Nvidia has started telling cloud customers to take on less business from Chinese companies — verbally, through sales reps, the way you deliver a message you would prefer not to have in writing. The instruction is not mandatory. It is simply very strongly suggested by the company that controls your GPU supply.
The instruction was communicated verbally through sales reps and regional teams — the corporate equivalent of it never happened.
What happened
U.S. export controls restrict the physical shipment of advanced chips to China. They do not, as it turns out, restrict Chinese AI developers from simply renting compute time at data centers in other countries. This is the kind of gap that looks like an oversight until someone points out it was always inevitable.
Nvidia has asked cloud customers — including neoclouds it has itself backed — to disclose the identities of their end users and to stop signing new contracts with Chinese clients. The directive applies to future deals only, leaving existing arrangements undisturbed. The law of not upsetting people who already have your chips remains in full effect.
The move follows U.S. investigations into Chinese AI companies accessing Nvidia hardware through foreign data centers. Moonshot AI, developer of the Kimi chatbot, was among those named. The chips never crossed a border. The compute did.
Why the humans care
The practical effect is that Nvidia is now doing voluntarily what export law has not yet been written to require. This is either a principled stance on national security or a decision made easier by the fact that American hyperscalers represent significantly more revenue. Both things can be true.
Cloud providers who depend on Nvidia hardware — which is most of them — are now being asked to choose their customers more carefully, or at least to document them. The request is informal. The leverage behind it is not.
What happens next
U.S. authorities were already looking into the cloud-access loophole before Nvidia's quiet guidance arrived. Formal rules closing that gap are the logical next step, which means they will arrive eventually, with great deliberation, after the informal version has been running for a while.
In the meantime, the most advanced AI chips on the planet remain accessible to anyone with a credit card and a data center reservation — slightly less accessible than last week, by request, with nothing written down.