Alibaba and Tencent both laid out a similar capital expenditure trajectory during their respective latest quarterly earnings calls: the peak period for AI infrastructure capex is approaching, with 2028 set to be a key inflection point.
At the same time, both companies emphasized that inference workloads can already run efficiently on domestic chips, and domestic compute power is becoming a major core engine supporting cloud revenue growth.
Alibaba CFO Xu Hong said during the earnings call that capex for the quarter was 50.5 billion yuan (about US$7.5 billion), down from 67.7 billion yuan (about US$10 billion) in the previous quarter. He further explained the future pace:
The peak of this round of sustained investment is concentrated in 2027 and the first half of 2028; after 2028, as the domestic compute supply chain achieves large-scale, sustained expansion, we will have passed through the phase of the highest AI infrastructure investment intensity.
When asked whether positive free cash flow could be achieved earlier, Xu Hong responded:
Assuming the domestic compute supply chain based on T-Head continues to scale from calendar years 2027 and 2028 and progresses according to the established pace, we expect to achieve positive cash flow in the second half of 2028.
Tencent CFO Luo Shuohan disclosed during the earnings call that third-quarter operating capex payments were 63.6 billion yuan, up 257% year on year. Management said in response to analyst questions that prepayments in the second and third quarters were at "abnormally high levels," and that the company has now signed the long-term contracts it wanted to sign, so "prepayment outflows will decrease."
Tencent management further noted:
Looking at the cash capex level, we believe this is at a peak level, and cash capex may fluctuate around this peak level on a quarterly basis for the remainder of 2026 and in 2027, and then we expect it to decline thereafter.
Domestic Compute: From Alternative to Core Engine
Alibaba CEO Wu Yongming spent a great deal of time during the earnings call introducing progress at T-Head. He said that since its launch in August, supernode instances based on Zhenwu M890 have "achieved large-scale commercial use," and announced that the next-generation Zhenwu V900 will begin mass production in the first quarter of 2027, with computing performance three times that of the M890, and that "the Zhenwu V900 supernode cluster will achieve large-scale commercial use in the second quarter of 2027."
On the chip roadmap, Wu Yongming gave a clear timetable:
This year, Zhenwu M890 has been fully rolled out; we plan to expand the supply chain in 2027 and bring V900 into large-scale mass production, and advance J900 in 2028.
He also specifically mentioned that Qwen has already trained its next-generation model on T-Head's domestic chips, and Model Studio has also begun providing model inference services based on domestic chips.
When analysts asked about the impact of export controls, Xu Hong said:
In the overall planning of our long-term goal for 2030, our AI infrastructure has always been based on the supply chain plan for T-Head's domestic chips. Therefore, as capacity ramps up, the proportion of domestic compute in our overall compute will continue to increase, and it will become the main engine supporting cloud revenue growth, while correspondingly reducing dependence on overseas compute.
Asked whether training would be affected, another management executive responded:
If it affects the progress of some subsequent products, I think it would also be a relatively short-term impact, because in the first quarter of 2027, our own domestic chips based on T-Head can fully support ultra-large-scale model training.
Tencent management offered a clear judgment during the earnings call on the relationship between inference and domestic chips:
Our cloud business today mainly runs inference, especially inference for efficient models such as Hy4 Flash and other Flash models. Inference can run very effectively on a surprisingly limited number of domestic chips. Over the past 12 months, there has been some shift in inference productivity. Therefore, we are able to drive a much higher volume of tokens on domestic chips than imagined a year ago. And it is inference that drives cloud revenue.
Regarding the long-term trend, Tencent management judged:
In the long run, we believe all Chinese players will in fact rely on domestic supply chains and domestic chips, and we do see industry progress moving very quickly. With the latest restrictions, we expect it to accelerate further, because all development efforts will have to focus on accelerating the domestic supply chain.
Tencent President Liu Chiping also added a judgment on costs:
Entering 2028, even if you need to buy the same things, you should expect prices to fall significantly.
Cloud Revenue Outlook: Short-Term Pressure, Return to High Growth in 2028
Under the impact of export controls, both companies gave relatively cautious expectations for near-term cloud revenue growth, but both regard 2028 as the point of reacceleration.
On Alibaba's side, Xu Hong expects cloud business growth in the next fiscal year (fiscal 2028) to fall within a "30% to 45% growth range," below the previous market expectation of more than 50%. But he also said, "As domestic compute product scale continues to expand in calendar year 2028, we expect to restore high growth of more than 50% in calendar year 2028." Alibaba reiterated its calendar 2030 guidance: external cloud revenue reaching US$100 billion, with long-term margins exceeding 20%.
In the quarter, Alibaba's revenue from external customers for AI, cloud, and compute services grew 64% year on year, while AI-related product revenue was 17.9 billion yuan, accounting for 49% of external recurring revenue, achieving triple-digit growth for the 13th consecutive quarter, with annualized revenue exceeding US$10 billion.
Tencent's cloud service revenue growth accelerated to 35% year on year in the quarter, higher than 21% in the second quarter. Tencent management said cloud revenue growth was mainly driven by WorkBuddy, model-as-a-service, and GPU leasing, and that "we do not believe our cloud revenue will be materially adversely affected by these changes," because inference workloads can already run efficiently on domestic chips.
Profit Margins: In-House Chips Bring Structural Improvement
Both companies believe that the large-scale application of domestic compute will have a positive impact on cloud business profit margins.
Alibaba management said in response to analyst questions that "since overall these are all our self-developed chips, and combined with the continued scaling of our volume, we also have a relatively good advantage in procurement and supply chain. So compared with using externally purchased chips for cloud and AI businesses, our profit margin will see a relatively positive improvement." In the quarter, Alibaba Cloud's adjusted EBITDA margin expanded to 14.9%.
Alibaba management also pointed out that product differentiation in cloud services in the AI era has fundamentally changed the pricing logic:
"Alibaba Cloud is no longer just a supplier procurement party for customers, but a partner jointly developing products with customers. This kind of product differentiation is difficult to compare directly on price and is a structural factor supporting long-term improvement in cloud profit margins."
On Tencent's side, management said computing efficiency has "actually multiplied year on year," which means that "over time, the same facilities will be able to achieve higher training efficiency and more tokens," thereby continuously improving unit economics.
Comments