Prices for Washington’s restricted Nvidia AI semiconductors have more than doubled over the past six months on China’s underground gray and black markets, according to a comprehensive report by the Financial Times.

The dramatic price surge highlights a severe supply crunch as intensified U.S. enforcement actions successfully narrow illicit shipping channels, colliding directly with an unyielding appetite for Nvidia hardware among Chinese artificial intelligence developers.

1. Tracking the Underground Premiums

Interviews with multiple Chinese chip traders reveal that the black-market costs of both enterprise data center racks and elite localized workstations have reached historic highs:

  • The Flagship Blackwell Server (DGX B300): The underground price of Nvidia’s flagship server platform—which houses eight advanced Blackwell graphics processing units—has skyrocketed to over 8 million yuan ($1.1 million). This is a massive jump from the 4 million yuan it commanded late last year, and nearly triple its standard U.S. retail price of roughly $400,000.
  • The Workstation Tier (RTX 6000 Pro): Highly coveted by early-stage Chinese AI startups for deploying large language models, the price of this professional workstation processor has jumped from about 50,000 yuan ($6,900) at the start of the year to as high as 130,000 yuan ($17,900).
  • Desperation for Legacy Silicon (A100): With the latest Blackwell units tightly bottle-necked, buyers are aggressively clearing out older architecture. The price for servers containing legacy A100 accelerators has spiked from 200,000 yuan to 600,000 yuan in just a few months.
                  [Official US Base Price]  VS  [China Black Market Price]
DGX B300 Server  ──► ~$400,000                  ──► ~$1,100,000 (8M RMB)
RTX 6000 Pro     ──► ~$13,250                   ──► ~$17,900    (130K RMB)

2. Why the Loopholes are Shrinking

The explosive price correction is a direct byproduct of shrinking supply, as regulatory authorities aggressively shut down backdoors that previously leaked up to $1 billion in hardware per quarter into mainland China:

  • The Subsidiary Clampdown: The U.S. Department of Commerce enacted strict new guardrails on May 31 to permanently plug a major leak, blocking overseas Chinese business entities and foreign subsidiaries from purchasing high-end chips to route them back home.
  • High-Profile Criminal Crackdowns: Global smuggling networks were thrown into disarray following massive U.S. enforcement actions. Most notably, a federal indictment in March charged a Supermicro co-founder and overseas contractors with running a massive operation that illicitly diverted $2.5 billion in Nvidia-configured hardware to black-listed Chinese buyers through shell firms in Southeast Asia.

3. The Inference Boom & The Software Lock-in

The extreme premium Chinese enterprises are willing to pay underscores two major structural realities within the regional technology landscape:

The Software Moat: Despite heavy political backing and massive state funding pushed toward domestic alternatives like Huawei’s Ascend lineup, local options remain constrained by raw processing performance and, crucially, less mature software stacks. Nvidia’s CUDA ecosystem remains the universal foundation for AI engineering, making alternative integration highly friction-heavy for developers under tight deadlines.

Furthermore, Chinese tech giants are currently locked in a fierce domestic race to commercialize public-facing autonomous agents and generative enterprise applications. This requires a massive, immediate fleet of “inference” chips to serve millions of active users concurrently.

Because building data center grids out of contraband hardware carries extreme long-term risks—with Nvidia publicly warning that smuggled systems are a “dead end” completely devoid of enterprise technical repairs, component warranties, or software support—the spiraling hardware costs have translated directly into the local cloud market. For the first time, the cost of renting advanced GPU compute capacity on independent Chinese platforms has parity with, and in some cases exceeds, standard Western enterprise cloud pricing.