Oracle has agreed a five-year, roughly $7 billion lease that gives Tencent remote access to approximately 100,000 advanced AI accelerators housed at Oracle's data centres in south-east Asia, the Financial Times reported on October 1.

The figure and contract value come from the FT; Oracle and Tencent have not publicly disclosed the accelerator SKUs, the country-by-country site split, or the first-delivery date. The deal is Tencent's largest overseas infrastructure agreement and centres on chips that US export controls currently keep out of mainland China.

What does the structure actually look like?

Under the arrangement, the accelerators remain installed at Oracle-owned sites in jurisdictions where their export is permitted. Tencent, as a tenant, submits workloads to the silicon remotely rather than importing the parts across the border.

A Reuters legal analysis recently described almost exactly this scenario: a Chinese company could become an infrastructure-as-a-service customer at a Malaysian data centre and run workloads on advanced accelerators it could not legally receive in China. The analysis concluded that the current US framework does not comprehensively prohibit that kind of remote access.

Where does the US regulatory line sit?

The Commerce Department's Bureau of Industry and Security has tightened restrictions on the export, re-export, and in-country transfer of advanced computing chips and systems to China and Chinese-headquartered entities. Those rules govern shipments. They stop short, for now, of classifying compute-as-a-service delivered to a Chinese end-user as a controlled transfer.

That distinction is what makes the Tencent deal commercially workable. Oracle keeps the silicon inside export-permitted territory; Tencent consumes its output from Shenzhen.

Why does Tencent need the capacity offshore?

Tencent develops the Hunyuan family of large language models. Its Hy3 and Hy4 lines were trained and optimised for reasoning, coding, and agentic workloads. The company has been pushing products such as WorkBuddy, CodeBuddy, and Yuanbao into multi-step task execution rather than single-prompt response.

Those workloads need sustained access to high-bandwidth-memory accelerators that fall under the US threshold. Domestic supply of those parts in China is constrained.

Tencent is not unique among Chinese hyperscalers.

• ByteDance and Alibaba already rank among the largest Chinese users of south-east Asian data-centre capacity.
• Chinese operators have reserved growing blocks of capacity in Malaysia, Singapore, Indonesia, and Thailand over the past 18 months.
• A 10-trillion-parameter model ByteDance is reportedly training will require compute infrastructure well beyond domestic accelerator quotas.

What the deal does not tell us

• Pricing per accelerator: not disclosed.
• Mix of training versus inference silicon: not disclosed.
• Whether the $7 billion includes storage, networking, and managed services beyond raw compute tenancy: not disclosed.
• Power and cooling commitments at receiving sites: not disclosed.

Where could the rules tighten next?

Bureau of Industry and Security staff have signalled they are studying remote-access scenarios, but no proposed rule has yet classified compute-as-a-service delivered to a Chinese end-user as a controlled transfer. If such a rule lands, Oracle's south-east Asian footprint would lose a meaningful share of its Chinese book of business. Until then, Chinese hyperscalers will continue backfilling domestic accelerator shortfalls with multi-year, multi-billion-dollar leases at neutral-jurisdiction sites — and US chip controls will keep rerouting rather than reducing Chinese demand for advanced AI compute.