In the past, a significant portion of increased revenue at internet companies could be converted into cash. Now, a growing amount of cash is being channeled directly into servers and data centers almost as soon as it's earned.
Just two or three years ago, capital markets worried that tech giants might miss out on the AI boom. Today, with hundreds of billions in capital already transformed into chips, servers, and data centers, investors are calculating a different equation: was this spending justified, and how long will it take to see returns? The latest earnings season is making this calculation increasingly tangible. Rapid growth in capital expenditure is pressuring free cash flow, bringing subsequent questions about depreciation and return on investment to the forefront. This leads us to launch the "AI Ledger for Tech Giants" series, which examines financial reports to understand how much AI is costing these companies and how it's reshaping their core businesses.
In the latest quarterly reports from Alibaba Group Holding Ltd (HKG: 9988) and Tencent Holdings Ltd (HKG: 0700), two numbers stand out. In the second quarter of this year, Alibaba's capital expenditure reached RMB 67.68 billion, a 75% year-on-year increase. Tencent's capital expenditure for the same period hit RMB 52.78 billion, up 176% year-on-year. Combined, these two investments exceed RMB 120 billion, with AI computing power, servers, and data centers being key destinations for the new investment.
This represents a significant departure from the familiar spending patterns of Chinese internet companies. In 2023, Tencent's total annual capital expenditure was only RMB 23.9 billion, accounting for 3.9% of its revenue. Before the large-scale AI push, Alibaba's capital expenditure also remained at low single digits as a percentage of revenue. By the second quarter of this year, capital expenditure for both companies had risen to roughly one-quarter of their quarterly revenue. Not all of this spending can be attributed solely to AI. Neither company separately discloses its AI-specific capital expenditure, but financial reports indicate that AI infrastructure is now a primary area for new investment.
The impact on free cash flow is even more pronounced. Alibaba's quarterly free cash flow was negative RMB 44.67 billion, while Tencent's also turned negative at RMB 13.8 billion. Tencent's upfront cash payments for computing resources were substantial enough to swing its entire quarterly free cash flow from positive to negative. Why have internet companies become such heavy spenders? In just one year, they are spending what previously took years.
Tencent's capital expenditure history over the past few years is illustrative. It was RMB 23.9 billion in 2023, then jumped to RMB 76.8 billion in 2024. The growth pace slowed in 2025, but the second quarter of this year saw a surge to RMB 52.78 billion, even as revenue grew only 11% during the same period. Alibaba's capital is also accelerating into infrastructure. Free cash flow was RMB 156.2 billion in fiscal year 2024, fell to RMB 73.9 billion in fiscal 2025, and turned negative to the tune of RMB 46.6 billion in fiscal 2026. The company has cited cloud infrastructure spending as a primary reason for the decline in free cash flow.
A portion of Tencent's cash outlay is reflected in prepayments for computing power. In the second quarter, its operating cash flow included substantial AI-related prepayments for model upgrades, WeChat AI features, programming products, and external cloud customers. Tencent disclosed that its free cash flow was negative RMB 13.8 billion for the quarter; excluding these computing resource prepayments, it would have been positive RMB 37.6 billion. This single item was enough to change the overall quarterly free cash flow trend. Similar developments are occurring in the United States. Microsoft's cash capital expenditure for fiscal year 2026 reached $115.9 billion, up from $44.5 billion two years prior. Over two years, revenue grew by about 35%, while capital expenditure more than doubled, and free cash flow slightly declined. Alphabet's second-quarter capital expenditure was $44.9 billion, and its free cash flow turned negative. Meta's quarterly capital expenditure was $31.1 billion, leaving only $784 million in free cash flow. Amazon's operating cash flow grew 33% year-on-year over the past twelve months, but capital investment grew even faster, ultimately resulting in no positive free cash flow.
These figures are difficult to compare directly in a single table. Microsoft and Amazon already have vast infrastructure from their cloud computing businesses, so AI is increasing their capital intensity from an already high base. In contrast, Meta, and previously Alibaba and Tencent, better illustrate the shift from an asset-light to an asset-heavy model. Previously, increased revenue at internet companies largely translated into cash. Now, cash is increasingly reinvested into servers and data centers as soon as it's earned. The income statement doesn't immediately reflect these dramatic changes. Machinery purchases aren't fully expensed in the current period, but the cash is spent at the time of purchase. Consequently, we're seeing a scenario uncommon for internet giants: revenue and profits continue to grow, but free cash flow declines first.
AI infrastructure involves much more than just buying chips. Servers must be housed in data centers, requiring network connectivity, storage, power, and cooling systems. Alphabet has disclosed that roughly 60% of its technology infrastructure investment is for servers, with the remainder going to data centers and networking. The International Energy Agency projects that global data center electricity consumption will more than double by 2030. A portion of this computing power is used for model training. Training large models requires concentrated computing power from numerous servers; once training is complete, this intensive phase ends. Using models daily presents a different, ongoing cost. Generating text, using an AI feature in WeChat, or having AI assist with coding – each operation requires server resources. As users and usage frequency increase, so does the computational load. Tencent's large-scale computing resource procurement in the second quarter is already aligned with the actual usage of its Hunyuan model, WeChat AI, programming products, and external customers. Alibaba also anticipates that as AI Agents become more prevalent, computing demand will continue to rise. A transaction on Taobao or a message on WeChat doesn't necessarily require proportional new hardware. With AI, however, a significant increase in usage directly translates to new computing demands. The internet's most profitable aspect was that revenue growth could far outpace fixed asset growth, but AI is narrowing this gap.
Technological advancements are working to reduce these costs. Chips are becoming faster, models are being optimized, and companies like Google, Amazon, and Alibaba are developing their own AI chips, which should lower the cost per use over time. While computing becomes cheaper, tech companies are buying more of it. As costs decline, scenarios previously deemed too expensive begin adopting AI. Search, advertising, office software, coding, and customer service are all integrating AI features. Savings from lower per-use costs are easily offset by increased overall usage. Another financial impact will appear in later years: server purchases consume cash upfront, but their cost is expensed gradually through depreciation on the income statement. Meta's depreciation and amortization increased approximately 46% year-on-year in the second quarter. Alphabet has also cautioned investors that data centers and servers built in recent years will continue to generate depreciation and operating costs. Furthermore, AI hardware becomes outdated quickly. Newer, more powerful, and efficient equipment may arrive before older servers are fully depreciated on the books. For companies investing tens of billions or even over a hundred billion dollars annually, the useful life and replacement cycle of this equipment ultimately impacts return on investment. Currently, free cash flow is absorbing the initial pressure. As depreciation increases, the income statement will continue to absorb the costs of equipment purchased in previous years.
When will the returns materialize? Two or three years ago, the market was quite tolerant of AI capital expenditure. When ChatGPT first emerged, the prevailing sentiment was "better to over-invest than miss out," with the main risk being missing the next technological platform. Microsoft, Google, Meta, and Amazon were all competing for GPUs; slow investment was more likely to worry investors. By 2026, after several rounds of server purchases and with AI revenue starting to appear in financial reports, investors have more data to analyze. Alphabet's cloud business grew rapidly in the second quarter, while the company also raised its full-year capital expenditure plan. During the earnings call, analysts directly questioned the return on capital and payback period for generative AI. Meta's second-quarter revenue grew 28%, but its free cash flow dropped significantly, prompting questions about how these massive investments will translate into profits. Alibaba is already able to show some revenue-side data. In the June quarter, its AI-related product revenue reached approximately RMB 12.4 billion, maintaining triple-digit growth for twelve consecutive quarters. The AI business is generating real revenue, but from a group cash flow perspective, large-scale infrastructure buildout is currently causing a significant cash drain.
For Tencent, this calculation is more complex. The computing power it purchases is not only sold to cloud customers but also integrated into WeChat, advertising, gaming, and various internal products. AI improving advertising efficiency, accelerating game development, or creating new service scenarios within WeChat can all generate economic value, yet these are difficult to consolidate into a single "AI revenue" line item. Looking at future financial reports from such companies, "AI revenue growth" will only answer half the question. The other half involves assessing how much capital was invested to achieve that revenue and efficiency. AWS, Azure, and the entire cloud computing industry have demonstrated that asset-heavy businesses can be highly profitable. However, when capital expenditure rises from a few percentage points of revenue to 20% or even 30%, the evaluation criteria for a company inevitably change. A 20% revenue growth is impressive, but if achieving it requires doubling capital investment, the equation needs re-evaluation. Today's investment intensity may not be the permanent norm. Data center construction has its peaks, server purchases have their cycles, and chips and models will continue to become more affordable. Tencent's capital expenditure surge in 2024, stabilization in 2025, and another jump in the second quarter of this year indicates this curve won't continue rising at the same pace indefinitely. Once the initial wave of data centers, power, and network infrastructure is built, some costs won't need to be repeated annually. However, there are currently few signs of returning to the capital expenditure levels of a few years ago. AI is integrating into more products, inference workloads consume computing power daily, older servers require replacement, and previously purchased equipment continues to depreciate. The eventual stabilization point is currently unknown. In the past, evaluating internet companies primarily involved looking at revenue growth, profit margins, and free cash flow. Going forward, another key metric must be considered: the annual investment required in servers and data centers to sustain that growth. The latest quarterly results from Alibaba and Tencent have put this exact ledger on the table for all to see.
Comments