US to allow Nvidia’s H200 chip shipments to China, sources say

Nvidia is set to receive US approval to export its H200 AI chips to China, according to sources cited by Bloomberg.

The decision follows weeks of discussions within the Trump administration, and comes after Nvidia CEO Jensen Huang met with President Trump to discuss export controls.

The H200 chip is less advanced than Nvidia’s Blackwell design, which the company had previously sought to sell in China.

An official announcement could come as early as December 8, Bloomberg reported.

This move would be a shift from previous US restrictions, which aimed to limit China’s access to advanced American technology.

Earlier, Nvidia was allowed to sell its less-powerful H20 chip in China, but Chinese authorities reportedly told local firms to avoid it, and buy domestic processors instead.

Some US lawmakers have voiced concerns that selling the H200 to China could strengthen the country’s military and weaken US tech leadership.

🔗 Source: Bloomberg

🧠 Food for thought

Implications, context, and why it matters.

Any H200 approval likely comes with conditions; Chinese AI clusters would still face cost gaps

Open questions cover blanket approval versus case-by-case licenses from the U.S. Commerce Department’s Bureau of Industry and Security (BIS), plus limits on volumes or end-user checks; BIS recently sent “is informed” letters that put chip-specific licenses on exporters, which makes conditions plausible 1“>1.
H200 beats current U.S. export control limits on total processing performance by almost 10x 2“>2.
Even with approval, H200 builds face higher costs; Chinese labs could hit U.S.-class training power at about 50% extra, while some Blackwell-class options land near 20% 2“>2.
High Bandwidth Memory (HBM) in HBM3e on H200 tops the H100’s HBM3 for bandwidth, which boosts inference efficiency while enabling advanced services without frontier training scale 2“>2.

Infrastructure vendors can benefit from China’s move to 800G Ethernet and liquid cooling under GPU limits

If shipments go ahead large AI clusters in China will lean on 400G or 800G Ethernet fabrics 3“>3. That shift lifts demand for switches, transceivers and high-speed active electrical cables 3“>3.
In China, hyperscale buildouts (mega-scale data centers run by cloud platforms and large internet firms) are moving to 400G or 800G optical backbones with liquid-cooled racks over 100 kW each, which reinforces demand 4“>4.
Suppliers offering low-latency 800G plus 1.6 terabit-per-second (1.6T) solutions include Marvell, Credo and Ciena 5“>5. Power-cooling firms such as Delta, Vertiv ship Coolant Distribution Units (CDUs) up to 2,000 kW 6“>6.