Nvidia Caps Chinese Cloud GPU Rentals at 49% as B300 Racks Hit $2.8M

Nvidia has ordered cloud partners worldwide to reserve at least 51% of their GPU services for non-Chinese customers, according to the Financial Times, tightening both routes Chinese AI labs use to obtain advanced compute. The same month, a B300 rack that sells for roughly $600,000 internationally fetched $2.8 million in China, the FT reported.

Chinese firms have also bought more than 1 million Nvidia gaming and workstation cards this year to cover the shortfall. Indicated orders for the export-compliant RTX Pro 5500 alone exceed 6 million cards, at least double the roughly 3 million units Nvidia could reportedly manufacture in 2027. The big three — ByteDance, Alibaba and Tencent — account for more than 3 million of those orders at about $11,400 each, a sticker total above $35 billion.

What does the 51% rule change?

Nvidia delivered the instruction verbally through sales and regional teams, and it applies globally, not only in Asia. Existing deals and leases stand; new ones must comply. Nvidia has not confirmed or denied the rule.

The cap lands on top of two earlier measures. Nvidia's whitelist more than halved its count of authorized Asian buyers, pushing Chinese groups toward leasing. And Fleet Intelligence tracking is now required for Southeast Asian providers seeking Nvidia Cloud Partner status. Together, the whitelist, the cap and the tracking could ease pressure from U.S. lawmakers at a moment when Chinese demand keeps rising.

Chinese tech groups have been the largest clients for data centers in Southeast Asia. One Singapore-based operator told the FT it is "difficult to find enough non-Chinese clients outside the US." Data centers that fail to replace those clients may have to pull back supply.

Who is renting what?

Moonshot has held talks to rent capacity from neoclouds hosting Nvidia GPUs in Europe, targeting Blackwell systems that export controls bar from China, one source said. The discussions covered clusters of 8,000 or more Blackwell GPUs for inference, and followed White House statements that Moonshot had accessed Blackwell chips through a facility in Thailand.

Tencent has a deal with Oracle to lease 100,000 GPUs in Southeast Asian data centers. ByteDance and Alibaba are pursuing similar long-term leases, and most of these overseas facilities run restricted chips such as the B300. MiniMax also leases significant compute at home and abroad.

One source told CNBC that Moonshot was asked to pay 30% to 50% more upfront than a non-Chinese entity. Neoclouds and their financiers are pricing in a possible ban on overseas cloud access, which drives that premium. Current U.S. law gives regulators no broad authority over remote cloud access; the Remote Access Security Act, which would extend export controls to that area, has passed the House but not the Senate.

Why does domestic supply fall short?

China's total AI chip supply this year will reach about 3.5 million units — Huawei at 1.6 million, around 1 million from Alibaba and Baidu, and 300,000 from Cambricon, plus roughly 1 million made overseas — against demand exceeding 5 million, a figure that could more than double next year.

The compliant cards Chinese firms buy, chiefly the discontinued RTX Pro 6000D, RTX Pro 5000 and RTX 5090, mostly serve simpler work such as prefill. The 6000D's price in China has reportedly doubled. Huawei's Ascend 910C outperforms them in most areas, but buyers still prefer Nvidia's CUDA to Huawei's CANN, since most AI frameworks and much in-house code target CUDA.

The new RTX Pro 5500 carries 84GB of ECC GDDR7 and 1,398 GB/s of bandwidth on a PCIe Gen 5 x16 link, draws up to 600W and splits into two MIG instances. A B300 still offers roughly 3.4 times the memory, 5.7 times the bandwidth and NVLink, which the 5500 lacks.

What happens next?

Beijing is withholding approval for large-volume H200 imports, leaving about 600,000 units reportedly stuck in Hong Kong storage. Only small quantities have reached the mainland in recent months, some through intermediaries, and an eight-card H200 server ended last month above $900,000 in China — about triple a year ago.

Partners who ordered 5500s may receive only 5% to 10% of their ask, and much of Nvidia's TSMC capacity shifts to Rubin from 2027. The first real test arrives with commercial 5500 deliveries in Q1 2027, after early samples in December. Nvidia reports Q3 results on Nov. 17, and CEO Jensen Huang — whose board is investigating reports he was targeted by an alleged Chinese spy — faces a closed-door lawmaker briefing a day later, where the 51% rule and his push for looser controls will both be tested.