NVIDIA’s H20 AI chips have run into a new roadblock—and this time it isn’t Washington. Despite a recent green light from U.S. authorities to resume certain AI chip exports, NVIDIA reported zero H20 sales to China-based customers in the second quarter, aside from deliveries tied to prior orders. The sticking point now is Beijing, where regulators are probing the H20 for alleged security backdoors and urging major tech firms to prioritize domestic alternatives.
The pause showed up plainly in NVIDIA’s latest results. China has been largely absent from the company’s earnings for several quarters, and Q2 was no different. Executives noted that a single “previously reserved” batch of H20 chips worth roughly $650 million was picked up by a customer, but that transaction was not included in Q2 figures. In short, fresh traction in China hasn’t materialized, creating a stalemate for a market that could be worth tens of billions over time.
According to industry commentary around the results, the company hasn’t shipped any H20 units into China and isn’t factoring H20 into its Q3 outlook either. If licenses are approved and the logjam breaks, NVIDIA believes it could book $2 billion to $5 billion in H20 revenue with inventory already on hand. The company is also continuing to seek approval to sell next-generation Blackwell-based accelerators into the market.
Beijing’s scrutiny has intensified after President Trump made security assurances a national priority within his AI action plan. That stance has fed skepticism inside China about foreign AI hardware, leading regulators to examine the H20 for potential backdoors. At the same time, government guidance has pushed large domestic platforms to build on fully homegrown stacks wherever possible. Chinese chipmakers, including Cambricon and leading system vendors, have stepped up efforts to fill the gap, while firms like Huawei are rapidly iterating on their own accelerators. Still, swapping out a mature ecosystem like NVIDIA’s doesn’t happen overnight. The performance, software tooling, and developer base that surround NVIDIA GPUs remain difficult to replicate quickly.
NVIDIA’s counterstrategy centers on making its offerings so compelling that local buyers can’t ignore them. The company is positioning Blackwell-based parts as a leap in performance and efficiency, hoping that, once approved, they will become the default choice for large-scale training and inference—even amid political headwinds. To facilitate that outcome, leadership has signaled a willingness to make concessions to U.S. authorities, including the possibility of sharing a portion of revenue related to China sales, in exchange for a stable and compliant export pathway.
The risk for NVIDIA is that time spent on the sidelines allows domestic competitors to gain share and lock in long-term deployments. Every quarter without clarity increases the odds that Chinese hyperscalers, research institutes, and AI startups standardize on homegrown silicon and software. That shift could rewrite procurement patterns for years and erode the dominance of U.S.-origin accelerators in one of the world’s largest AI markets.
What happens next hinges on three levers:
– The outcome of Beijing’s investigation into alleged backdoors and any follow-on certification requirements.
– U.S. licensing decisions governing both H20 and Blackwell shipments.
– How quickly Chinese vendors can close the performance and ecosystem gap if the stalemate persists.
For now, the bottleneck has moved from Washington to Beijing. NVIDIA’s China ambitions remain on hold, and the company is navigating a delicate balance: comply with evolving U.S. rules, address China’s security concerns, and keep its technology roadmap attractive enough that, once the gates open, customers are ready to buy. The stakes are enormous—NVIDIA estimates immediate billions on the table for H20 alone, with much more tied to the broader transition to Blackwell. Whether that potential turns into booked revenue will depend on geopolitics as much as on gigaflops.
