H20 Still Under Scrutiny as NVIDIA Prepares New China-Specific Chip

Despite ongoing concerns over potential backdoor vulnerabilities in the H20 accelerator, NVIDIA is reportedly developing a fresh AI chip for the Chinese market. Sources speaking to Reuters on August 19 indicate that this new processor, provisionally called the B30A, will be built on NVIDIA’s latest Blackwell architecture and is expected to deliver performance surpassing that of the H20.
Insiders describe the B30A as adopting a single-die design, integrating all major components onto one silicon wafer, in contrast to the dual-die configuration of NVIDIA’s top-tier B300. Early estimates suggest the B30A’s raw compute power will be roughly half that of the dual-die B300, yet it will still feature high-bandwidth memory and NVIDIA’s NVLink interconnect technology to facilitate rapid data exchanges between processors. By comparison, the H20—recently cleared for Chinese sales—remains based on the older Hopper architecture.

Although NVIDIA has yet to finalize the B30A’s specifications, the company plans to begin delivering sample units to selected Chinese customers as soon as next month for evaluation. In a statement, NVIDIA emphasized that it is “evaluating multiple products for inclusion in our roadmap to ensure readiness to compete within the bounds of government regulations,” and asserted that “all products we offer have received full approval from relevant authorities and are intended solely for legitimate commercial use.” Nevertheless, industry observers caution that securing U.S. export licenses remains far from guaranteed.

Recent reports in the Financial Times allege that NVIDIA and AMD agreed to remit 15% of their China chip-sales revenue to the U.S. government in exchange for export permissions—a claim later confirmed by President Trump. According to Trump, he originally sought a 20% levy but settled on 15% after discussions with NVIDIA CEO Jensen Huang.

The H20 itself was introduced at the end of 2023 as a pared-down AI accelerator tailored for China under U.S. export restrictions. Trading at only 15%–30% of the flagship H100’s performance, the H20 still accounted for 80% of NVIDIA’s China revenue, leading to a $5.5 billion impairment when the export ban first took effect. In April, the U.S. government prohibited its sale in China without special authorization; by July, NVIDIA announced that the H20 would return to the Chinese market, alongside a new GPU fully compatible with local regulations.

Subsequent disclosures that NVIDIA had effectively paid a 15% “protection fee” to obtain the export license prompted swift criticism from U.S. lawmakers and national security officials. President Trump defended the decision by labeling the H20 “essentially an old chip” with “outdated” performance and noting that Huawei already produces similar components. Analysts interpret Trump’s stance as an admission that banning the H20 no longer served strategic interests and that resuming sales would help NVIDIA preserve market share in China.

In China, the H20’s reinstatement was met with a muted response. CNN observed that Beijing’s restrained reaction underscores its confidence in forging a self-sufficient semiconductor ecosystem and in the rapid progress of its domestic chip industry. Several executives from local chip firms told 36Kr that leading domestic manufacturers had sold out their production capacity in the first half of the year, and that native GPU vendors were seizing the opportunity to challenge the H20’s position.

This competitive momentum is reflected in IDC data showing that China’s homegrown compute accelerators grew their market share from 14% in 2023 to 34.6% in 2024. Projections suggest domestic AI chips could claim 55% of the market by 2027, up from 17% in 2023, while U.S. suppliers like NVIDIA and AMD would see their share shrink from 83% to 45%.

Security concerns continue to shadow the H20. On July 31, China’s Cyberspace Administration summoned NVIDIA to explain alleged backdoor vulnerabilities and submit related documentation. Social media investigator @玉渊谭天 has highlighted U.S. discussions about embedding government-accessible “backdoors” in AI chips as a condition for relaxed export controls for “low-risk” Chinese clients, condemning the H20 as neither advanced nor secure.

Xiang Ligang, chairman of the Zhongguancun Information Consumption Alliance, told foreign media that China views these U.S. tactics as patently unfair. “You refuse to sell us what we truly need, yet you’re dumping outdated products into our market to seize share—do you really think we’re that naïve?” he asked.

Amid these tensions, reports emerged on August 18 that NVIDIA plans to offset the 15% fee by hiking H20 prices by roughly 18%. Gene Munster, co-founder of Deepwater Asset Management, suggested that this cost transfer is necessary to maintain NVIDIA’s margins. Yet industry insiders argue this strategy overlooks the H20’s precarious standing in China: until NVIDIA addresses the security doubts surrounding the chip, neither state-owned enterprises nor sensitive private firms are likely to prioritize its procurement, casting doubt on its future sales performance in the world’s largest AI market.