Nvidia Instructs Cloud Partners to Cap Chinese Client Access at 49 Percent

Nvidia has instituted a new quota for its global cloud infrastructure partners, requiring them to allocate at least 51 percent of their computing services to non-Chinese companies. The directive targets a widely used workaround where Chinese tech firms rent access to advanced artificial intelligence hardware located in foreign data centers.[1][6]

The instruction was delivered verbally through Nvidia's regional sales teams rather than in formal written contracts, according to multiple industry insiders. It applies to future lease agreements worldwide, though it disproportionately impacts operators in Southeast Asia where Chinese demand is heavily concentrated.[1][6]

Under current United States export controls, Nvidia is prohibited from shipping its most advanced accelerators, such as the Blackwell-based B300, directly to China. However, the regulations do not prevent a chip physically installed in Singapore, Malaysia, or Japan from being accessed remotely over the internet.[1][6]

Chinese technology giants, including Tencent and Alibaba, have increasingly relied on this remote access model to train large AI systems. By capping Chinese tenancy at 49 percent per provider, Nvidia is attempting to limit how much of a single data center's capacity can be absorbed by these firms.[1]

The Rise of Neoclouds

The mandate specifically pressures "neoclouds"—newer infrastructure providers that specialize in renting out GPU clusters for AI workloads rather than offering broad enterprise IT services. Many of these operators have built their entire business models around the surging demand from Chinese developers starved of domestic compute.[1]

Finding alternative tenants may prove difficult for these specialized hosts. One Singapore-based provider told the Financial Times that compliance is a significant commercial hurdle, noting that "it is difficult to find enough non-Chinese customers outside the U.S." to meet the threshold.[1][6]

Nvidia's verbal directive follows a broader compliance tightening initiated earlier this year. In July 2026, the chipmaker reportedly slashed its list of approved Asian buyers by more than half, introducing stricter vetting protocols that include physical site inspections, contract verifications, and end-user interviews.[1]

These measures align with guidance issued by the US Commerce Department in May 2026, which aimed to curb the diversion of advanced technology through third countries. The effort specifically targets the Blackwell series to prevent access by Chinese intermediaries operating elsewhere in Asia.[1]

Black Market Premiums

The scarcity of direct hardware access continues to drive extreme pricing in unauthorized markets. A single rack of B300 chips, which retails globally for approximately $600,000, reportedly changed hands in China last month for 18.5 million yuan, or roughly $2.8 million.[1]

That 470 percent markup echoes the record black market premiums seen earlier this year and underscores the desperation of Chinese firms to secure cutting-edge hardware. Despite strict restrictions in place since 2022, smuggling remains a persistent issue across the global supply chain.[5]

Several high-ranking employees at server assembler Super Micro were recently charged with illicitly routing $2.5 billion worth of Nvidia servers to China. The incident highlights the immense financial incentives driving the grey market for advanced artificial intelligence infrastructure.[5]

While Nvidia frames the quota as a proactive compliance measure, the lack of a written policy leaves the actual enforcement mechanism highly ambiguous. It remains unclear whether the 51 percent threshold will be measured by gross revenue, allocated computing capacity, or raw customer headcount.[1][2]

The distinction between a formal contractual obligation and verbal guidance is significant in enterprise hardware sales. Without written terms, cloud providers face uncertainty over how compliance will be audited and what specific penalties might apply for breaching the threshold.[1][3]

Shifting Revenue Streams

The broader context is Nvidia's near-total loss of direct Chinese data center revenue, which accounted for a massive share of its enterprise sales prior to the October 2022 export controls. The company has been forced to navigate a complex regulatory landscape to maintain any commercial presence.[1][4]

The compliance push reflects mounting pressure from Washington to seal the remaining side doors to advanced computing. US lawmakers have increasingly scrutinized overseas cloud arrangements, arguing that remote access undermines the intent of the original export bans.[1][5]

For Nvidia, the quota serves as a preemptive defense mechanism. Industry analysts note that voluntary de-risking by a dominant supplier often precedes formal government restrictions, allowing the company to shape the rules rather than having them imposed abruptly.[1][6]

The Sovereign Compute Race

The restriction on cloud leasing highlights the evolving nature of computing power as a sovereign resource. As artificial intelligence models require exponentially more hardware to train, the physical location of the silicon is becoming less relevant than who holds the administrative keys and network access.[1]

For Chinese developers, the new quota suggests a potentially prolonged period of limited access to the most efficient training hardware. This friction could weigh heavily on their operational capabilities, forcing them to rely on less efficient domestic silicon or highly fragmented overseas clusters.[1][6]

The enforcement of the 51 percent rule relies entirely on Nvidia's control over its supply chain. Because global demand for AI accelerators still vastly outstrips manufacturing capacity, the chipmaker holds the leverage to dictate operational terms to its partners, even when those terms are only spoken.[1][3]

Cloud providers now face a strict mathematical reality in their sales pipelines. A failure to attract enough non-Chinese tenants means losing their allocations of Nvidia's highly coveted hardware, a penalty that would immediately cripple their viability in the competitive AI infrastructure market.[1][6]