Tencent has reportedly leased about 100,000 advanced AI chips from Oracle for roughly $7 billion over five years, in data centers in Southeast Asia. The Financial Times reported the deal, citing people familiar with the matter, and TrendForce summarized the Reuters relay on October 1, 2026. The FT described it as Tencent's largest overseas lease. We found no official confirmation from either company.
The story matters less for the dollar figure than for the mechanism: it shows how Chinese AI labs get access to frontier hardware they cannot buy, and why that route is still open.
The Tencent-Oracle lease at a glance
| Question | Answer |
|---|---|
| Who reported it? | Financial Times, relayed by Reuters; Data Center Dynamics also covered it |
| What is leased? | About 100,000 advanced AI chips |
| Where? | Multiple Oracle data centers in Southeast Asia (countries not named) |
| How long and how much? | Five years, about $7B, roughly 30% upfront |
| Chip model? | Not publicly named; reports do not explicitly say Nvidia |
| What for? | Training Hunyuan models, then inference, possibly resale later |
| Confirmed by companies? | Not that we found |
Why leasing works when buying does not
Chinese tech companies are restricted from directly buying the most advanced AI chips. According to the reports, US rules do not currently prohibit them from leasing computing capacity overseas. A lease puts the hardware in a data center outside China, owned and operated by a US provider, with the customer renting access.
That distinction is the loophole at the center of this story. The chips never cross into China, so the sale-and-shipment controls do not apply in the same way. The rules are about the product leaving the US or going to certain buyers, while the rented compute is a service. Whether regulators close that gap is a policy question we cannot answer from these reports.
This is not the first sign of the route. TrendForce, citing Tom's Hardware and the FT, notes that ByteDance's Singapore-based subsidiary Spring has gained access to 2,304 Nvidia B200 GPUs in a data center run by UK cloud provider Nscale in Norway. Both are rented, offshore, and legal on the reported facts.
The contrast: smuggling versus leasing
Leasing is the lawful option. The illegal one still appears in court: see our coverage of the Super Micro contractor guilty plea over Nvidia AI server smuggling. The two together show demand that export controls have not removed, only redirected. For the other side of the border, China's approval of Nvidia RTX Pro 5500 purchases for ByteDance and Alibaba shows a more mixed picture of what Chinese firms can and cannot get.
Why Tencent needs it: capex and Hunyuan
TrendForce, citing Tencent's second-quarter 2026 results, says capital expenditure rose 176% year over year to RMB 52.8 billion, and free cash flow was negative at RMB 13.8 billion amid large AI-related prepayments. DCD, citing the FT, reports the cash-flow figure as Tencent's first negative result in more than a decade, which its CFO attributed to large AI infrastructure capex and AI-related prepayments.
Executives said the compute is intended for training bigger and better Hunyuan models, then inference, with possible later rental as bare metal or model-as-a-service. Per Investing.com as cited by TrendForce, Tencent's latest Hunyuan models have improved substantially in recent months, narrowing the gap with DeepSeek and Alibaba models. Tencent is also expanding AI across WeChat and other products. For a sense of Tencent's developer-facing AI work, see its open BrowserSkill agent and TeamAI CLI.
What this means for Oracle
If accurate, the deal is a large multi-year customer for Oracle's cloud, with 30% reportedly paid upfront. The reports do not describe Oracle's financing, and we draw no conclusions about it.
What this means for people who build or buy compute
This is a finance-adjacent story, so the practical lines are specific:
- Compute follows the loophole. If you plan around the global GPU market, treat Southeast Asia as a growing supply and demand center. Capacity in the region may tighten as Chinese buyers sign multi-year blocks.
- Price pressure on short-term capacity. A 100,000-chip commitment, if accurate, removes a lot of supply from the spot market. See how a provider recently moved prices in our post on the AWS Capacity Blocks GPU price hike.
- Policy risk is real. Rules that allow leasing today could change. Contracts with a five-year term carry that risk for both the lessee and the data center operator.
- Compliance diligence matters. If you operate GPU capacity, know who your customers are and what the rules say about them, the same issue that sent a contractor to court in the smuggling case.
Questions people are asking
Does this mean Tencent is training on Nvidia's newest chips? The reports do not say. They call the hardware "advanced AI chips" and no one has named a generation. Treat any claim about Blackwell or Rubin parts in this lease as speculation until a company says so. The only named hardware in the related reporting is the 2,304 B200 GPUs ByteDance's Spring unit reportedly reached in Norway, which is a separate arrangement.
Why Southeast Asia? The reports name the region but not the countries. The practical logic is easy to see even without details: it needs power, cooling, network links, and a legal environment where a US provider can operate large GPU clusters and a Chinese customer can pay for them. Our coverage of the power and siting fights around AI data centers shows that finding all four at once is hard anywhere, which is part of why a lease with one established operator is attractive.
Why a five-year term? Training clusters take months to stand up, and operators need committed revenue to justify the build. A long term with a meaningful upfront payment, reported as roughly 30%, lets the operator finance capacity. For the customer it locks in scarce hardware at a time when short-term capacity prices can move sharply.
Is this a one-off? Not on the evidence. The ByteDance arrangement in Norway, also reported by the Financial Times and relayed by Tom's Hardware, points to a pattern of Chinese firms renting offshore capacity from non-Chinese operators. The Tencent deal is described as the largest of its kind for the company, not the first.
What could change it? A rule from US regulators covering remote access to controlled chips would hit this structure directly. So would pressure on host countries, or on Oracle as the operator. None of these has been reported as planned in the sources we reviewed, so we flag them as risks, not forecasts.
A short timeline for context
- Q2 2026: Tencent reports capex of RMB 52.8 billion, up 176% year over year, and negative free cash flow of RMB 13.8 billion.
- October 1, 2026: TrendForce publishes the Reuters-relayed FT report of the lease.
- October 2026: the cases and policies around chip diversion, including the Super Micro contractor plea, continue to surface alongside this lawful route.
The point of the timeline is not that these events cause each other. It is that each is a piece of the same picture: demand for frontier compute that outruns what any single policy can contain, and capital that keeps finding ways to reach it.
Claimed versus verified
Reported, not confirmed by the companies: the $7 billion value, the 30% upfront, the five-year term, the size of 100,000 chips, the Southeast Asia location.
Reported by the company in its own results: Tencent's capex growth and negative free cash flow, per its Q2 2026 results as cited by TrendForce.
Not known: chip generation and mix, which countries host the data centers, exact pricing per GPU-hour, and whether Washington will respond. Data Center Dynamics said it contacted Oracle and Tencent for comment; we did not find a reply.
Green computer chip in front of three server towers, representing Oracle Southeast Asia data centers hosting leased AI chips
Context: the wider China compute picture
China's compute buildout is large and also uneven by chip quality. Our analysis of the SemiAnalysis comparison of China's 24 GW against US 56 GW frames the domestic side of the same race. Leasing abroad is one way to close the gap in the hardware that matters most for frontier training. Nvidia's own roadmap, including the Vera Rubin open designs shown this month, widens what is out of reach for direct purchase.
What to watch next
Official confirmation or denial from Oracle or Tencent; any US Commerce statement on remote access to controlled chips; and whether other Chinese firms sign similar blocks. A policy move to restrict leased access would change the economics of this entire route.
Related reading
-
NTT Data reportedly close to building Anthropic's first Japan data center (up to 500 MW, Osaka)
-
Super Micro contractor guilty plea: Nvidia AI server smuggling
-
China approves Nvidia RTX Pro 5500 for ByteDance and Alibaba
Sources: TrendForce News, October 1, 2026 and Data Center Dynamics, both citing the Financial Times and Reuters.
Reporting is as of October 13, 2026. The deal terms rest on anonymous sources and may change.