Nvidia has instructed global cloud partners to reserve at least 51% of their services for non-Chinese clients, according to a report by the Financial Times. This directive applies to new agreements and leases, while existing contracts remain unchanged. The move aims to address U.S. regulatory concerns as Chinese demand for advanced computing power continues to grow.
Chinese artificial intelligence companies are facing a dual squeeze. They are struggling to secure export-compliant graphics processing units domestically while finding it increasingly difficult and expensive to rent advanced chips from data centers abroad. The new quota rule, delivered verbally through sales teams, effectively halves the pool of authorized Asian buyers, forcing many firms to seek alternative routes for their training and inference workloads.
Surging Prices for Compliant Hardware
Domestic supply of Nvidia chips is falling far short of demand. Chinese firms have ordered more than 6 million units of the export-compliant RTX Pro 5500 card. This figure alone exceeds the estimated global manufacturing capacity for the card, which is projected to be around 3 million units in 2027. The big three tech groups, ByteDance, Alibaba, and Tencent, have accounted for more than 3 million of these orders, purchasing the cards at approximately $11,400 each.
The total sticker price for these orders exceeds $35 billion. However, partners may only receive 5% to 10% of their original requests due to production constraints. Nvidia’s TSMC manufacturing capacity is expected to shift toward the upcoming Rubin architecture starting in 2027, further limiting availability for current models. Commercial deliveries of the RTX Pro 5500 are scheduled to begin in the first quarter of 2027, with samples arriving in December.
Prices for older compliant models have also skyrocketed. The discontinued RTX Pro 6000D has seen its price in China double. Meanwhile, a rack of Nvidia Blackwell B300 GPUs, which sells for about $600,000 internationally, reportedly fetched $2.8 million in China last month. This surge reflects the desperation of Chinese labs to access the most powerful chips barred from direct export.
Overseas Rentals and Regulatory Risks
With domestic supply insufficient, many Chinese groups have turned to overseas leases. Data centers in Southeast Asia have become the primary beneficiaries of this demand. However, the new 51% non-Chinese quota threatens this model. A Singapore-based group told the Financial Times that it is difficult to find enough non-Chinese clients to meet the requirement, suggesting that some data centers may be forced to pull back supply.
Major firms are actively seeking capacity in Europe and other regions. Moonshot has held discussions to rent clusters of 8,000 or more Blackwell GPUs in Europe to handle inference workloads. These talks occurred after reports emerged that Moonshot accessed Blackwell chips through a facility in Thailand. Tencent has secured a deal with Oracle to lease 100,000 GPUs in Southeast Asian data centers, while ByteDance and Alibaba are pursuing similar long-term leases.
The legal landscape for overseas cloud access remains uncertain. The Remote Access Security Act, which would extend U.S. export controls to remote access of U.S.-based hardware, has passed the House but has not yet cleared the Senate. Despite the lack of final legislation, neoclouds and financiers are pricing in a potential ban, which drives up upfront premiums for Chinese clients.
Domestic Chip Limitations
Chinese domestic chip production is not yet capable of meeting the massive demand. Huawei is expected to produce 1.6 million chips this year. Alibaba and Baidu are expected to contribute about 1 million combined, and Cambricon is expected to provide 300,000. The total supply, including about 1 million units made overseas, is capped at 3.5 million for the year. This falls well short of the demand, which exceeds 5 million and could more than double next year.
Even when compliant cards are available, they often lack the performance of the banned high-end models. The RTX Pro 5500 offers 84GB of GDDR7 memory and 1,398 GB/s of bandwidth. In contrast, the Blackwell B300 provides about 3.4 times the memory and 5.7 times the bandwidth, along with NVLink connectivity. Many Chinese firms prefer Nvidia’s CUDA platform over Huawei’s CANN, leading them to covet lower-end Nvidia options despite their limitations.
Regulatory approval for large-scale imports remains a bottleneck. Beijing is withholding approvals for large-volume imports, leaving about 600,000 H200 servers stuck in storage in Hong Kong. The price for an eight-card H200 server in China reached over $900,000 by the end of last month, triple the cost a year ago. Only small quantities have entered the mainland recently, often through intermediaries.
Nvidia CEO Jensen Huang faces increased scrutiny over these supply issues. He is scheduled to attend a closed-door briefing with lawmakers on November 18, just one day after Nvidia reports its third-quarter results on November 17. The company’s board is also investigating allegations that Huang was targeted by a Chinese spy, adding to the pressure as the new overseas cloud quota rules take effect.
Source: Tom’s Hardware
