Tencent Is Reported to Have Rented 100,000 Chips It Cannot Buy. Export Controls Regulate Where Compute Sits, Not Who Uses It

The Financial Times reported on September 30 that Tencent has agreed to lease roughly 100,000 advanced AI chips from Oracle, a five-year arrangement worth about $7 billion with close to 30% paid upfront, with the hardware sitting in Oracle data centres across Southeast Asia. Reuters carried the report the same evening. Neither company has confirmed the terms, so the numbers rest on the FT's sources rather than a filing, and should be read that way.

The structure is not in question, though, because the structure is what US policy currently allows. Tencent cannot buy these chips. It can rent them.

Buying is controlled, access is not

Export controls restrict the sale and shipment of advanced accelerators to Chinese companies. They do not restrict a Chinese company from logging into a data centre somewhere else and using chips that never move. Nvidia put the position plainly when an earlier version of this arrangement surfaced:

By design, the export rules allow clouds to be built and operated outside controlled countries by approved firms.

That is an accurate description of the rules, and it is the whole story. The control is on the location of the metal, not on who computes with it.

Tencent has used this route before. In December 2025 the FT reported that it was renting Nvidia Blackwell B200 capacity through Datasection Inc., a Japanese firm operating data centres near Osaka, with most of an initial batch of 15,000 chips contracted to Tencent for three years. Tencent said at the time that its use of cloud computing services was "transparent and legal." The reported Oracle deal is the same mechanism at a different order of magnitude.

Fifteen thousand chips for three years is a workaround. A hundred thousand for five years, with $2 billion or so paid before the first token is trained, is procurement. It is the kind of commitment a company makes when it has stopped treating offshore rental as a stopgap and started treating it as its compute strategy.

Why Tencent is buying this much time

Tencent is behind at home and knows it. Its Hy3 model — 295 billion parameters with 21 billion active, Apache 2.0, 256K context — is a credible open-weights release, but DeepSeek and Alibaba have set the pace in Chinese AI for two years. On September 21 Tencent released HY Image 3.5 Preview and claimed results on par with ByteDance's Seedream 5.0 Pro and slightly ahead of Google's Nano Banana Pro and Alibaba's Qwen-Image-3.0 Pro. Those comparisons came from Tencent's own panel of in-house designers, not an independent benchmark, and should be treated as vendor numbers. The market took them anyway: the shares rose more than 7% in Hong Kong.

The FT also notes what this costs. Heavy AI spending has been pressing on Tencent's margins, and a leasing contract converts what would be capital expenditure into a multi-year obligation with a large cash payment at the front. Tencent is buying compute it is not allowed to own, at a premium, because the alternative is falling further behind.

Washington has noticed, and has not acted

This gap is not an oversight that nobody has spotted. The House passed the Remote Access Security Act (H.R. 2683) on January 12, 2026 by 369 votes to 22. It would give the Bureau of Industry and Security explicit authority to license and penalise remote access to controlled technology, closing exactly this route. It has not cleared the Senate, and until it does, remote access remains legal.

BIS enforcement has meanwhile been reviewing how Chinese AI firms reach offshore chips, with reported cases running through Thailand, Malaysia and Singapore — including Alibaba accessing Nvidia hardware in Malaysia via the Singaporean firm Megaspeed, through a structure that traces back to an entity Alibaba owns, and a White House official accusing Moonshot AI of using an unnamed Thai intermediary. Reviewing is not the same as prohibiting. The legal position has not changed.

What the deal actually settles

Two things follow, and neither is comfortable for the policy as written.

The first is that frontier compute now has a price for Chinese labs rather than a prohibition. $7 billion over five years is expensive, but it is payable, and it is payable by exactly the handful of Chinese companies — Tencent, Alibaba, ByteDance — that can already afford it. The control has not stopped the largest players. It has raised their cost base and eliminated everyone smaller.

The second is that every deal like this builds a constituency against closing the loophole. Oracle has now reportedly booked years of Asian capacity against a Chinese customer, with cash in hand. If the Senate passes H.R. 2683, somebody has to unwind that. The longer the gap stays open, the more American revenue depends on it staying open — which is a strange outcome for a policy whose stated purpose was to deny Chinese labs the compute in the first place.

Sources: Reuters via Investing.com — Tencent leases 100,000 chips from Oracle for $7bn, FT reports · Reuters — China's Tencent leases 100,000 chips from Oracle to accelerate AI push, FT reports · Benzinga — Tencent taps Nvidia Blackwell AI chips through Japan cloud deal · eeNews Europe — US House moves to extend AI chip export controls to the cloud · TNW — US reviews China's remote access to Nvidia chips · Bloomberg via Yahoo Finance — Tencent releases AI image model to catch ByteDance, Alibaba · Tencent — Tencent Hunyuan officially releases Hy3 · Tencent-Hunyuan/Hy3 on GitHub