- The Institute for AI Policy and Strategy says US-designed compute is still legally rentable to foreign entities barred from buying the hardware, the South China Morning Post reports.
- The Institute for Progress estimates China's 2027 AI compute at 8.8 million H100-equivalents, about 65 per cent on US-designed chips, and urged tighter limits on overseas cloud access.
- Providers and buyers of cloud AI capacity should ask who verifies customer identity and ownership, and what their contracts say if remote-access rules change.
Rules on chip sales cover hardware. Renting works differently. A customer in one country can use chips that sit in another. Two Washington policy groups have published research touching on this, as reported by the South China Morning Post (SCMP). One of them, the Institute for AI Policy and Strategy (IAPS), says the rental route is still open in law.
What the two reports say
The Institute for AI Policy and Strategy published its report on Tuesday. It says US-designed compute can still be legally rented by foreign entities that are otherwise barred from buying the hardware.
IAPS wants the US Bureau of Industry and Security to limit Chinese access to cloud capacity big enough to train frontier AI models or run them at scale. It also wants providers to check more carefully who their customers are and who owns them.
The Institute for Progress (IFP) is a separate group with a separate report. It urged US policymakers to tighten controls on overseas cloud access, chip smuggling and legal exports. The SCMP account does not say IFP described rental as legal.
The numbers behind the concern
IFP's report measures compute in "H100-equivalents". That unit is benchmarked against Nvidia's data-centre graphics processing units (GPUs). It lets analysts count many different chips as if they were one model.
By IFP's estimate, Chinese entities have about 4.6 million H100-equivalents of compute, and its figure for 2027 is 8.8 million. About 65 per cent of that 2027 capacity would come from US-designed chips, which IFP reads as continued reliance on foreign hardware.
Remote access is a part of this. IFP puts it at roughly 1 million H100-equivalents this year and at 2.8 million by the end of 2027.
These are estimates. The SCMP summary does not say how IFP converts other chips into H100-equivalents, so the totals carry that uncertainty.
Why renting is the weak point
Buying a controlled chip means passing an export check. Renting capacity is a service contract. A rule on renting depends on the provider knowing who is behind an account. That is why IAPS pairs its access limits with identity and ownership checks.
The SCMP account describes these as calls for action. It does not report any change to the rules on cloud access.
A large compute gap, a small capability gap
IFP puts the US lead in AI training compute at about 16 times China's. The same report finds that leading Chinese models trail their American rivals by only 6.5 months.
The summary does not explain how both can hold. Together the two figures warn against treating chip counts as a direct measure of who is ahead.
Chip sales and the domestic alternative
Direct chip sales are also under review. In January, Washington opened case-by-case export reviews for Nvidia's H200, Advanced Micro Devices' MI325X and comparable processors. Two conditions apply. American sellers must show the sales leave US customers' supply intact. Chinese buyers must put formal export compliance procedures in place and send the chips for independent testing inside the US.
A Shenzhen-based company reported more than 1,000 Atlas 900 A3 SuperPoD units already deployed. It also said its Atlas 950 system has moved into commercial service at scale. These designs link thousands of individual processors so they can share heavy AI workloads.
Scaling that hardware is hard. Epoch AI, a San Francisco research group, named high-bandwidth memory as the main constraint on Huawei's output in an updated report last month. This is the fast memory that feeds data to AI chips. Epoch AI put Huawei's chip output this year at below 4 per cent of the computing power Nvidia ships worldwide.
What to ask your team
If you run or resell GPU capacity, ask how you verify who a customer is and who owns it. The IAPS proposal puts those checks at its centre.
If you buy AI capacity, ask where your provider's hardware sits and which compliance terms your contract carries. Ask what happens to your service if rules on remote access change.
If your board tracks AI competition, ask whether the numbers in front of it measure chips, access or results. The IFP figures show these can point in different directions.
The lesson is simple. Controls on chips decide who can own them. Controls on accounts decide who can use them.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: South China Morning Post.
