How NVIDIA Allegedly Plans To Dodge Blackwell AI Chip Export Restrictions To China
NVIDIA is planning to ship a cut-down version of its RTX Pro 6000 server-class graphics card with key specification downgrades to China, in order to sidestep U.S.-imposed artificial intelligence (AI) chip export restrictions, according to multiple sources who spoke with Reuters on the matter. It's essentially the same strategy NVIDIA has taken with its consumer graphics card lineups.
For example, during the Ada Lovelace era, NVIDIA shipped a GeForce RTX 4090D to China with fewer CUDA cores (14,592 versus 16,834), Tensor cores (456 versus 512), and RT cores (114 versus 128) compared to the regular GeForce RTX 4090. It also scaled back the total graphics power (TGP) rating (425W versus 450W), but kept the clock speeds and memory configuration the same.
NVIDIA similarly offered a downgraded version of its Blackwell-based GeForce RTX 5090 to China, though with a less aggressive neutering, at least in some parts. Compared to the 5090, the Geforce RTX 5090D sports the same number of CUDA cores and the same 32GB of GDDR7 linked to a 512-bit bus, but it offers considerably fewer AI TOPS (2,375 versus 3,352). That amounts to a 70% reduction in AI performance.
Just as it's done in the consumer GPU segment, NVIDIA is evidently planning to ship a partially neutered Blackwell GPU called B40 to China, in place of its banned Hopper-based H200 and H100 GPUs.
According to the aforementioned sources, the B40 chip will be based on NVIDIA's RTX Pro 6000D server-class solution, but with GDDR7 in place of high bandwidth memory (HBM) for the VRAM. The sources also say the cut-down part will skip TSMC's advanced Chip-on-Wafer-on-Substrate (CoWoS) packaging.
Other details are not yet known, though we can assume that other key specifications will be downgraded as well, so as not to run afoul of existing chip export rules. This will also be reflected in a cheaper price tag, with the cut-down part reportedly set to sell for between $6,500 to $8,000 versus $10,000 to $12,000 for the H20, the sources claim.
It will be interesting to see how this plays out. NVIDIA CEO Jensen Huang recently went on record condemning the U.S.'s ban on AI chip exports "a failure." Of course, NVIDIA has a vested interest in the AI chip race, with its data center division raking in tens of billions of dollars each quarter, including a record $35.58 billion last quarter. For comparison, NVIDIA's gaming division generated $2.54 billion in the same time frame.
