An unconfirmed report with outsized consequences: Nvidia has reportedly told global cloud service providers that at least 51 percent of compute capacity in newly signed contracts must be sold to customers outside China.

Existing contracts are not affected. The supply-side constraint is nevertheless tightening.

What the rule targets

Not chip exports. Compute rental.

For the past two years, with direct imports of high-end accelerators restricted, some Chinese companies rented capacity from overseas cloud providers to train models. That path sidestepped hardware controls while still consuming controlled compute.

Requiring providers to sell a majority of new capacity abroad compresses that path at the contract level.

Why now

Per the reporting, this is Nvidia adjusting proactively to US regulatory pressure.

The company faces a genuine dilemma: China is a large market, and compliance risk is high. Moving the rule up to the cloud contract layer is a form of self-certification — we sell a service, and we can document who buys it.

Who feels it

Three groups most directly.

Teams running large pretraining runs that depended on overseas cloud capacity need to re-plan resources. Inference-serving companies are less exposed near term, since inference can run on domestic cards or existing inventory. Companies building their own compute face higher procurement costs.

The indirect effect is price. Tighter supply pushes rental rates up, which eventually shows up in API pricing and product economics.

How far domestic substitution has come

Brokerage research argues that policy and demand together are pushing domestic compute from "usable" to "good." Public progress includes Iluvatar CoreX's Gen4 architecture and Biren's new generation reaching silicon bring-up, with Cambricon, Hygon and MetaX all shipping faster.

Measured data supports partial parity: domestic cards now match or slightly exceed H200 on inference in specific setups. Training is closing but still behind.

What companies should do

1. Build multiple compute sources. Do not depend on a single channel.
2. Reserve headroom for critical training runs instead of buying at the deadline.
3. Run domestic cards on non-critical workloads first to accumulate migration experience.
4. Read contract terms carefully, especially clauses on workload and geography restrictions.

A caveat worth keeping

This is reporting plus brokerage analysis. Nvidia and the cloud providers have not publicly confirmed the details. Treat it as a risk signal that deserves a contingency plan, not as settled fact.

The direction, though, is clear: access to compute is becoming a core competency for AI companies, not merely a cost line.