Stock Markets April 30, 2026 02:46 AM

China’s Short Supply Pushes Price of Nvidia B300 Servers Near 7 Million Yuan

Crackdown on diversion and surging AI compute demand lift China prices to roughly $1 million per unit

By Maya Rios NVDA AMD
China’s Short Supply Pushes Price of Nvidia B300 Servers Near 7 Million Yuan
NVDA AMD

Tighter controls on the grey market for high-end chips, along with robust demand for AI compute in China, have driven the price of Nvidia’s B300 servers to about 7 million yuan ($1 million) apiece. The devices, designed for intensive AI inference workloads, are restricted from sale in China and face limited official supply, creating a scarcity premium that has nearly doubled local prices since late last year.

Key Points

  • Nvidia B300 servers are trading at about 7 million yuan (~$1 million) in China, driven by tightened controls on grey-market supply and strong AI compute demand.
  • A U.S. B300 server, equipped with eight B300 GPUs, is priced near $550,000 in the United States, up from roughly $500,000 late last year; China’s near-doubling in price reflects a scarcity premium.
  • Rising token consumption by Chinese AI models and export uncertainties around H200 chips are intensifying demand for alternative solutions, including rentals, which have reached 190,000 yuan per month on short-term contracts.

Demand for high-performance artificial intelligence hardware in China has pushed the market price of Nvidia’s most advanced server, the B300, to approximately 7 million yuan each - roughly $1 million - according to industry sources. The escalation in price follows a crackdown on channels that previously supplied grey-market units, and is occurring against a backdrop of elevated computing needs from Chinese technology firms.

Industry participants say the B300’s sticker price in China has climbed steadily this year, and surged after enforcement actions narrowed the flow of units through unofficial routes. Sources requested anonymity because of the sensitivity of the topic.

Nvidia has told partners that the B300 is restricted from sale in China and emphasized the need for strict compliance. In a statement it warned that "As systems become increasingly large and complex, unlawful diversion is a recipe for failure." Nvidia added: "Nvidia does not provide any service or support for such systems, and the enforcement mechanisms are rigorous and effective."


Scarcity premium and price comparators

In the United States, a B300 server - configured with eight B300 GPUs - carries a price tag of about $550,000, up from around $500,000 late last year, according to two industry sources. By contrast, the near-doubling of the going rate in China - from roughly 4 million yuan late last year to about 7 million yuan now - reflects a scarcity premium that market participants attribute to tighter U.S. controls on exports.

That scarcity is occurring at a time when Chinese firms are racing for the most cost-efficient hardware to process tokens - the basic units of text that AI models consume - enabling monetisation of their models and computing stacks. For some buyers, the spike in purchase prices has placed new equipment out of reach, prompting exploration of rental options on shorter-term contracts. Rental rates have climbed as high as 190,000 yuan a month on a one-year, short-term agreement, industry contacts said.


Market dynamics and competing technologies

The price squeeze came into sharper relief after U.S. authorities pursued a high-profile prosecution in March of Yih-Shyan "Wally" Liaw, a co-founder of Nvidia partner Supermicro, creating additional uncertainty for the grey market. Sources say uncertainty around exports of H200 chips has also contributed to upward pressure on B300 prices. Although the H200 has reportedly received approvals from both governments for exports, shipments to China have not yet taken place as the two sides remain at odds over the conditions governing sales.

Chinese AI companies and domestic chipmakers, including the technology giant Huawei, have moved to capitalise on the frictions. These firms aim to chip away at Nvidia’s market-leading share in China, which sits at 55%, while competitor AMD holds about 4% of that market, according to the information provided by industry sources.


Technical profile

The B300 server boasts 288 GB of high-bandwidth memory and delivers 14 petaFLOPS of compute at FP4 precision, placing it among the most powerful current options for AI inference workloads. Nvidia and its partners, including Supermicro, began shipping the chip last September.

At the same time, usage of Chinese AI models has expanded rapidly. Data cited by market analysts show that Chinese models’ share of global token usage rose to 32% in March 2026 from 5% a year earlier, reflecting advances in areas such as coding and agentic capabilities. Specific Chinese models and platforms, including MiniMax, Zhipu and Alibaba’s Qwen, each reported token usage that increased by as much as six to seven times in February and March compared with December, according to a note from a U.S. investment bank.


The combination of constrained legitimate supply, enforcement against diversion, and rapidly rising domestic demand has created a local market where the B300 commands a premium not reflected in U.S. pricing. As some firms seek alternatives to direct ownership of restricted hardware, rental markets have tightened and prices for short-term access have risen in step with purchase costs.

(Exchange rate used in reporting: $1 = 6.8302 Chinese yuan renminbi)

Risks

  • Regulatory enforcement against unlawful diversion of restricted hardware could further limit grey-market supply, maintaining or increasing price pressure on high-end servers - this affects technology, cloud and AI infrastructure sectors.
  • Ongoing uncertainty over conditions for H200 chip exports, despite reported approvals, may prolong supply constraints and keep prices elevated for Nvidia-based systems - impacting vendors, renters, and enterprise AI operators.
  • Sharp price increases for critical AI hardware may push some firms toward short-term rentals or alternative domestic providers, potentially disrupting procurement strategies and investment plans in AI infrastructure.

More from Stock Markets

Brockman Reveals Near-$30 Billion OpenAI Stake and Financial Links to Altman During Musk Trial May 4, 2026 California Launches Probe into Federal Deal That Scrapped Central Coast Offshore Wind Project May 4, 2026 Pilots Union Praises Kirby’s Merger Vision, Stops Short of Endorsing Deal May 4, 2026 Embraer Sees Follow-On Middle East Defense Sales After UAE C-390 Agreement May 4, 2026 Intel hires long-serving Qualcomm executive to oversee PCs and physical AI unit May 4, 2026