Tencent has signed a five-year lease worth about $7 billion with Oracle to use roughly 100,000 advanced AI chips at data centers in Southeast Asia, the Financial Times (FT) has reported. It is described as the company's largest overseas lease and expands its access to computing resources that are difficult to secure inside China. According to the FT report, the deal requires an upfront payment of about 30% of the contract value, so even a lease of overseas facilities involves a large initial outlay. Reading Tencent's earnings materials alongside U.S. export-control guidance shows that evaluating this deal requires looking beyond the number of chips to the financial burden and the conditions that affect long-term use.
The contract value and chip count are based on the FT's own reporting, not on details confirmed by the two companies. The chips are said to be at multiple Oracle facilities in Southeast Asia, but the chip model, the host countries and the start date for operations have not been disclosed. Nor is this a contract in which Tencent buys chips and brings them into China. What has been reported is an agreement to use computing resources in facilities located outside the country.
A 51.4 billion yuan prepayment burden in the earnings
In its second-quarter 2026 results, released on August 12, Tencent said free cash flow for April–June was negative 13.8 billion yuan. Free cash flow is one measure of how much financial room a company has after business operations and investment, but the company said it would have been 37.6 billion yuan excluding prepayments for computing resources.
Simply comparing the two figures in Tencent's second-quarter 2026 results gives an impact of 51.4 billion yuan from prepayments for computing resources. However, this is not the amount paid under the Oracle contract alone.
| Metric for the same quarter | Amount | What the figure means |
|---|---|---|
| Actual free cash flow | −13.8 billion yuan | The figure reported by the company |
| Value excluding computing-resource prepayments | 37.6 billion yuan | A supplementary figure the company gave, excluding the effect of prepayments |
| Difference | 51.4 billion yuan | The prepayment impact, calculated as 37.6 − (−13.8) |
The source is page 2 of the earnings release. All the figures cover the same Tencent consolidated scope, the same quarter and the same currency. However, the 37.6 billion yuan excluding prepayments does not represent cash that actually remained on hand. The fact that the actual figure was negative but would have been positive without prepayments shows that the timing of payments for computing resources heavily influenced this quarter's cash flow.
As for the Oracle contract itself, multiplying the FT-reported $7 billion total by the roughly 30% prepayment ratio gives a simple estimate of about $2.1 billion upfront. The 51.4 billion yuan is derived from earnings figures reflecting Tencent's overall prepayments for computing resources, while the $2.1 billion is a rough estimate based on the terms reported for a single contract. The scope and currency differ. Nothing in the earnings materials explains the 51.4 billion yuan as a payment to Oracle.
Tencent's capital expenditure in the same quarter was 52.8 billion yuan, up 176% from a year earlier. While capital spending is surging, prepayments to secure computing resources are also weighing on cash flow. But the 52.8 billion yuan is a reported capex figure and a separate item from the 51.4 billion yuan prepayment impact. The two cannot be added together to arrive at the total cost of the Oracle contract.
Cash flow and profit also need to be considered separately. The company's non-IFRS operating profit was 75.6 billion yuan, up 9% from a year earlier. This is the company's own profit measure, calculated by excluding certain one-off items and others. The fact that free cash flow was negative therefore cannot, on its own, be taken to mean that the core business or AI services are loss-making. What the figures show is the heavy burden of committing large sums to secure computing resources before usage growth and monetization have progressed.
Computing power is needed for inference, not just training
Tencent said the uses of the prepayments include not only improving its Hy models but also inference for the business-support service WorkBuddy and the coding-assistance tool CodeBuddy, AI initiatives at Weixin, and cloud use by external customers. Inference refers to running a developed AI model in practice to respond to users' questions and work requests. It requires computing resources in different situations from training, which is used to build new models.
In other words, Tencent's demand for computing resources does not stem solely from competition in model development. Once AI is built into products and services and people actually use it, continuous computing capacity is needed to process the responses. Judging from the uses cited in the earnings, Tencent is securing computing resources for both AI development and service operations.
This, however, is an explanation of Tencent's overall prepayments and computing demand. There is no public plan assigning the roughly 100,000 chips said to be leased from Oracle to training a particular model or to CodeBuddy inference. The chip count alone therefore cannot be used to predict when a new model will be released or how much its performance will improve.
Tencent's management has indicated it intends to expand its procurement of computing resources and turn growing use of applications and AI models into future revenue. That is an outlook for future monetization, however, and separate from a record of investment returns. After securing usage rights through prepayments, how much actual demand the company can convert into paid service usage will be important in assessing the financial burden.
Is the 100,000 in Southeast Asia one giant cluster?
In October 2024, Oracle announced plans to invest more than $6.5 billion in Malaysia and open a public cloud region. Its policy of expanding AI and cloud infrastructure in Southeast Asia had been made clear before this contract was reported.
That plan alone, however, does not identify the facilities Tencent will use. Also, the figure of more than $6.5 billion is Oracle's investment plan amount and is an entirely different number from the $7 billion reported as the Tencent contract value.
The roughly 100,000 chips reported by the FT should likewise be understood as a total across multiple facilities. It has not been reported that all 100,000 will be configured as one giant compute cluster, meaning a facility in which many chips are interconnected to jointly run a single workload. The total number of usable chips is a separate matter from the scale of computing that can be committed together to a single training run or similar task.
Without the chip model and system configuration, its computing power also cannot be directly compared with other companies' GPU clusters. Moreover, because the service content and usage-time terms of the contract have not been disclosed, it would not be appropriate to divide $7 billion by the number of chips and compare the result with typical GPU purchase prices or hourly cloud rates.
What the report shows is only that Tencent is trying to secure a very large amount of computing resources. How much computing it can actually run for that money cannot be judged until the contract details emerge.
Even when rented overseas, regulatory conditions remain
In guidance dated May 13, 2025, the U.S. Commerce Department's Bureau of Industry and Security (BIS) explained that, depending on the purpose, export-control licenses may be required for activities that train AI models for companies headquartered in China and elsewhere. This can apply, for example, where there is knowledge that the model will be used for purposes or by users connected to weapons of mass destruction or military intelligence activities. The guidance also cites as examples the supply of equipment to overseas data centers, changes in the use or users of existing facilities, and support or services provided by U.S. persons.
The guidance does not uniformly prohibit Chinese companies from using overseas cloud services. At the same time, it cannot be said that every transaction falls outside the regulations simply because the chips are not brought into China and are used at overseas data centers. Multiple conditions are involved, including the covered equipment, the purpose, the users and the service provider's awareness. Publicly available information does not make it possible to determine whether the reported Tencent contract requires a license or whether any approval has been obtained.
Legislation to bring remote use itself under export controls is also advancing. On January 12, 2026, the U.S. House of Representatives passed the Remote Access Security Act (H.R.2683). The bill would allow the Commerce Secretary to regulate "remote access," meaning the use of equipment in distant locations through networks or the cloud, if the Secretary determines it poses a serious risk to national security or other interests.
According to the status posted on Congress.gov, the bill was referred to a Senate committee on January 13 and has so far only passed the House. It should not be treated as enacted law.
Existing purpose-based regulations and a bill seeking to extend regulatory authority to remote access need to be considered separately. Even so, in a contract securing usage rights over five years, the possibility that regulations and service-provision conditions will change during that period cannot be ignored.
For Tencent to turn the computing resources it has secured into revenue, it will need to allocate them stably to the AI services that require them, and it will also need to keep using them over a long period while satisfying the applicable regulations and contract terms.