Nvidia cooks up new China-specific AI chips after H20 ban

Nvidia is pulling out its drawing board once again.

Just weeks after Washington effectively killed off Nvidia’s H20 GPU, the chip giant is reportedly designing a brand-new China-specific processor, according to The Information. Nvidia has already briefed major Chinese tech players, including Alibaba, Tencent, and ByteDance, about the incoming compliant hardware.

Here is what you need to know:

Threading the regulatory needle for the third time

Giving up on China is simply not an option for Jensen Huang. Last fiscal year, the market generated $17 billion, accounting for 13% of Nvidia’s total revenue. Analysts warn that recent restrictions could slash its data center sales in the region to nearly zero.

The latest crisis struck on April 15, when the U.S. Commerce Department imposed an "indefinite license" requirement on Nvidia’s H20 and AMD’s MI308 for shipment to China, Hong Kong, and Macau. The rule effectively barred the H20, which Nvidia had specifically downclocked to a performance density of roughly 2.9 TFLOPS/mm² (compared to 19.4 on the flagship H100) to clear previous export hurdles.

That policy move came with a brutal price tag: Nvidia took a massive $5.5 billion charge against quarterly earnings to cover stranded inventory and unfillable supply commitments.

Yet Huang wasted no time pushing back. By April 17, he was in Beijing meeting Chinese trade officials and Vice Premier He Lifeng to reaffirm Nvidia’s commitment to the market. Nvidia also clashed publicly with Anthropic over proposed diffusion rules, with an Nvidia spokesperson mockingly dismissing claims that massive enterprise servers could somehow be smuggled in "baby bumps" or "alongside live lobsters."

Huawei is stepping into the vacuum

Nvidia’s urgency is completely justified because domestic rival Huawei is aggressively capitalizing on the opening.

Reports indicate Huawei is beginning mass shipments of its Ascend 910C processor this month. Built using advanced chiplet packaging to combine two older 910B dies, the 910C reportedly rivals Nvidia’s H100 in sheer compute. Major state-backed telecom operators and AI giants like ByteDance are already ratcheting up orders.

Huawei is also moving rapidly on next-generation hardware. It announced the Ascend 920 shortly after the H20 ban, aiming for mass production later this year with over 900 TFLOPS of BF16 compute and 4000 GB/s bandwidth via HBM3. An even faster Ascend 910D is also waiting in the wings.

On the cluster side, Huawei is deploying the CloudMatrix 384, a massive rack combining 384 Ascend 910C chips. While it burns significantly more power than Nvidia’s GB200 NVL72, lower domestic industrial electricity rates make that trade-off practical for Chinese data centers.

Why it matters

We think Nvidia's endless game of cat-and-mouse with export controls is reaching the point of diminishing returns.

For Chinese tech giants, CUDA and Nvidia software remain deeply convenient, but relying on them has become an unacceptable supply-chain risk. Alibaba and ByteDance will gladly test whatever new compliant silicon Nvidia produces, but their long-term infrastructure budgets are clearly pivoting elsewhere.

For the broader industry, Washington's moving goalposts are creating an unintended consequence: forcing Chinese hyperscalers directly into Huawei’s arms. By banning the H20, the U.S. eliminated the "good enough" fallback that kept developers tethered to the Western software stack.

It looks like Nvidia is caught in a trap where each redesign delivers worse cost-performance while local chips get steadily better. If this next China-specific chip is throttled any further, it may not matter whether Washington approves it or not, because developers won’t bother buying it.