Nvidia, 중국 기업들이 게이밍 및 워크스테이션 GPU를 100만 개 넘게 구매함에 따라 클라우드 파트너의 중국 사업을 49%로 제한할 수 있다고 밝혔다. 이는 중국 기업의 해외 클라우드 AI GPU 임대에 대한 단속 조치다.

Nvidia-based compute is becoming scarce and expensive for Chinese AI labs on both of their main routes: buying export-compliant cards and renting capacity abroad. The Financial Times reports that Nvidia's own crackdowns are partly responsible. Nvidia has told cloud partners worldwide that at least 51% of their services must go to non-Chinese firms. Meanwhile, Chinese companies have bought more than 1 million Nvidia gaming and workstation cards this year to cover the shortfall. A B300 rack that sells for about $600,000 internationally reportedly fetched $2.8 million in China last month.

Chinese customers have ordered more than 6 million of Nvidia's export-compliant RTX Pro 5500 cards. That is at least twice the 3 million the company could reportedly manufacture in 2027. One source told CNBC that Moonshot was proposed to pay 30% to 50% more upfront than a non-Chinese entity.

Many Chinese groups have relied on overseas leases for advanced AI work, including model training, and these technology groups have been the biggest clients for data centers in Southeast Asia. Data centers unable to find non-Chinese clients could be forced to pull back supply. "It's difficult to find enough non-Chinese clients outside the US," one Singapore-based group told the FT. Nvidia has neither confirmed nor denied the new rule. Implementing it, however, would ease criticism from U.S. lawmakers just as Chinese demand is growing.

According to the reporting, Nvidia communicated the requirement through verbal instructions to sales and regional teams, and it applies globally rather than just in Asia. Existing deals and leases are not changed, but new ones must comply. Nvidia's whitelist more than halved its authorized Asian buyer count, which pushed more Chinese groups toward leasing over other options, including smuggling. In addition, Fleet Intelligence tracking is now required for Southeast Asian providers that want to become Nvidia Cloud Partners. Together, the whitelist, the cap, and the tracking may help alleviate U.S. concerns.

Among companies that have moved toward renting, Moonshot has held discussions to rent capacity from neoclouds hosting Nvidia GPUs in Europe. It is specifically looking for Nvidia Blackwell systems, which export controls bar from China. One source said the talks covered clusters of 8,000 or more Blackwell GPUs in Europe. These clusters would handle inference, serving models to end users. The European talks took place after the White House said Moonshot had accessed Blackwell chips through a facility in Thailand. The firm has been vocal about its intention to secure more compute.

Moonshot is not the only Chinese company seeking Nvidia chips for AI training. Tencent has a deal with Oracle to lease 100,000 GPUs in Southeast Asian data centers. ByteDance and Alibaba are also pursuing similar long-term leases, and most of these overseas data centers run chips like the banned B300. MiniMax, like Moonshot, leases significant compute both at home and abroad.

The main reason is that domestic AI chip supply is simply not up to the task, despite considerable effort. Huawei is expected to produce 1.6 million chips this year, with about 1 million from Alibaba and Baidu and 300,000 from Cambricon. Total supply this year could reach 3.5 million, including about 1 million made overseas, against demand that far exceeds 5 million. That gap could more than double next year.

Part of the problem is that the export-compliant cards Chinese firms have been buying, mostly the RTX Pro 6000D, RTX Pro 5000, and RTX 5090, are useful mainly for simpler work such as prefill. Even so, the discontinued 6000D's price in China has reportedly doubled. Huawei's Ascend 910C outperforms these cards in most areas, but the lower-end options remain highly sought after, partly because of a preference for Nvidia's CUDA over Huawei's CANN. CUDA is supported by most AI frameworks and libraries, and much in-house code is optimized for it.

The RTX Pro 5500 has 84GB of GDDR7 memory with ECC and 1,398 GB/s of memory bandwidth. It runs on a full PCIe Gen 5 x16 interface, draws up to 600 W, and can be split into up to two MIG instances. It still falls well short of the B300, which offers about 3.4 times the memory and 5.7 times the bandwidth, along with NVLink, which the 5500 lacks.

| Card | Memory | Bandwidth | Interconnect | Status for Chinese buyers |

|---|---|---|---|---|

| RTX Pro 5500 | 84GB GDDR7 | 1.4 TB/s | PCIe Gen 5 x16 | Export-compliant; samples in December, deliveries from Q1 2027 |

| H200 | 141GB HBM3E | 4.8 TB/s | NVLink | Licensed by the U.S. case by case; Beijing withholding approval for large-scale imports |

| B300 | 288GB HBM3E | 8 TB/s | NVLink (1.8 TB/s) | Banned for sale to China; rented abroad |

Indicated orders from China are not precise, but the total of over six million includes more than three million from ByteDance, Alibaba, and Tencent. The big three reportedly ordered at about $11,400 per card, for a total sticker price over $35 billion. That is difficult to fill when there may be fewer than three million produced next year, and partners might actually receive only 5% to 10% of the original request. Starting in 2027, much of Nvidia's TSMC capacity may shift to Rubin, taking up more production, while the RTX Pro 5500's early samples are due in December.

On the legal side, renting is currently easier than owning, as current U.S. law does not give regulators broad authority over overseas cloud access. The Remote Access Security Act (RASA), which would extend U.S. export controls to that area, has passed the House but has not yet passed the Senate. Neoclouds and related financiers, who are directly affected, are pricing in a possible ban regardless, which drives the upfront premium.

China's side is not clean either. About 600,000 H200s are reportedly stuck in storage in Hong Kong, awaiting shipment to the mainland. Beijing is withholding approvals for large-volume imports such as this one. Only relatively small quantities have arrived in recent months, for some large internet companies, while others have come through intermediaries and indirect routes in small amounts. At the end of last month, an eight-card H200 server cost over $900,000 in China, about three times the cost a year earlier. This is despite the fact that H200 sales are already under U.S. license.

The first test of how many cards will actually land comes in Q1 2027, when the RTX Pro 5500 is expected in commercial deliveries. For rentals, whether the Remote Access Security Act becomes a factor depends on a Senate vote, which is not yet scheduled. Nvidia's Q3 results come out soon, on Nov. 17, and could provide some clarity.

CEO Jensen Huang is already under scrutiny over reports that he was targeted by an alleged Chinese spy, a matter Nvidia's board is now investigating. He will face even more pressure a day later, when lawmakers have asked him to attend a closed-door briefing. He has pushed for looser controls and greater access, so the assumed 51% rule will also be put to the test.