Tencent Leases 100,000 AI Chips From Oracle to Bypass US Export Rules

Tencent has arranged to lease 100,000 AI chips from Oracle, according to a report by Startup Fortune, using a cloud rental structure that keeps the hardware outside the reach of United States export controls on China.

The deal, as described by the outlet, does not move chips across any border. Tencent would buy compute capacity on Oracle's cloud infrastructure, built outside China, rather than take physical delivery of the accelerators. That distinction matters because US rules restrict the sale and shipment of advanced AI chips to Chinese entities, not the sale of compute time on servers already running in other countries.

How does the arrangement work?

Under the reported structure, Oracle owns the hardware and hosts it in its own data centers. Tencent pays for access to the resulting processing power. No Chinese company imports a controlled chip, so no export license is triggered under the rules as they stand today.

This is the same logic that has made cloud compute a work-around across the industry. A chip that cannot be sold to Shanghai can still serve a customer from a data center in, say, Texas or Tokyo. Startup Fortune frames the Tencent-Oracle deal as the largest single instance of this pattern reported so far, at 100,000 chips.

For Tencent, the commercial motive is direct. Its Hunyuan large language models and its AI services across WeChat and its cloud unit need Nvidia-class accelerators in volume. Domestic alternatives exist, but they trail in performance and in the maturity of the software stack that surrounds them.

What does it change for the supply chain?

The report points to a gap that regulators have discussed for more than a year. Export controls tighten the flow of physical chips; they do not yet restrict a Chinese company's ability to rent the same silicon at a distance. US officials have periodically floated controls on cloud access for Chinese customers, but no rule of that kind is in force, and Startup Fortune does not report any imminent change.

For Oracle, the deal fits its strategy of renting out AI infrastructure at scale. The company has spent heavily on GPU clusters to compete with Amazon Web Services, Microsoft Azure and Google Cloud for AI training workloads. A single 100,000-chip customer would rank among the largest AI compute commitments on record, if the figure holds.

For Nvidia, the arrangement is double-edged. Oracle still has to buy the accelerators, so demand is not lost. But the transaction routes the value through an American intermediary, and it hands Oracle, not Nvidia, the recurring revenue from the Chinese end user.

Is the model durable?

The weakness is regulatory. Any future rule that treats cloud compute as an export would break the model overnight, and US policy in this area has moved in one direction since 2022. Tencent, Oracle and every other party using this structure is building on ground that Washington can rezone.

Startup Fortune reports no comment from Tencent or Oracle on the arrangement, and the contract terms remain undisclosed. Until regulators close the cloud gap, or the parties confirm the deal publicly, the 100,000-chip figure stands as a reported number rather than a verified one, and the market will keep pricing in the risk that the loophole is temporary.