Alibaba and Tencent Turn to Domestic Chips as a US Rule Tests Alibaba's 53% Year-on-Year AI Cloud Growth
Tencent's inference revenue looks safer than Alibaba's year-on-year AI Cloud and Compute Services growth. Domestic chips may preserve existing services before enough capacity arrives to sustain that growth.
A server abroad can remain switched on after 17 February 2027 and still become unusable to a Chinese AI customer. Under a new American rule, certain overseas data centers must secure licenses to continue supplying covered compute. Without one, the machinery works while the commercial arrangement fails. Alibaba and Tencent are turning to domestic chips to escape that dependency. Keeping an existing model running, however, is different from securing the processors needed to build the next one.
We would decline to value Alibaba as though its 53% September 2026 AI Cloud and Compute Services growth will last through 2027. Tencent sounds calmer because much of its cloud business runs inference on relatively few domestic chips. Alibaba sells compute to outside AI labs and must keep adding capacity. Substitution Lag is the period in which domestic processors preserve existing AI services before enough additional high end compute arrives to sustain growth. We expect that lag to put Alibaba's March 2027 AI Cloud and Compute Services growth at 45% or less.
The Bureau of Industry and Security issued the rule on 16 November 2026; it took effect on 19 November. Covered advanced chips shipped to data centers providing companies headquartered in China access to more than 12 million aggregate Total Processing Performance require a license. That is roughly 760 Nvidia H100s, and linked centers count together. Existing non US providers supplying significant compute face a 17 February 2027 licensing deadline, according to Latham and Watkins. Licenses are a condition rather than an automatic shutdown, but the rule makes future usable capacity depend partly on an administrative decision.
Alibaba's September 2026 results put AI Cloud and Compute Services revenue at RMB60,973 million, up 53% from RMB39,914 million a year earlier. Capital expenditure reached RMB50,477 million, an increase of 60%, and free cash flow was an outflow of RMB34,472 million. Alibaba held RMB487,929 million in cash and liquid investments at 30 September 2026. That is 9.67 quarters of expenditure at September's rate (RMB487,929 million divided by RMB50,477 million), before other demands on cash. Alibaba has money to order hardware; a full fabrication line still determines when it arrives.
Chief executive Eddie Wu gave a harder number on the 19 November 2026 earnings call. He targeted 30% to 45% growth in broader cloud revenue next year and said the rule would affect next year's products and services. In a scenario applying the upper 45% rate to September's narrower AI Cloud and Compute Services base, RMB60,973 million times 1.45 gives RMB88,411 million, against RMB93,288 million at 53%, an illustrative quarterly gap of RMB4,877 million. That is our scenario across different revenue definitions, neither Alibaba's quarterly forecast nor revenue attributed to the rule. Even its upper bound assumes a slower rate of expansion.
Tencent offers the strongest near term objection. Chief strategy officer James Mitchell told analysts, according to Fortune, "We actually don't see an impact because our cloud business today is primarily running inference". Tencent's September 2026 results recorded 35% cloud growth as it allocated more compute to model services and GPU rental. Mitchell says inference runs on a surprisingly limited number of domestic chips. Customers buying answers from existing models have less reason to care which processor produces them, provided the service continues.
Yet Tencent has spent heavily to secure that service. September 2026 capital expenditure was RMB81,433 million; net cash of RMB58.2 billion at 30 June became net debt of RMB13.7 billion at 30 September. Negative free cash flow of RMB69.1 billion would have been positive RMB16.9 billion without compute prepayments, implying RMB86.0 billion paid ahead (RMB69.1 billion plus RMB16.9 billion). President Martin Lau said training Tencent's next Hunyuan model would take longer under the restrictions. A company can defend current inference revenue and still find its next training cycle painfully expensive.
Alibaba's best answer is a manufacturing timetable. Wu says current T-Head chips already support large and ultra large model training, while Zhenwu V900 mass production is planned for the first quarter of 2027 and commercial scale clusters for the second. The Institute for Progress projects accessible Chinese compute rising from 4.6 million H100 equivalents at the end of 2026 to 8.8 million at the end of 2027. Even if all 2.8 million projected offshore equivalents disappear, 8.8 million minus 2.8 million leaves 6.0 million, still 1.4 million above the previous year. That is a meaningful case for continued growth despite licensing.
The same 2027 projection limits the reassurance. Offshore access represents 2.8 million over 8.8 million, or 31.8% of projected compute, and domestic stock is already counted in that total. Robyn Mak at Reuters Breakingviews writes that SMIC expects shortages through 2027 and Chinese foundries operate at or near capacity. Wu's own 30% to 45% guidance covers that year. More aggregate compute at year end can coexist with slower annual revenue growth when commercial training clusters arrive after demand has materialized.
The previous restriction points the other way. On 9 April 2025 Nvidia learned that H20 chip exports to China required a US license. Yet Alibaba's September 2025 cloud results showed RMB39,824 million in revenue against RMB29,610 million a year earlier, growth of 34%. That episode makes resilience a real possibility. The H20 restriction covered a chip shipped into China; the November 2026 rule reaches access to chips installed in third countries. The earlier workaround is now part of the exposure, making a repeated growth surprise harder to underwrite.
Tencent needs enough domestic inference capacity to keep earning revenue from existing models; Alibaba needs expanding clusters to sell computing power to customers building new ones. A chip capable of training a large model is commercially different from a cluster delivered, installed and rentable at the moment a customer needs it. The distinction puts a limit on what cash reserves and chip road maps can accomplish. It also explains why a slowdown would first appear in Alibaba's growth guidance rather than as an immediate halt to AI services. Substitution Lag turns chip availability into a timing risk borne by investors paying for future revenue.
We would avoid paying for another year of Alibaba's 53% AI Cloud and Compute Services growth, while treating Tencent's domestic inference capability as protection for current revenue rather than an answer to future model training. We would be wrong if Alibaba's March 2027 quarter results show more than 45% year on year growth in AI Cloud and Compute Services on a consistent reporting basis when we review them on 31 May 2027.