The Financial Times reported on Friday that Beijing quietly added Nvidia's RTX 5090D V2 to its list of banned imports on May 15, the same week Jensen Huang was riding Air Force One into Beijing as part of Trump's state visit. The chip had been engineered specifically to satisfy US export controls, a Blackwell- derived gaming GPU with less VRAM and lower bandwidth, sold to Chinese gamers and 3D artists from August onward. Chinese AI developers had been quietly using it too, with the H200s and the proper Blackwell AI accelerators off the table.

That is the part that should land. Nvidia built a chip to comply. China banned the chip anyway.

The geopolitical theatre of this is bleak in a way I find genuinely interesting. The 5090D V2 was Nvidia's attempt to play both sides: meet US export rules, keep Chinese revenue flowing, accept the haircut on VRAM and bandwidth as the price of access. The whole point of the SKU was that Beijing was supposed to want it. The chip's specs were tuned to a regulatory compromise that already conceded most of the high-end AI use case. Then Beijing decided it didn't want the compromise either.

The timing matters. Customs added the chip to the banned list on May 15. Huang boarded Air Force One in Alaska that same week, a late addition to the entourage. The summit happened. By the time anyone outside Beijing knew the chip was banned, the CEO of the company whose chip it was had been physically in the country and back. It reads as a signal sent with diplomatic precision: we are not interested in the de-fanged version of your tech, and we are happy to tell you while you are still here.

What sits underneath the signal is the harder thing. China has been telling its own technology companies to prioritise domestic chips for a while now, and the numbers suggest the message is landing. Zero H200s have been imported despite the US clearing roughly ten Chinese firms to buy them last week. The Huawei Ascend 910B is doing more work than the export-control story usually acknowledges, and DeepSeek's pivot to Ascend silicon for V4 was the kind of move that, if it generalises, eats into the long thesis Nvidia has been quietly leaning on: that even gated access is still access.

Banning the 5090D V2 also closes a quieter loophole. Hobbyist Chinese AI developers, cut off from the proper Blackwell stack, had been using the consumer card to run open-source models on Blackwell compute. The chip was a backdoor that wasn't really a backdoor, a way for capability to leak in via the gaming SKU while the AI SKU stayed off-limits. Cutting it off pushes those developers toward domestic alternatives faster, which is the consistent thread running through everything Beijing is doing on chips right now. The export-control regime didn't manage to keep advanced silicon out of Singapore; Beijing managing to keep it out of its own market is a different problem with a different shape, and one Washington has less leverage to alter.

Nvidia's pitch to its own investors this week was that the $200 billion CPU market it now forecasts includes China. Huang said as much in Taipei on Saturday, on his way out. The hope is real, the licences exist on paper, and the H200 nominally has a route in. The other side keeps banning the chips that route would carry. At some point that becomes less an export-control story and more a buyer story, a market that has decided it no longer wants what is on offer, and what is on offer was already the compromise.

Sources: