US said to weigh easing curbs on Nvidia H200 exports to China

The US government is weighing whether to permit Nvidia to sell its H200 AI chips to China, with the Commerce Department reviewing a possible change to its export restrictions.

The move follows talks between President Donald Trump and Chinese leader Xi Jinping that led to a truce in their trade and tech dispute.

Washington has previously limited exports of advanced AI chips over concerns they could support Chinese military development.

The H200 chip contains more high-bandwidth memory than its predecessor and is considered faster at processing data.

Currently, Nvidia’s H20 is the most advanced chip allowed for export to China.

Earlier this week, the US approved shipments of up to 70,000 Nvidia Blackwell chips to Saudi Arabia and the United Arab Emirates.

🔗 Source: Reuters

🧠 Food for thought

Implications, context, and why it matters.

H200 exports hinge on chip thresholds

October 2023 BIS set chip-level Total Processing Performance (TPP, the export-control performance metric) plus performance density thresholds that block advanced GPU exports to China 1. December 2024 kept those limits and added High Bandwidth Memory (HBM) plus Semiconductor Manufacturing Equipment (SME) rules 1.
Nvidia’s H20 sits below these caps and leads legal shipments to China 2.
H200 carries more HBM than H100, which already fails the October 2023 thresholds 1.
The new 3.3 GB/s/mm² HBM memory bandwidth density rule targets standalone HBM sales 1 3. Co-packaged HBM on GPUs like H200 falls under chip-level TPP plus performance density caps, not the HBM density control 1 3.
A policy shift that mirrors the proposed B30A path (a mooted Blackwell-series configuration intended to comply with export rules) could permit chips 12 to 17 times stronger than H20 4. That would undercut the current export control plan 4.

HBM3e and CoWoS capacity decide who benefits if H200 ships

Samsung and SK Hynix race to lift HBM output to 150,000 wafer starts per month (a standard measure of fab throughput) by end of 2025, while Micron aims for 90,000 by 2026 5.
TSMC’s Chip-on-Wafer-on-Substrate (CoWoS) packaging is used to integrate HBM with GPUs 6. Capacity should nearly double by 2025, with Nvidia already using over half of total supply 6.
If H200 wins approval, data center operators plus cloud providers would spark a rush for HBM3e 12-high stacks (12-layer memory packages) that are moving into mainstream output, with yields still being tuned 7.
AI data center power needs push grid-scale battery energy storage to 92 GW/240 GWh by 2025 7. Growth runs at 25 to 33 percent each year 7.