According to multiple reports, AI giant NVIDIA will launch a “new special-supply” AI chip for the Chinese market, featuring its latest Blackwell architecture. This chip will be priced significantly lower than the restricted H20 and is expected to enter mass production as early as June.
It’s understood that this custom-made Blackwell GPU for the Chinese market is estimated to cost $6,500-$8,000 (approximately ¥46,674-¥57,445 RMB), which is considerably less than the H20’s price of $10,000-$12,000.
Furthermore, its specifications will be significantly reduced. Sources indicate that this Blackwell-architecture chip will be based on NVIDIA’s server GPU RTX Pro 6000D, utilizing traditional GDDR7 memory instead of advanced HBM high-bandwidth memory. It also won’t employ TSMC’s advanced CoWoS packaging technology.
Currently, its final name remains unconfirmed. GF Securities stated in a report that the new GPU might be called the 6000D or B40, with specifications falling precisely within export control limits.
In response, an NVIDIA spokesperson stated that the company is still evaluating its limited options, saying, “We are actually locked out of the $50 billion Chinese data center market until we identify a new product design and gain approval from the U.S. government.”
China remains a massive market for NVIDIA, accounting for 13% of its sales in the last fiscal year. Sources reveal that after the U.S. banned the H20 in April, NVIDIA initially considered developing a downgraded version of the H20 for China, but that plan was unsuccessful.
NVIDIA CEO Jensen Huang recently stated that, under current U.S. export restrictions, the older Hopper architecture used in the H20 can no longer accommodate further modifications.
Huang also mentioned that NVIDIA’s market share in China has plummeted from a previous 95% to the current 50%, with its main competitor being Huawei, which produces the Ascend 910B chip.
He further warned that if the U.S. continues to impose export restrictions, more Chinese customers will purchase Huawei’s chips.
NVIDIA’s AI Chips Find New Demand Despite China Market Loss
Although NVIDIA’s AI chips lost a significant portion of the Chinese market, resulting in a $15 billion loss in a single quarter, their unique products continue to attract large customers.
Reportedly, Oracle has purchased as many as 400,000 of the latest Blackwell GB200 AI chips from NVIDIA, totaling approximately $40 billion, with an average unit price of $50,000!
These AI chips will be used to build a new AI data center for OpenAI in Abilene, Texas, which is the launch site of the U.S. AI infrastructure project “Stargate.”
At the call of Donald Trump, OpenAI, Oracle, SoftBank, and the UAE plan to invest up to $500 billion in this project.
The new AI data center will cover approximately 3.5 square kilometers and include eight buildings. Construction began in June last year, and Oracle has leased the site for 15 years for OpenAI to train its next-generation AI large language models.
OpenAI currently relies on Microsoft’s data centers but can no longer meet its demands.
Recently, OpenAI also announced the “Stargate UAE” project, which will involve building a massive data center cluster in Abu Dhabi, planned to be equipped with over 2 million NVIDIA GB200 chips, expected to be operational by 2026.
Additionally, Elon Musk’s xAI Colossus 2 supercomputer is planned to feature 1 million GPUs.
