Limited Deliveries Signal Major Shift in GPU Market

After months of restricted access, Nvidia’s H200 accelerator chips have begun entering mainland China under a tightly controlled licensing system. ByteDance and Tencent each received approximately 10,000 units in recent weeks, marking the first meaningful arrival of these processors since export approvals were granted in December. However, the quantities pale in comparison to what these companies are authorized to purchase overall, and the restrictions reflect a fundamental change in how the Chinese AI computing market operates.

The deliveries come with significant strings attached. Each purchase requires individual approval from China’s National Development and Reform Commission, mirroring the case-by-case review process the U.S. Commerce Department implemented for Nvidia exports. Most of each company’s licensed allocation must remain outside mainland China, primarily in Hong Kong, before import inspections occur. Of the 400,000 units that major Chinese tech firms collectively won approval to buy, less than 3 percent have reached the mainland so far.

Why Nvidia Lost 95 Percent of Its Market Share

tech worker computer lab equipment
Photo by Chidera Faustina Okeke

The contrast between Nvidia’s former dominance and current position is stark. Just years ago, the company controlled approximately 95 percent of China’s high-end GPU market. By mid-2024, that share had collapsed to zero due to export restrictions and a government campaign questioning the safety and performance of Nvidia’s H20 chips. State-controlled data centers were barred from using foreign accelerators, and the company found itself virtually locked out of the world’s second-largest economy.

Chinese semiconductor companies capitalized on this forced exclusion. Industry projections now show domestic GPU makers claiming nearly 90 percent of China’s high-end AI chip market this year, with their shipments growing 83 percent year over year. Huawei’s Ascend chips and other homegrown designs have rapidly filled the void, giving Beijing exactly the semiconductor independence it sought through years of export restrictions.

The Frontier Training Bottleneck

Despite this domestic progress, a critical gap remains: Chinese AI labs cannot yet train frontier-level models using only domestic silicon. Evidence suggests that leading companies like DeepSeek are still relying on Nvidia hardware for training operations, while using domestic Huawei accelerators for inference tasks that don’t require as much raw performance. The H200, with its 141GB of memory and processing power roughly six times greater than the restricted H20, fills this exact gap.

The 10,000-unit allocations appear carefully calibrated by Beijing officials. Large enough to enable flagship AI training projects comparable to systems behind GPT-4 generation models, the quantities remain small enough that inference remains a protected market for domestic chipmakers. This policy allows Chinese companies to develop competitive AI models without fully surrendering the inference market to Nvidia, creating a segmented approach where foreign hardware supports training while domestic chips handle the deployment phase.

What This Means for GPU Prices and Availability

manufacturing semiconductor factory production
Photo by TruckRun

For buyers outside China, these developments carry mixed implications. Nvidia retains a 25 percent royalty on every H200 sold under the new export regime, and the company is managing an inventory of roughly 500,000 units worldwide. Market pressures from competing technologies continue to reshape pricing strategies across the GPU sector, with domestic alternatives increasingly viable for specific workloads.

The restricted Chinese market represents a profound loss of revenue for Nvidia, but the company maintains strength in Western markets where no such limitations exist. U.S. and European data centers continue building large-scale AI infrastructure predominantly around Nvidia chips. The 10,000-unit caps on individual Chinese buyers ensure that no single company can accumulate the massive H200 clusters that would be needed to eliminate dependency on Nvidia for frontier training work.

The Export Control Strategy Proves Its Point

The very fact that Beijing is now approving H200 imports validates the argument behind American export controls. Four years of restrictions forced Chinese companies to develop domestic alternatives, which they’ve done with remarkable speed. Yet those alternatives still cannot match Nvidia’s capabilities for the most demanding training tasks, proving that the controls succeeded in creating the dependency they aimed to limit.

Going forward, watch for whether China’s domestic GPU makers can close this final performance gap. If Huawei and others can deliver chips comparable to Nvidia for training work, the need for H200 imports will diminish. If they cannot, Beijing faces a choice between maintaining the export restrictions that protect its domestic industry or loosening controls to keep Chinese AI labs competitive globally. For GPU shoppers and enterprises planning infrastructure investments, this geopolitical tension will continue shaping hardware availability and pricing for years to come.