Lead: A collapsed $7 billion U.S. compute deal has exposed a new challenge for Chinese internet giants seeking overseas AI compute.

For Tencent, AI competition is entering a more sobering phase: models can keep training and products can keep iterating, but the high-end compute underpinning all of this may no longer be obtainable as originally planned.

In the past, internet giants solved their compute problems mainly by purchasing chips, building data centers and renting cloud services. As AI large models sent demand for computing resources soaring, overseas compute leasing also became an important supplementary channel.

But that path now faces new variables.

On November 4, according to a report by the UK's Financial Times, Oracle canceled an AI chip leasing agreement with Tencent estimated to be worth $7 billion. Under the original plan, Tencent would have leased about 100,000 advanced AI chips through Oracle's data centers in Southeast Asia over a five-year term.

Had the deal gone through smoothly, it would have become an important supplement for Tencent in securing overseas AI compute. But the plan has now been disrupted.

More noteworthy is that this is not simply a failed commercial transaction. The report linked it to the U.S. government's concerns about advanced chips flowing to China, while also mentioning that China is pushing tech companies to use more domestic chips.

What appears to be an ordinary compute leasing project actually involves three questions: Where else can Chinese internet giants obtain high-end compute? Can overseas cloud service providers still serve as compute suppliers for Chinese companies? And to what extent can domestic AI chips absorb this demand?

That is what makes the $7 billion deal truly worth attention.

01 Why was the $7 billion deal suddenly canceled?

Let's first go back to the end of September.

On September 30, Reuters, citing a previous Financial Times report, disclosed that Tencent had reached a five-year overseas compute leasing agreement with Oracle.

According to information disclosed at the time, the deal was expected to be worth about $7 billion, involving about 100,000 advanced AI chips deployed across multiple Oracle data centers in Southeast Asia.

This was also Tencent's largest overseas compute leasing transaction to date.

It should be noted that the $7 billion was not the amount for a one-time chip purchase, but the estimated value of this five-year leasing agreement. Tencent would obtain access to advanced AI chip resources by leasing computing capacity from overseas data centers.

This arrangement had its practical background.

U.S. restrictions on exports of advanced AI chips to China make it difficult for domestic companies to directly obtain some high-performance processors. At the same time, China also has its own approval and industrial policy considerations regarding imports of some advanced chips.

For internet companies that need to continuously train large models, domestic procurement, overseas leasing and self-built data centers were originally several paths that could be pursued simultaneously.

Especially overseas compute leasing.

It does not require chips to be shipped directly into China, but instead provides computing services through overseas data centers. For companies hoping to obtain higher-performance GPUs while finding it temporarily difficult to meet demand through traditional procurement channels, this is an attractive option.

Tencent clearly saw the value in it.

However, things changed in early November.

The Financial Times reported on November 4 that U.S. government officials had expressed concerns to Oracle about the deal. At the same time, officials in Guangdong had also visited Tencent to ask it to support domestic chips.

Against this backdrop, Oracle ultimately canceled the deal.

Reuters subsequently relayed the report but said it could not independently verify it. Tencent, Oracle and the White House all did not respond to requests for comment.

In other words, as of now, the specific process of this incident still mainly comes from media citations of people familiar with the matter and has not yet been officially confirmed by both parties to the transaction.

But one thing is relatively clear: overseas compute leasing is beginning to face an increasingly complex policy environment.

In the past, cloud service providers mainly considered customer demand, compute utilization and contract returns. Now, who the customer is and for whom the chips ultimately provide computing power may also become important factors in whether a transaction can proceed.

For U.S. cloud service providers like Oracle, a $7 billion commercial contract is certainly attractive, but if it triggers sustained attention from the U.S. government, the compliance costs and policy risks behind the transaction also need to be reassessed.

A transaction that had already entered the commercial cooperation stage ultimately still failed to materialize.

02 Why does Tencent need compute so badly?

Looking only at the amount of this transaction may not be enough to understand its significance to Tencent.

What is truly noteworthy is that Tencent's AI investment is clearly accelerating, and compute has become an unavoidable part of this round of investment.

Since the beginning of this year, Tencent has continuously increased investment in the Hunyuan large model, AI-native products and WeChat AI capabilities.

From the company's disclosed 2026 interim report, in the second quarter, Tencent's FinTech and Business Services revenue reached 60.3 billion yuan, up 9% year on year. Within this, growth in the business services segment was mainly driven by higher cloud service revenue, supported by factors including increased AI-related demand, overseas business expansion and an improved pricing environment.

A more intuitive change comes from capital expenditure.

In the first half of 2026, Tencent's capital expenditure reached 84.72 billion yuan, compared with 46.583 billion yuan in the same period last year, an increase of about 82% year on year.

Within this, second-quarter capital expenditure alone reached 52.784 billion yuan, significantly higher than 19.107 billion yuan in the same period last year.

Of course, capital expenditure is not entirely for AI; it also includes investment in IT infrastructure, computing equipment, data centers and so on. But judging from the company's explanation of cash flow, AI has become an important direction of fund use.

Tencent disclosed in its interim report that second-quarter capital expenditure payments reached 59.3 billion yuan, and free cash flow for the quarter was negative 13.8 billion yuan.

The company also specifically mentioned that operating cash flow included a large amount of AI-related prepayments, mainly used to support Hunyuan model upgrades, inference demand for WorkBuddy and CodeBuddy, WeChat AI projects, and the application of AI capabilities across various products and services, while also meeting demand brought by cloud service business growth.

This shows that Tencent's demand for compute has already landed in specific business scenarios.

Hunyuan needs compute for training and iteration, WorkBuddy and CodeBuddy need compute to support inference, WeChat AI needs compute to handle massive user demand, and Tencent Cloud also needs more computing resources to serve enterprise customers.

AI is no longer just a technology investment by Tencent's R&D department; it is entering multiple business links such as advertising, cloud services, office tools and social products.

This also explains why Tencent wants to seek overseas compute.

In fact, Tencent management had previously publicly discussed the GPU supply issue.

At the fourth-quarter 2025 earnings call in March this year, Tencent management said it would continue to increase compute supply in 2026, with main sources including leased compute, purchasing high-end imported GPUs and increasing procurement of domestic GPUs.

Management also mentioned that if supply chain and regulatory conditions permit, the company would prefer to buy compute equipment directly because leasing requires additional costs; but when supply is constrained and regulatory requirements are complex, leasing is sometimes a necessary way to obtain compute.

This statement is quite interesting.

It actually shows that Tencent has not treated overseas leasing as its only option, but when there is a gap in high-end compute supply, leasing is indeed an important supplement.

Moreover, training and inference have different requirements for chips.

Tencent management specifically explained this at the earnings call: training large models requires obtaining chips with leading performance as much as possible in order to continuously improve model capabilities; inference focuses more on cost, and China already has multiple chip supply options.

This is also why, although domestic AI chips have made considerable progress, it still cannot simply be assumed in the short term that all high-end training compute demand can be fully replaced by domestic chips.

For Tencent, whether the Hunyuan model can continue to iterate is directly related to its position in the next round of AI competition.

If training compute is insufficient, the pace of model iteration may be affected; if inference costs are too high, the commercialization of AI products will also face pressure.

Therefore, the value of this $7 billion deal was not only about increasing access to 100,000 chips, but also about providing Tencent with a relatively independent overseas compute resource.

Now, this resource has temporarily fallen through.

03 Why is the path to overseas compute becoming harder to travel?

Tencent's experience actually reflects the dual constraints Chinese tech companies face in obtaining high-end AI compute.

The first constraint comes from the United States.

U.S. export restrictions on advanced AI chips are no longer simply about limiting chips from directly entering China.

As Chinese companies begin leasing computing resources through overseas data centers, the U.S. government is also paying increasing attention to this remote access model.

In August this year, The Information reported that the U.S. Department of Commerce was studying new AI rules that could further restrict Chinese companies from remotely using advanced AI chips through data centers in places such as Thailand and Singapore.

At the time, the relevant policies were still under study, and whether they would be introduced and how they would be implemented remained unclear.

But this has already sent a signal: U.S. restrictions on advanced AI compute may in the future extend from the physical movement of chips to the remote use of computing power.

This has a particularly direct impact on cloud service providers.

In the past, overseas data centers could generate revenue by renting out compute as long as they could find customers and sign contracts.

Now, the customer's nationality, end use and service targets may all become part of compliance reviews.

The Financial Times previously also reported that Nvidia had asked some cloud service customers to ensure that at least 51% of services are provided to non-Chinese customers. This also shows that overseas compute suppliers are facing stricter customer structure and compliance requirements.

Even if the relevant leasing transactions are not necessarily illegal under existing rules, policy pressure is enough to affect corporate business decisions.

The second constraint comes from within China.

There is a detail worth noting in this report: Guangdong also asked Tencent to support domestic chips.

This is consistent with China's current direction of promoting independent development of the semiconductor industry.

For domestic internet giants, purchasing overseas high-end chips can solve immediate compute needs, but if they rely on imported products for a long time, the development space for the domestic chip industry will also be affected.

Therefore, the orientation of domestic policy is also changing.

On November 5, Reuters reported another noteworthy situation: about 600,000 Nvidia H200 chips are currently stranded in Hong Kong, waiting to enter mainland China.

The report cited people familiar with the matter as saying that although some Chinese internet companies had obtained a small number of H200 chips in recent months, large-scale imports had still not been approved.

The H200 is not Nvidia's latest-generation product, but its performance and software ecosystem are still superior to some current domestic alternatives, especially in the field of large model training.

This matter is not the same event as the Tencent-Oracle deal, but it reflects a common problem: Chinese internet companies still have strong demand for high-end compute, but the channels for obtaining compute are affected by multiple factors.

On one side, the United States restricts advanced chips and remote compute services; on the other side, China promotes domestic chip substitution.

Chinese tech companies need to find a balance between these two policy constraints.

For Tencent, the problem is no longer just whether it can buy chips, but also whether it can obtain sufficient computing resources in a long-term, stable and compliant manner.

04 What other options does Tencent have next?

The cancellation of the $7 billion deal does not mean Tencent's AI compute will come to a standstill.

Judging from management's previous statements, Tencent itself has been adopting multiple paths to increase compute supply.

But different paths can solve different problems.

The first is increasing procurement of domestic chips.

This path has relatively higher certainty. Especially in inference scenarios, domestic chips have already formed multiple supply options. As model inference costs gradually become an important variable in AI commercialization, domestic chips have opportunities to undertake computing tasks in more businesses.

But there is still a gap between training and inference. For Tencent, which hopes to continuously improve foundation model capabilities, how to obtain training chips with leading performance and stable supply remains a problem that needs to be solved.

The second approach is to keep seeking overseas compute resources.

The cancellation of the Oracle deal does not mean that all overseas cloud service providers will exit the Chinese customer market. Tencent can still secure resources through other cloud service providers, data centers in different regions, and diversified supply agreements.

However, the costs and uncertainties of this path may increase.

If the U.S. government further tightens restrictions on remote compute access, overseas suppliers will face higher compliance risks when accepting Chinese customers. Even if new partners are found, they may not be able to secure the same volume of chips and terms as originally planned.

Finally, there is improving compute efficiency.

This may matter more than simply adding more chips.

For Tencent, compute investment ultimately needs to translate into model capabilities, product experience, and commercial revenue.

The Hunyuan model requires continuous iteration, but businesses such as WeChat, advertising, gaming, and cloud services also need to leverage AI to improve efficiency and generate revenue.

If more tasks can be completed under the same compute conditions through model optimization, inference scheduling, and software-hardware coordination, then dependence on external compute may gradually decrease.

However, this takes time and also requires sustained R&D investment.

From a financial perspective, Tencent already has fairly strong cash flow and investment capacity, but whether AI investment can create new revenue growth drivers remains to be seen.

In particular, AI-native products are still in the investment stage. Tencent management has previously acknowledged that there is a time lag between investment and returns for new products.

This means that the compute issue will ultimately feed into the commercialization issue as well.

If compute supply is unstable, model training and product iteration may be affected; if compute costs remain high, AI products will need to find paid use cases faster in order to support continued investment.

For investors, what is worth watching next is not only how many chips Tencent has purchased and how many data centers it has built, but also AI-related capital expenditure, cloud service revenue, the pace of model iteration, and whether AI products can truly deliver commercial returns.

The failure of the $7 billion deal will not change Tencent's direction of continued AI investment in the near term.

But it makes one issue more concrete: in global AI competition, having capital, talent, and users does not mean one can obtain enough high-end compute at any time.

In the past, compute was primarily a funding question. Now, it has also become a supply chain question, a policy question, and a business model question.

Tencent has lost one overseas compute channel. What it truly needs to make up for is a more stable and sustainable compute supply system.