Tencent Is Reported to Be Renting 100,000 AI Chips From Oracle, a Sign That Cloud Leases Can Route Around Chip Export Limits

For four years, Washington has tried to keep its most advanced AI chips out of China. A reported deal suggests a simpler path for Chinese firms: do not buy the chips, rent them. If confirmed, it would show that the rules govern where silicon ships, not who uses it.

Our research desk summarized a report carried by The Straits Times on October 1. It said Tencent, China's largest games and social-media company, had leased 100,000 advanced AI chips from Oracle in a deal worth about $7 billion. The chips would stay in Oracle's data centers. Tencent would pay over time to train and run large AI models on them.

The desk paired this with a second item: DeepSeek and Huawei releasing opensource programming tools for Huawei's Ascend AI chips. The aim is to reduce Chinese developers' dependence on Nvidia's CUDA software, the standard toolkit for AI chips. The desk read the two moves as complementary: rent American capacity where allowed, build a domestic stack where not.

It also cited an Andreessen Horowitz market presentation framing big-tech capital spending as approaching $1 trillion a year.

On a second check, the deal's origin is clear. The Financial Times broke the story, citing people familiar with the matter, and outlets including The Next Web relayed it on October 2. We found no statement from Tencent or Oracle, so the deal rests on one well-sourced report.

The FT's account adds detail our desk lacked. The lease runs five years, Tencent pays about 30 percent upfront, and the chips sit in several Oracle data centers in Southeast Asia. The chip models were not named. Tencent executives have said the compute will first train larger versions of its Hunyuan models.

The DeepSeek and Huawei release is confirmed. Reuters reported on September 30 that DeepSeek open-sourced programming tools for Huawei's Ascend chips, including a version of its TileLang language and compute and communication libraries, built with Huawei's support. DeepSeek pitches TileLang as simpler than CUDA; no independent benchmarks have been published. We could not confirm the a16z capital-spending figure.

Some background is well established. Since October 2022, U.S. export rules have restricted sales of the most advanced AI accelerators to China and tightened several times since. Those rules focus on physical shipments. Remote access to chips sitting in another country's data centers has been a recognized gray zone, and lawmakers have debated closing it. According to the FT, current U.S. rules permit overseas cloud leases like this one.

Oracle is a major cloud provider with large AI hosting deals. Tencent runs its own cloud business and has strong reasons to secure compute any way it can. So the reported deal is plausible, and the FT's sourcing is strong. Until a company confirms it, the size and terms remain reported, not official.

The software point deserves weight. Hardware alone does not make a chip useful; developers need mature tools, and Nvidia's CUDA has had nearly two decades to mature. Open tools for Huawei's chips, if they work well, would matter over years more than any single lease matters this quarter.

We could not find any comment from Oracle or U.S. regulators.

Most medical AI our readers will meet, from imaging reads to drug discovery to analysis of genomic and wearable data, runs on the same scarce accelerators. When the world's largest firms compete to rent every available chip, prices rise and smaller health-AI developers wait longer.

There is also a privacy angle. If compute increasingly crosses borders by lease, the question of where health data is processed, and under whose laws, becomes harder to answer. Patients and clinicians should expect more questions about data location, not fewer.

• Single-source report. Every account we found traces back to the FT; neither company has confirmed. • "End-run" is an interpretation. The FT says the structure is permitted today. Calling it a workaround is a judgment about the rules' intent, not their letter. • Chip type is unclear. "Advanced AI chips" could mean the most restricted parts or lesser models already allowed. • Open question: whether regulators move to cover remote access, which would change the economics of deals like this.

• A statement from Oracle or Tencent, or a filing, confirming or denying the lease. • Any U.S. Commerce Department guidance on remote cloud access by Chinese firms.

• Rental prices for AI compute, which set how quickly health-AI startups can scale. Not financial advice. Facts and context only.

Claim What We Found Tier

Could Embryos Be Blind to Some Rhythms? A Sampling Idea for the Spine AI Battery-Life Forecasts Keep Improving on Paper. The Real Test Is a Battery the Model Has Never Seen.