Tencent has signed a five-year agreement with Oracle to lease roughly 100,000 advanced AI chips for $7 billion, according to a Financial Times report published September 30. Reuters picked up the story the same day but said it could not independently verify the figures. Tencent is paying around 30% of the deal value upfront. The chips will run across Oracle's data centers in Asia.

The contract is the largest overseas compute lease Tencent has ever signed. It also exposes the central fault line in China's semiconductor policy: Chinese companies cannot buy advanced AI chips outright under U.S. export rules, but those same rules still permit leasing compute capacity hosted in data centers abroad. Tencent is using that gap directly. The hardware never touches mainland soil, and the transaction never crosses the desk of China's National Development and Reform Commission, which must approve chip imports.

Tencent is not the first Chinese company to take this route. ByteDance already ranks among Oracle's largest GPU customers in the Asia-Pacific region. OpenAI also leases Oracle capacity under the broader $300 billion arrangement that has turned Oracle into one of the dominant infrastructure landlords of the AI buildout. Tencent's entry into that customer list signals that even China's best-capitalized technology company prefers renting American compute by the hour over staking its frontier AI roadmap entirely on domestic silicon.

The deal lands during a limited thaw in U.S.-China chip trade rather than a genuine opening. Nvidia has begun shipping H200 chips into China under Beijing's supervision, and ByteDance and Tencent have each received about 10,000 units over the past few weeks, according to the FT. U.S. licensing rules allow each buyer to import up to 75,000 H200s, so the 10,000 delivered so far represent roughly 13% of Tencent's legal ceiling. The NDRC must sign off on every additional shipment, and it is not moving quickly. Beijing would rather direct that demand, and Tencent's money, toward Huawei, Cambricon, and the rest of the domestic chip industry it has spent years building.

Domestic supply cannot yet close the gap. CXMT, China's leading memory manufacturer, is expected to reach roughly 300,000 wafers per month of production by the end of 2026, but it is prioritizing Huawei's orders over all other customers, including global PC makers now seeking allocation. The binding constraint is high bandwidth memory. Huawei needs HBM to pair with its Ascend 910C chips, and CXMT's planned output of roughly 2 million HBM stacks this year translates into only 250,000 to 300,000 Ascend-equivalent chips — far short of what a company of Tencent's scale would need to train frontier models. The shortage became acute enough that Huawei, Cambricon, MetaX, and Iluvatar CoreX all raised prices within the same few weeks of September, according to a Reuters exclusive on the HBM crunch.

Faced with that arithmetic, Tencent chose to write Oracle a check rather than wait years for Huawei's supply chain to mature. The lease delivers 100,000 chips' worth of training capacity immediately and legally, under rules that were never designed to block this form of access.

The deal does not mean U.S. export controls have failed. The leased chips still never enter China; Tencent rents compute abroad precisely because it cannot import that volume at home. The controls are functioning as designed — they are redirecting Chinese AI spending offshore rather than halting it. Washington can state that the hardware stayed out of Chinese territory. Beijing is left watching its own champions rent American silicon instead of trusting domestic chips with their AI roadmaps. And Oracle collects $7 billion from a customer the U.S. government never had to approve.