Quick overview

---

The number of US export licenses for Nvidia will be limited by the limited memory chip supply to sell its H200 AI processors to Chinese consumers, according to the top Republican on the House China committee, citing a rule this week from the Commerce Department.

Representative John Moolenaar described the lack of dynamic random-access memory, or DRAM, as an “immediate challenge” under the new licensing conditions in a letter to Commerce Secretary Howard Lutnick. DRAM is a critical input for AI accelerators. Exporters must certify that authorized shipments from China won’t result in a shortage in the US market as part of these requirements.

Moolenaar wrote, referring to a type of high-bandwidth memory, “Due to severe supply constraints, chips equipped with HMB3E bound for China represent an opportunity cost when it comes to HMB3E being utilized by American customers.”

The regulations, which Moolenaar mainly applauded, require businesses to confirm that their sales in China won’t delay US buyers of AI chips or divert foundry capacity that could be used to fulfill US orders.

According to the letter, Moolenaar supported GAIN AI legislation last year, but Nvidia and the White House can successfully lobby for its removal from a crucial defense bill. Three major companies produce the high-bandwidth memory needed for AI accelerators, which is made of stacked DRAM: Samsung Electronics Co., SK Hynix Inc., and Micron Technology Inc., which have all issued warnings about a shortage in recent months.

The AI data center boom has caused demand for the components, which were previously considered a commodity, to soar.
NVIDIA CEO Jensen Huang acknowledged the memory shortage at the CES conference earlier this month, but he insisted that his company is protected because it is the only purchaser of the newest memory generation, HBM4, which is utilized in its upcoming line of processors designed by Vera Rubin.