Author ｜ Zhang Zhi

Editor ｜ Li Chentong

In early October, Tencent inked a monumental five-year, $7 billion computing power leasing deal with Oracle, securing access to approximately 100,000 cutting-edge AI chips stationed across Oracle's Southeast Asian data centers.

This agreement stands as Tencent's largest-ever overseas computing power lease, with a significant 30% upfront payment already reflected in its Q2 financial results. Notably, free cash flow dipped to -13.8 billion yuan, while capital expenditures soared to 52.78 billion yuan, marking a 176% year-on-year increase.

Due to export controls, these chips remain stationed in overseas data centers, with enterprises accessing them remotely through cloud services.

The computing power is primarily earmarked for training and inference tasks related to Tencent's Hunyuan large model. Tencent management also mentioned that if internal usage falls short of expectations, the computing power could be recouped at cost and subsequently leased externally through Tencent Cloud.

I. Why Opt for Leasing Over Purchasing?

This decision hinges on economic considerations.

Industry estimates suggest that operating a self-built data center at full capacity for three years incurs costs ranging from approximately $0.55 to $1.10 per GPU-hour.

In contrast, mainstream cloud platforms charge between $2.80 and $5.19 per hour for on-demand usage, offering flexibility at a premium.

One-year reserved instances, while relatively cheaper at $2.45 to $2.75 per hour, still cost two to four times more than self-built options.

Self-built costs represent fixed expenditures, with the industry consensus placing the breakeven point at 55-60% utilization. Above this threshold, self-built facilities can recoup costs within 11-14 months; below 40% utilization, leasing proves more economical.

Given the volatile nature of AI demand and the evolving architectures of AI models, self-built data centers may not operate at full capacity throughout the year.

Tencent's $7 billion investment, spread over five years and encompassing 100,000 chips, amounts to $14,000 per chip annually or roughly $1.60 per GPU-hour—a cost that is competitive when compared to alternatives.

Tencent initiated the deployment of its Hunyuan large model in 2023, with capital expenditures reaching 23.9 billion yuan.

Full-year 2024 expenditures climbed to approximately 77 billion yuan, further escalating to 76.76 billion yuan in 2025 (a 221% year-on-year increase), though this still accounts for just 10% of revenue—below institutional estimates ranging from 86.4 to 108 billion yuan.

Martin Lau clarified that supply constraints for advanced GPUs, rather than budgetary limitations, were the driving force behind this approach.

Alibaba, in early 2025, announced plans to invest 380 billion yuan over three years in cloud and AI infrastructure, with 190 billion yuan already completed by June.

In Q2 2026, Alibaba's capital expenditures reached 67.678 billion yuan (a 75% year-on-year increase), while free cash flow outflow hit 44.67 billion yuan for the second consecutive quarter.

In August, Alibaba raised 80 billion HKD through new share placements to fund full-stack AI capabilities and infrastructure.

ByteDance, as early as 2020, was compelled to store U.S. TikTok user data in Oracle facilities, relocating European data to Norway and Ireland.

Reports indicate that ByteDance deployed around 36,000 advanced AI chips in Malaysia in March 2024, with hardware costs exceeding $2.5 billion; its approved Thai data center project saw investments surge to $25 billion in May.

Additionally, ByteDance constructed six large domestic AI computing centers in 2024, with total investments reaching 35.7 billion yuan.

Market estimates project ByteDance's 2026 capital expenditures at 160 billion yuan (a 40% year-on-year increase), with nearly the entire sum allocated to AI computing power.

While domestic chips continue to advance, significant gaps remain.

With Hunyuan at a critical juncture to catch up with industry leaders, any delay in investment would only escalate future costs.

Financial data reveals that Tencent spent 84.72 billion yuan in H1 2024 (an 82% year-on-year increase), including 52.78 billion yuan in Q2 (a 176% year-on-year increase). As of Q2, cash and time deposits exceeded 510 billion yuan.

Institutions estimate Tencent's 2026 full-year capital expenditures to range between 95 and 110 billion yuan—a conservative projection given the progress made in H1.

II. Tencent's High-Stakes Gamble

Since 2026, U.S. export controls have tightened further, shifting focus from chip destinations to actual users.

In late May, the U.S. Commerce Department issued enforcement guidelines stating that controlled advanced computing products delivered to third-country entities may still require licenses if final control rests with specific national enterprises. Subsequent rules barred firms with Chinese ultimate controllers—regardless of registration location—from leasing advanced training chips.

Legislatively, the U.S. House passed the Remote Access Security Act in January 2024 by a vote of 369-22, proposing to include remote access in export controls. In June, lawmakers introduced the Cloud Security Act, mandating cloud providers to verify user identities.

July's new rules imposed fresh authorization frameworks on major cloud providers, including Amazon, Google, Microsoft, Meta, OpenAI, and Oracle.

The U.S. now channels computing power flows through reviewable, trackable platforms, with its control framework resembling a three-tier funnel:

Tier 1: The A:5 Alliance (NATO members, Japan, South Korea, EU). On July 10, 2024, the UAE became the first Arab nation to join, granting approved entities license-free access to advanced chips.

Tier 2: Includes Malaysia, Singapore, Thailand, India, and Saudi Arabia, requiring case-by-case approvals and end-use monitoring.

Tier 3: Comprehensive restricted list.

For Chinese firms, A:5 countries—core U.S. allies—will not open data centers for remote Chinese access. Even Gulf states are pursuing "de-China-fication" in chip procurement. The UAE's approved G42 and Core42 firms face automatic license expiration after 270 days unless they become U.S. entities.

While many countries remain nominally unrestricted, practical options for Chinese entities are severely limited.

China also enforces stringent data exit safeguards. Domestic user data transfers necessitate security assessments and cannot be freely trained overseas. Data stored on U.S. infrastructure falls under the jurisdiction of the U.S. CLOUD Act.

Enterprises, therefore, adopt a separation approach: domestic operations and user data remain in onshore facilities, while specific workloads like model training occur offshore.

The U.S. Department of Justice arrested a California tech executive accused of smuggling $300 million in controlled AI servers through Malaysia and Singapore from 2023 to August 2024.

Compliant computing power has become increasingly scarce and expensive.

The freedoms witnessed in 2024 began to narrow in 2025 and tightened further in 2026. Tencent's $7 billion commitment came just before these closures.

Further policy tightening or contract disputes could render Tencent's investment worthless.

III. Tencent Bails Out Oracle

S&P downgraded Oracle to BBB- in July 2024, with shares plummeting to $27 earlier this year.

Though Oracle appears to benefit from AI demand, its 2026 fiscal year saw cloud infrastructure revenue grow by 77%, while remaining performance obligations (RPOs) surged by 363% to $638 billion—nearly half tied to a five-year, $300 billion OpenAI contract, revealing a heavy client concentration.

Annual capital expenditures jumped from $21.2 billion to $55.7 billion, with free cash flow turning negative at -$23.7 billion. Debt reached approximately $129.5 billion.

Oracle faces $260 billion in 15-19-year data center lease commitments, while client contracts typically last five years—a classic maturity mismatch.

Tencent's reliance on this highly leveraged supplier carries inherent risks.

Conclusion

Once China's most profitable asset-light company, Tencent now navigates the capital-intensive AI era, where computing power has transformed into a form of "digital rent."

This data center competition will ultimately manifest in everyday technology—whether WeChat's AI assistant can book tonight's restaurant may hinge on those 100,000 chips in Southeast Asia.

The entry fee for AI is clear: $7 billion over five years.

-end-