. Second, the CUDA software lock-in disadvantages Chinese companies. Forced to migrate from Nvidia’s CUDA to Huawei’s CANN, Chinese tech companies face higher engineering costs, slower development cycles, and time-to-market delays. US companies using CUDA natively have no such friction. This creates a persistent productivity gap that favors US AI development. Third, the shortage through 2030 means Chinese frontier models (built by Alibaba, ByteDance, Tencent, DeepSeek, Zhipu) will continue relying on US infrastructure. The models being trained on overseas Nvidia clusters are dependent on US supply chains. If the US decided to pressure cloud providers or restrict offshore access, it could instantly impair Chinese AI development. That asymmetric vulnerability is a geopolitical liability. Finally, the surge in Nvidia’s business is a strategic win for the US. Despite export controls, Nvidia’s market share in China remains dominant through workaround products, and 2+ million pre-orders for the RTX 5500 suggest that dominance will persist. Nvidia has effectively circumvented export controls through product categorization.
What’s Next?
Monitor US export control policy closely. If the US tightens restrictions on workstation chips or high-volume consumer GPU products, it directly impacts the RTX 5500 pre-orders and forces Chinese companies back to pure domestic alternatives that are less capable. If restrictions remain unchanged, Nvidia’s dominance persists through the workaround strategy. Track Huawei chip production ramp-ups. If Huawei and other domestic chipmakers accelerate timelines and achieve better-than-expected performance, it could reduce Nvidia dependency. If they slip or disappoint on performance, the supply gap widens further. Monitor Chinese government regulatory responses. The article mentions regulators asking tech companies whether RTX 5500 demand is “necessary” and companies answering “yes.” If regulators decide that necessity doesn’t override strategic independence goals, they could impose restrictions or quotas on Nvidia purchases. Such action would be economically painful but politically significant. Watch overseas lease agreements announcements. If major Chinese tech companies publicly disclose large-scale Tencent-Oracle-style arrangements with US cloud providers, it signals that offshore compute reliance is becoming structural. If companies instead announce domestic-centric infrastructure plans, it signals commitment to reducing offshore dependence. Track Chinese frontier AI lab funding and capex. DeepSeek and Zhipu are raising massive capital for data center buildouts with Huawei chips. If these buildouts succeed and models trained on Huawei infrastructure match or exceed US-trained models, it validates the path forward. If results disappoint, it validates concerns that Huawei chips can’t compete. Finally, monitor Alibaba, ByteDance, and Tencent earnings guidance on capex and AI model development. If they moderate capex due to supply constraints, it suggests they’ve accepted the shortage and are building within constraints. If they maintain aggressive capex, they’re betting on continued offshore leases and Nvidia workarounds.
Affected Tickers and Coins: NVDA | TCEHY | BABA
Source: Financial Times
