Nvidia B300s Hit $1M in China as Export Controls Reshape AI Economics
US chip restrictions drive 3-4x price premiums for restricted hardware while spurring $5.6 billion pivot to Huawei alternatives—cloud providers absorb structural cost disadvantage even as domestic AI models gain market share.
Nvidia’s B300 AI accelerators are selling for approximately $1 million per unit in China—a 3-4x premium over US pricing—as export controls imposed since October 2022 create artificial scarcity that is simultaneously inflating compute costs and accelerating indigenous chip development.
The pricing surge, confirmed by industry sources to Reuters, reflects the downstream economic impact of US semiconductor restrictions that have progressively tightened through December 2024. Strong demand for AI infrastructure combined with intensified crackdowns on grey-market chip smuggling have nearly doubled B300 server prices to 7 million yuan, compared to estimated US pricing of $250,000-$350,000 for equivalent hardware. The delta represents a quantifiable competitive disadvantage for Chinese cloud providers—one they are addressing not by curtailing AI ambitions, but by restructuring supply chains and absorbing cost increases while accelerating domestic alternatives.
Cloud Providers Pass Costs Through as Supply Tightens
China’s major cloud platforms have begun raising AI compute prices in response to hardware cost inflation. Tencent Cloud announced a ~5% price increase for AI services effective May 2026, per Trendforce. Alibaba Cloud implemented increases starting 18 April, with Baidu following suit, according to Caixin Global. The coordinated pricing adjustments signal that 30-40% compute cost increases—driven by restricted access to cutting-edge Nvidia hardware—are now being absorbed across the Chinese AI stack.
This cost pressure emerges even as Chinese AI models have tripled their global footprint. Morgan Stanley data cited by Reuters shows Chinese models captured 32% of worldwide token usage in March 2026, up from just 5% a year earlier—a sixfold expansion occurring precisely as hardware access constricts.
ByteDance, Alibaba Commit $5.6 Billion to Huawei Chips
The pricing asymmetry is accelerating procurement commitments to domestic alternatives. ByteDance has committed $5.6 billion to Huawei’s Ascend 950PR chips, with Alibaba and Tencent placing substantial orders, bringing total committed procurement beyond 500,000 units, Aitoolsbee reported. The Ascend 950PR, priced around $16,000 per unit—roughly 6% of the inflated B300 price—entered mass production in April 2026 with Huawei targeting 750,000 units this year.
“Gaps still exist between Huawei and Nvidia in terms of individual chips and ecosystems, and we are working to close these gaps. But we have the capability to build SuperPoDs and SuperClusters — which are the source of our confidence.”
— Eric Xu, Huawei Rotating Chairman
Huawei’s broader Ascend roadmap includes manufacturing 600,000 Ascend 910C units in 2026—double the ~300,000 produced in 2025—with total Ascend production potentially reaching 1.6 million dies, according to RCR Wireless citing Bloomberg. The aggressive scaling reflects both demand pull from cloud providers and strategic Chinese government support, including subsidies covering up to 50% of electricity costs for data centers deploying Huawei and Cambricon chips, as investingLive reported in November 2025.
DeepSeek V4: Software Efficiency as Strategic Response
The hardware cost disadvantage is driving parallel advances in software optimization. DeepSeek released V4 on 24 April 2026—the company’s first model optimized specifically for Huawei’s Ascend architecture, MIT Technology Review reported. The release demonstrates efficiency gains that partly offset hardware limitations: V4-Pro achieves performance targets using just 27% of the floating-point operations required by the prior V3.2 generation and 10% of the key-value cache at 1-million-token context windows.
This software-hardware co-optimization strategy represents a structural adaptation to constrained access. Rather than attempting to match Nvidia performance on equivalent tasks, Chinese labs are redefining architectural trade-offs—prioritising inference efficiency and reduced memory overhead over raw compute density. The approach leverages China’s strength in large-scale model deployment and algorithmic innovation while working around semiconductor manufacturing constraints imposed by Export Controls implemented under the October 2022, 2023, and December 2024 expansions.
| Model | China Price | Production Status |
|---|---|---|
| Nvidia B300 | ~$1,000,000 | Export-restricted |
| Huawei Ascend 950PR | ~$16,000 | 750K units (2026 target) |
| Huawei Ascend 910C | Undisclosed | 600K units (2026 target) |
Policy Effectiveness Remains Ambiguous
The export control regime has unambiguously inflated Chinese AI compute costs—a 3-4x hardware premium translates directly into higher capital expenditure for cloud infrastructure and longer return-on-investment timelines for model training. Yet the strategic outcome remains uncertain. Controls have not prevented Chinese models from capturing nearly one-third of global token usage, nor have they curtailed infrastructure investment by ByteDance, Alibaba, and Tencent.
Instead, the restrictions appear to be triggering a medium-term decoupling: Chinese firms are absorbing near-term cost disadvantages while building indigenous supply chains that—if successful—will reduce long-term dependency on US semiconductor exports. Huawei’s production ramp, subsidised by Beijing and backstopped by multi-billion-dollar procurement commitments, suggests the policy is accelerating exactly the outcome it ostensibly seeks to prevent: a vertically integrated Chinese AI stack insulated from external controls.
US export controls targeting advanced Semiconductors began in October 2022 and have been expanded twice—October 2023 and December 2024—to restrict sales of cutting-edge AI accelerators to China. The controls aim to slow Chinese military AI development by denying access to the most advanced training hardware, but rely on enforcement mechanisms vulnerable to smuggling and workarounds.
What to Watch
Monitor Huawei’s quarterly Ascend shipment data through year-end 2026 to assess whether production targets of 750,000 950PR units and 600,000 910C units are achievable given SMIC’s 7nm manufacturing constraints. Track cloud pricing adjustments from Alibaba, Tencent, and ByteDance as indicators of whether cost increases stabilise or accelerate—persistent upward pressure would signal supply-demand imbalances are worsening. Watch for additional DeepSeek or similar releases optimised for Huawei hardware; if efficiency gains continue at V4’s pace, software may be compensating for hardware gaps faster than export controls anticipated. Finally, observe whether Chinese AI model token share continues expanding beyond 32%—sustained growth despite cost headwinds would confirm that export controls are reshaping economics without curtailing competitive position.