Banned Nvidia AI chips double to $1.1M on China black market as US crackdown tightens supply

Nvidia AI Chips Surge to Record Black Market Prices in China

The price of banned Nvidia AI chips on the China black market has more than doubled over the past six months, with the company's flagship DGX B300 server now fetching over 8 million yuan, or approximately $1.1 million1

1

2

2

. This represents a dramatic increase from 4 million yuan just six months ago, and stands in stark contrast to the system's US retail price of around $400,0003

3

. The DGX B300, which Nvidia describes as an "AI factory for AI reasoning," packs eight Blackwell Ultra GPUs, Intel Xeon 6776P CPUs, and over 2TB of GPU memory—making it a highly sought-after platform for AI workloads despite US export restrictions that prohibit its sale to China1

1

.

The Nvidia RTX 6000 Pro workstation card has experienced similarly dramatic price escalation. These Blackwell GPUs, which feature 96GB of memory and serve as the professional version of the RTX 5090, have climbed from roughly 50,000 yuan at the start of the year to as much as 130,000 yuan, or approximately $18,000 to $20,0001

1

3

3

. The massive memory capacity makes these cards particularly attractive to entry-level AI firms seeking to deploy large language models, a capability that neither the consumer RTX 5090 with 32GB nor the China-specific 5090 D V2 with 24GB can match3

3

.

US Export Restrictions and Smuggling Crackdown Drive Supply Crunch

The dramatic price increases stem from a perfect storm of tightening US chip controls and enforcement actions that have severely constrained both legal and illegal supply channels. Washington intensified enforcement at the end of last year, and in March, a Supermicro co-founder was charged in connection with an alleged $2.5 billion scheme to route smuggled Nvidia AI servers to Chinese buyers—marking the largest case by US law enforcement related to AI chip exports2

2

4

4

. Following this action, authorities in Taiwan and Malaysia opened their own smuggling investigations, effectively drying up the re-export routes that traders had previously relied upon2

2

.

The Trump administration has worked to close loopholes for smuggled systems, and Nvidia has made clear that any systems found operating in unauthorized territories would not receive official support or service1

1

. Nvidia told the Financial Times that building data centers from smuggled chips is a "dead-end" strategy2

2

. Yet these warnings have done little to dampen demand from Chinese tech firms desperate for access to cutting-edge AI hardware.

Chinese Customs Restrictions Block Legal Alternatives

Adding to the supply crisis, Chinese customs restrictions have closed off what few legal channels remained. After the Trump administration approved exports of older H200 Hopper GPUs to China, Chinese customs authorities were instructed to block these chips at the border2

2

. Commerce Secretary Howard Lutnick later confirmed that Nvidia hadn't sold a single H200 to a Chinese company months after approval, effectively rendering the relaxed export rules meaningless2

2

. This dual squeeze—from both US export restrictions on newer chips and Chinese customs restrictions on approved older models—has pushed buyers toward increasingly desperate measures.

Older Hardware and Gaming GPUs Repurposed as Alternatives Scarce

The supply crunch has extended far beyond Blackwell GPUs to encompass virtually every tier of AI hardware. A100 servers, based on a data-center GPU Nvidia launched five years ago in 2020, have seen their prices roughly triple since late last year, climbing from about 200,000 yuan ($22,300) to as much as 600,000 yuan ($82,000)2

2

. Chinese companies are also buying up older A100 systems and even repurposing gaming graphics cards to run AI workloads, according to chip traders interviewed by the Financial Times1

1

3

3

.

GPU rental rates inside China now match or exceed US prices, reversing the discount that abundant smuggled supply once provided2

2

. Rising memory prices are compounding these challenges, with one trader noting that moving away from Nvidia hardware had become harder as component costs climbed—a knock-on effect from the DRAM and HBM shortage now working through every tier of the AI hardware stack2

2

.

Domestic AI Chip Alternatives Struggle to Fill the Gap

Geopolitical factors have pushed China to mandate that AI firms use at least 50% domestically produced inferencing chips, leading to explosive growth for companies like Huawei and Cambricon Technologies1

1

. Huawei has positioned its Ascend 950PR, launched in March, as the inference chip of choice for domestic firms2

2

. The 950PR is currently undergoing testing at large data center clients in China, but output remains limited, and its native CANN software stack substantially trails Nvidia's industry-standard CUDA platform2

2

.

While domestic alternatives have shown performance and efficiency improvements in recent years, they aren't yet able to compete directly with banned Nvidia GPUs in China, particularly given the maturity and ecosystem advantages of CUDA1

1

. Until Huawei scales production of the 950PR—which will take considerable time—or Beijing greenlights H200 imports, which appears highly unlikely given current tensions, prices for remaining A100 inventory and other smuggled Nvidia AI servers in China will likely continue their upward trajectory2

2

. The situation underscores how US chip controls and Chinese customs restrictions have created a market dynamic where demand far outstrips supply, leaving Chinese tech firms caught between inadequate domestic options and prohibitively expensive black market alternatives.