Alibaba has turned 380 billion yuan of its spending into assets.

Last night, I listened to Alibaba's earnings call, and by the end, only one sentence was left in my mind. Our total investment of 380 billion yuan can be recouped in 3 years. To be honest, I was stunned for a moment after hearing that. Why?

01

380 billion yuan, 3 years, payback period — these three terms put together, it is rare in the past two years to see a player dare to speak to the market in such a tone.

Don't believe it, just take a look around.

Cloud vendors all over the world are pouring money into building computing power, and the capital poured in is like water thrown into a river, with the profit statement eroded by depreciation by a large margin. There is no doubt about that, right?

Alibaba's capital expenditure this quarter reached 67.6 billion yuan, up 75% year on year, which would surprise anyone. Yet the management worked out the accounts directly in front of analysts during the earnings call.

How did they calculate it? Let's start with the demand side.

There is a figure in the financial report: the quarterly revenue of AI-related products reached 12.376 billion yuan, maintaining triple-digit growth for 12 consecutive quarters. 12 quarters are exactly three years, and in each of these three years, the revenue more than doubled compared with the same period of the previous year.

Three years ago, this figure was only a tiny fraction of what it is now, and now it exceeds 10 billion yuan per quarter.

At this annualized growth rate, the annual revenue hits 49.5 billion yuan, equivalent to about 7.3 billion US dollars. On a daily basis, only the AI business of Alibaba Cloud generates 136 million yuan in revenue per day.

This scale is already an independent profitable business, and many listed companies have lower annual revenue than this single segment.

This only counts AI-related products, whose proportion in the external commercial revenue of Alibaba Cloud has risen from 30% to 35%. In other words, for every 100 yuan of cloud revenue received now, 35 yuan is earned from AI business.

Where does the demand come from? The management mentioned several points during the earnings call, which I will sort out for you.

First, the explosion of AI Agents.

Now enterprises are embedding intelligent agents into their own systems. When one Agent runs tasks, it consumes a large amount of Tokens, queues up for GPU resources, and drives up the demand for CPU, storage and network at the same time. The demand is not a single rising river, but several rivers swelling simultaneously.

Second, the model service business is booming.

On Alibaba's Bailian platform, the annualized MaaS revenue has exceeded 16 billion yuan in August, which translates to 44 million yuan per day.

Enterprises purchasing model services is completely different from purchasing servers in the past. In the past, enterprises bought a server and used it slowly on their own, but now they pay by Tokens, consume as much as they use, and there is no upper limit on the consumption volume.

Here is a more direct example to help you understand.

In the past, the largest line item in the budget sheet of enterprise CIOs was purchasing servers and storage. That is no longer the case now.

For the same enterprise, they used to buy storage from Alibaba Cloud last year, but now they buy Tokens. The management mentioned in the earnings call that the procurement needs of customers have changed, expanding from cloud resources to computing power, model services, development platforms, and even Agent deployment services.

Third, C-end applications have covered hundreds of millions of users.

Since the launch of Tongyi App, 250 million users have experienced AI shopping through its intelligent agent function. They can ask questions, compare prices and place orders on Taobao with the whole process assisted by AI. The synergy between AI and e-commerce has taken effect, and this is only the performance on the C-end.

On the B-side, Alibaba's ATH, Tongyi Office, Qoder and a series of other AI applications have a solid user base.

According to data from research institution Synergy, the global cloud infrastructure market reached 143 billion US dollars per quarter, equivalent to 1 trillion yuan, up 43% year on year, the fastest growth in eight years.

Amazon's AWS just released its Q2 earnings report, showing a 37% year-on-year growth, the fastest in 18 quarters. CEO Jassy said that AI demand is so strong that the production capacity cannot meet sales, and the supply gap will last until 2027. Google Cloud's revenue grew by 82% in the last quarter.

Why is this happening all over the world?

Because AI workloads are continuously migrating from local deployment to the cloud, which is not a temporary trend. The situation is even more prominent in the Chinese market: Alibaba Cloud accounts for 38.1% of the domestic AI cloud market, ranking first. You can see that demand is outstripping supply all over the world.

Therefore, Alibaba dares to make this huge investment, laying the foundation with 380 billion yuan. Three years of development, hundreds of millions of users, and tens of billions of orders have all been verified repeatedly by the company. It will not start large-scale construction until the passenger flow is confirmed.

The passenger flow is still growing. The management said that the annualized revenue of AI products in the next quarter will approach 10 billion US dollars, equivalent to nearly 70 billion yuan.

02

The question is, there are so many players, and the demand is so huge that everyone wants to seize the opportunity. Why Alibaba? AWS has an annual budget of 220 billion US dollars, Google Cloud is growing faster than anyone else, how can Alibaba's relatively small investment achieve payback earlier than others?

I got the answer after listening to the earnings call: Alibaba gets more value from the same amount of money.

If you don't believe it, look at the revenue side first.

The AI cloud and computing power service segment earned 5.628 billion yuan in a single quarter, with a profit margin of 12%, 133% more than the same period last year.

During the earnings call, the management stated directly:

The profit margin of AI-related products is significantly higher than the average level of cloud products. Selling storage and bandwidth in the cloud only earns thin profits, selling computing power is more profitable than selling storage, and selling model services is more profitable than selling computing power. The higher up the value chain you go, the thicker the profit margin becomes.

An analogy can make this easier to understand:

Selling computing power is equivalent to selling land, which is priced by plot with average profit. Selling model services is equivalent to selling fully-furnished properties, and the profit from decoration is much higher than the profit from the land itself.

Alibaba is not only selling land, but also selling fully-furnished properties, and the decoration business is growing larger and larger.

There is a set of data here:

AI-related products account for 35% of external cloud revenue, and the improvement of their profit margin has raised the overall profit margin of the entire cloud segment by 4.4 percentage points. Relying only on the AI segment, Alibaba has greatly increased the overall profitability of its entire cloud business.

Then look at the expenditure side. Everyone spends the same yuan to buy computing power, but Alibaba gets it at a different cost.

What is the difference? Alibaba has its own chip factory.

T-Head independently develops self-developed chips covering GPU, CPU, storage and network. More than 500,000 chips of the previous generation have been manufactured already.

The latest Zhenwu M890 chip has been deployed on a large scale on Alibaba Cloud's super nodes in August, serving more than 20 industries and over 650 external customers, covering autonomous driving, internet and financial services sectors.

The original words of the management:

Large-scale deployment of self-developed chips avoids the premium of purchasing high-end commercial chips from external suppliers.

In plain terms, when other enterprises build their buildings, their building materials are imported, and most of the profits from bricks and cement are earned by the material suppliers. Alibaba has its own brick factory, it burns bricks by itself, so it has full control over its own cost.

Moreover, among the domestic enterprises that can mass-produce chips and deploy them on super nodes, there are only two, and Alibaba is one of them.

These 650 customers use its chips for model training and inference. The more self-developed chips they use, the fewer external chips they need to purchase, and every penny saved becomes the gross profit of the cloud business.

There is another point that is easily ignored: speed.

Alibaba Cloud has compressed the delivery cycle of large AI data centers to 100 days, which the management says is the leading level in the world. This year, the production efficiency of self-developed modular data centers will be doubled again.

What does 100 days mean? It is a little more than one quarter. When other players are still laying the foundation, Alibaba's buildings have already been put into operation for lease.

By the way, many people say that Tongyi is still burning money, losing 13.8 billion yuan per quarter, how to explain that? I think you need to see what the money is used for.

It needs to be clarified that MaaS and Tongyi App are two completely different businesses.

MaaS is the B-end model service, sold by Tokens on the cloud, which is profitable. Tongyi App is a C-end application that is still investing heavily to acquire users. The two belong to different business divisions of the cloud and application sectors, and should not be confused together.

Moreover, you need to see what Tongyi has achieved through its investment.

Tongyi runs on Alibaba Cloud. Every time a user uses it, the consumed computing power and Tokens are counted into Alibaba Cloud's revenue. The money spent by the decoration team finally flows back to the account of the construction team, which is a very cost-effective deal no matter how you calculate it.

Therefore, this is the full-stack layout:

Chips, data centers, cloud, models, applications, layer upon layer connected, the capital circulates in a closed loop. High revenue, low cost, fast delivery, these three advantages stacked together form the profit model that other players cannot replicate.

In short, every cycle reduces the overall cost and shortens the payback period, the cycle runs faster and faster. Everyone can say this, but not everyone can achieve it.

03

Yes, not everyone can achieve it, but Alibaba can. How does Alibaba calculate its accounts?

The management has calculated this account during the earnings call.

In February 2025, Alibaba announced a three-year investment plan of 380 billion yuan for cloud and AI infrastructure. By the end of June this year, 190 billion yuan has been invested, and the progress is in line with the plan without any deviation.

When the plan was announced last February, the market had doubts about it. One and a half years later, the investment progress is 100% on schedule.

What is the concept of 380 billion yuan? Spread over three years, it is 350 million yuan per day. For three years, every day when you open your eyes, you have to spend 350 million yuan first, and then start calculating how much you earn that day.

Where did the 190 billion yuan go?

The management gave three reasons: hardware procurement has a natural cycle, the rise of AI Agents drives the demand for CPU, and chip components are still rising in price. The money is not spent randomly, every sum has a clear destination.

There is a simple arithmetic formula:

Payback period = total investment / annual net profit. The total investment of 380 billion yuan is fixed, and every 1% increase in revenue will shorten the payback period by 1%. Therefore, the management said that the payback cycle can be shortened to two and a half years, or even close to two years.

Why dare they make such a calculation? Because this "building" has been collecting rent for eight years already.

The V100 chips purchased in 2018 and A100 chips purchased in 2020 are still running at nearly full load today. Back in 2018, before the emergence of large models, many companies that bought this batch of machines at that time thought the money was wasted.

Alibaba kept these machines and has been using them until now. According to the accounting books, these machines have already been fully depreciated and should have been eliminated, but they are still generating revenue.

They are still making profits even after full depreciation. There is a trick here: the cost of the machine needs to be amortized into the cost year by year. On the day the amortization is completed, its book value returns to zero.

After the book value returns to zero, every dollar it earns is pure profit, no more cost needs to be repaid. The management dares to say that long-term positive cash flow will be achieved after payback, and the V100's 8 years of operation has demonstrated this in advance.

For eight years, the AI industry has experienced two boom cycles. Some enterprises have replaced their machines twice, but Alibaba's old machines are still running. The book records depreciation, but in reality, they keep generating revenue like a building that can collect rent continuously after completion.

The management also reminded:

Do not multiply the single-quarter number by 4 to get the full-year estimate. The high expenditure this quarter is only due to concentrated delivery of goods, and Alibaba will not spend at this level every quarter.

What happens after the payback period?

These data centers can continue to generate revenue for many years after the payback. According to their model, when the growth rate drops below 33%, the cash flow will turn positive. But Alibaba does not choose to slow down, it chooses to seize the market first.

In other words, even if the AI growth rate drops below 30% one day, this investment will only generate profits more slowly, and will not collapse.

This choice is very characteristic of Alibaba: the accounts are calculated clearly, and the pace of investment never stops. The original words of the management are that the AI industry is still in a very early stage. If you do not seize the opportunity now, there will be no opportunity left for you later.

Moreover, do not think Alibaba only spends money without generating cash. This quarter, Alibaba recovered 22.9 billion yuan in cash from operating activities alone, 11% more than the same period last year. While investing heavily, its self-sustaining cash generation capability has never stopped.

I also checked the Morgan Stanley comment after the earnings release, which also did the same calculation:

The current profit margin of Alibaba Cloud still includes many traditional cloud businesses with low profit. When the proportion of AI and model services continues to increase, the overall profit margin still has much room to rise.

You can see that the profit margin is still rising, so the payback period can be further shortened.

In addition, not all the construction cost of these data centers is paid by Alibaba itself. The management said that they will bring partners to jointly build computing power centers, where partners contribute capital and Alibaba contributes technology and operation, and the revenue is shared, so Alibaba's own capital expenditure can be further reduced.

This logic is also shared by international leading manufacturers.

Amazon CEO Jassy once said that a data center can collect rent for 30 years, and servers can achieve payback within three years. AWS has an annual budget of 220 billion US dollars, Microsoft and Google also have huge investment budgets, everyone is building their own data centers, and everyone is waiting for payback.

The difference is that some players keep changing their statements during construction, while Alibaba has calculated the exact payback date clearly in advance.

There is also a consensus in the industry that the demand for AI computing power will remain in short supply before 2030, which is the judgment of the entire industry. As long as the demand does not decline and the revenue does not drop, the payback days will only become shorter and shorter.

Alibaba's uniqueness is that it has worked out the full schedule in advance. Personally, I think once a schedule is announced, it should be delivered as promised.

But in front of the AI industry, everything is full of variables. The 3-year payback period is a calculated result, and it may even be achieved in less than 3 years.