The artificial intelligence (AI) race increasingly has two parallel problems: Companies can't get enough computing power, while governments are deciding who should be allowed to buy it. Nowhere is that tension greater than in China.
U.S. export controls have restricted sales of advanced AI processors, while Beijing has pushed domestic companies toward homegrown alternatives. Yet China's appetite for AI compute keeps expanding faster than local manufacturers can supply it. That imbalance could crack open a door that seemed firmly shut — and Nvidia (NVDA) may be preparing to walk through it.
More Top Stocks Daily: Go behind Wall Street’s hottest headlines with Barchart’s Active Investor newsletter.
www.barchart.com From 95% to Almost Nothing
Nvidia once controlled roughly 95% of China's advanced AI-chip market. CEO Jensen Huang says U.S. export restrictions subsequently drove its market share effectively to zero as domestic suppliers, led by Huawei, filled the gap.
Both Washington and Beijing have since shown some willingness to loosen restrictions, but Nvidia has gained little meaningful ground. H200 shipments to China represented less than 1% of data-center revenue in Nvidia's fiscal second quarter, and the company still isn't assuming any China data-center compute revenue in its current-quarter forecast.
However, The Information reports that China's Ministry of Industry and Information Technology has asked companies including Alibaba (BABA) and ByteDance how many of Nvidia's new RTX Pro 5500 GPUs they want and how they intend to use them. Beijing has reportedly indicated it could approve purchases, although neither quantities nor timing have been finalized. ByteDance alone is considering buying 1 million chips.
A Workstation GPU With Data Center Ambitions
The RTX Pro 5500 isn't a B200-style data-center accelerator. It's a Blackwell workstation GPU with 84 GB of ECC GDDR7 memory, 1,398 GB/s of memory bandwidth, PCIe Gen 5 connectivity, and up to 600 watts of power consumption. Nvidia also offers a liquid-cooled RXM configuration.
Yet the RTX Pro 5500 isn't an ordinary workstation card Chinese hyperscalers would need to jerry-rig into a data center. Nvidia designed it for rack-mounted deployments, offers air- and liquid-cooled configurations, and gave it 84 GB of ECC GDDR7 memory. It even supports Multi-Instance GPU (MIG), allowing each card to be divided into two isolated compute environments with up to 42 GB of memory apiece.
The Information says prospective Chinese buyers plan to install the GPUs directly into servers, with at least one considering eight RTX Pro 5500s per machine for running larger AI models. In other words, companies like ByteDance don't necessarily need to bootleg workstation hardware into something it was never designed to be. Nvidia has already engineered away several of the practical barriers that would normally make deploying workstation GPUs by the hundreds of thousands difficult.
Admittedly, that doesn't turn an RTX Pro 5500 into a B200. The workstation GPU lacks the high-speed GPU interconnects that allow Nvidia's dedicated data-center accelerators to pool thousands of GPUs efficiently into enormous AI clusters. Chinese hyperscalers would instead need to rely more heavily on high-speed Ethernet networking and customized software to coordinate huge numbers of individual PCIe GPUs.
But scale can compensate for some of that disadvantage. One million Blackwell GPUs — each carrying 84 GB of memory — would represent 84 petabytes of aggregate GPU memory. Even if such a cluster were less efficient than one built around Nvidia's purpose-built data-center accelerators, that's an enormous pool of AI computing power.
Ironically, the RTX Pro 5500's lower capabilities could also make the chip easier to export. The U.S. government has not publicly ruled on the RTX Pro 5500, but The Information says some industry executives expect it to remain outside existing restrictions.
That could change, though. Nvidia's second-quarter 10-Q, for example, says U.S. controls consider several characteristics, “including the total processing performance of a chip, the ‘performance density’ of a chip, the interconnect bandwidth of a chip, and the memory bandwidth of a chip.” If President Donald Trump's administration concludes Chinese hyperscalers are using workstation GPUs to circumvent restrictions on advanced AI accelerators, Washington could tighten the rules again.
The Numbers Could Get Big Fast
The RTX Pro 5500 reportedly costs roughly 85,000 to 90,000 yuan, or about $13,000, in China. A 1-million-chip purchase from ByteDance could therefore approach $13 billion at reported market prices. Nvidia reportedly plans to supply China with approximately 500,000 chips per quarter beginning in late December, meaning ByteDance's contemplated order alone could consume two quarters of planned supply.
That's real money, but investors should keep perspective. Nvidia generated $96.2 billion of revenue last quarter alone, including $89 billion from the data-center unit.
Bottom Line
In short, Nvidia isn't suddenly reclaiming China. A huge ByteDance order would instead represent something potentially more important: Proof that Chinese demand for Nvidia's ecosystem remains alive despite years of efforts to replace it.
For NVDA stock shareholders, China should still be treated as upside rather than part of the core thesis, as the company's own guidance assumes zero China data-center compute revenue. If RTX Pro 5500 sales survive regulators on both sides of the Pacific, Nvidia could regain a valuable piece of a market it once dominated — but one workstation GPU won't turn the clock back to 95% market share.
www.barchart.com On the date of publication, Rich Duprey did not have (either directly or indirectly) positions in any of the securities mentioned in this article. All information and data in this article is solely for informational purposes. For more information please view the Barchart Disclosure Policy here.