China's AI Chip Industry Splits: Markets Read New US Restrictions as Favoring Domestic Semiconductor Suppliers
Chip suppliers rose while Alibaba and Baidu fell, but those moves have more than one cause. Alibaba's December quarter capital spending against operating cash flow, assessed in February 2027, will test the buyers' financing burden.
On 17 November 2026, SMIC rose 6.81% on its Hong Kong listing and Hua Hong gained 5.95% in Shanghai. Cambricon, the Shanghai listed AI chip designer, fell 0.22%, complicating any story of a uniform domestic chip rally. At their Nasdaq closes, Alibaba's ADR lost 4.38% and Baidu's ADR fell 3.77%. Baidu's own results and a global retreat from risk also moved the stocks. Markets nonetheless read the new U.S. restrictions as favoring domestic semiconductor suppliers; Washington's rule explains only part of the divergence.
We would pass on a blanket bet on domestic chip shares and favor companies delivering complete systems with memory. Memory Lag is the interval between an order for domestic AI chips and delivery of a functioning system equipped with high bandwidth memory. Chip demand can change in a trading session, while the buyer still has to procure and finance equipment capable of generating cloud revenue. The difference matters because China's leading cloud customers are also investors in local processing capacity. Their balance sheets carry the cost before any advantage from that capacity reaches their customers.
The U.S. Bureau of Industry and Security's rule, announced on 16 November 2026, requires licences for specified chip exports serving overseas data centers offering more than 12,000,000 aggregate TPP to companies headquartered in Country Group D:5 or Macau. Advanced chip exports to Indonesia, Malaysia, Singapore, Thailand and Brazil also require licences, with an exemption for U.S. headquartered companies. Existing providers have 90 days after the rule takes effect to obtain licences, and commercial recommender systems are exempt. Publication is scheduled for 19 November 2026. Licences alter the options for buying foreign compute; they add no wafers or memory to China's factories. The rally prices an expected shift in orders rather than a government purchase.
The market had reasons to worry about cloud buyers before that announcement. Alibaba's ADR fell 17.77% from $105.38 on 9 November 2026 to $86.65 on 17 November, already reaching $94.36 by 12 November. Baidu released results on 16 November showing online marketing revenue down 22% year on year. Its advertising weakness and the wider market retreat accompanied the reaction to Washington's rule. Treating every percentage point of the subsequent decline as an export control verdict would turn a complicated trading day into an implausibly tidy investment thesis.
SMIC's third quarter figures help explain why the factory attracted buyers. Revenue was $3,190.0 million against $2,381.8 million a year earlier, gross margin reached 29.9%, and utilization was 97.3%. Monthly capacity rose 1.85%, from 1,096,500 in the second quarter to 1,116,750 standard logic eight inch equivalent wafers in the third. Those wafers cover all nodes, so full factories alone say little about advanced AI system deliveries. SMIC guided fourth quarter revenue growth of 4% to 6% sequentially, yet a gross margin of 26% to 28%. More fabrication volume still comes with price and margin discipline.
The most attractive argument for the suppliers comes from TrendForce, whose October account says "domestic solutions expected to capture nearly 90% of the country's high-end AI chip market this year". It paraphrases Huawei's Eric Xu as saying Ascend had likely surpassed Nvidia in China and Huawei was selling everything it could produce. Huawei's 950DT SuperPoD is expected in volume around the end of 2026 or early 2027. If buyers already want every available chip, suppliers can fill new lines quickly and reserve future orders. Winning nearly all of a constrained domestic market offers substantial revenue and bargaining power.
China has done something similar before. SemiAnalysis counted 507,000 Huawei Ascend units shipped in 2024 and 805,000 in 2025, a 58.8% increase. SMIC shipped 3,033,683 wafers in the third quarter of 2026, up 21.4% from 2,499,465 a year earlier across its product mix. Earlier controls coincided with growing local deliveries, and customers already had evidence that domestic suppliers could expand. But Huawei's jump drew partly on a stockpile of TSMC manufactured dies that SemiAnalysis expected to run out. A bank of chips fabricated earlier supports a rapid ramp without guaranteeing a replacement supply of complete systems.
The 2026 and early 2027 numbers narrow that optimism. The Institute for Progress, a U.S. policy think tank advocating tighter controls, estimates 2026 Chinese AI chip production at about 986,000 H100 equivalents, within a range of 662,000 to 1.4 million. Its 2027 projection is 1.3 million. It also estimates that only 16% of Huawei's 2026 chip output uses domestic high bandwidth memory, rising to 39% in 2027. This percentage measures domestic HBM sourcing, not total memory available for finished systems. SemiAnalysis separately identifies HBM availability overall as Huawei's limiting component, even with adequate SMIC logic capacity. Memory Lag therefore remains a delivery risk on the same timetable as the expected 950DT ramp.
Alibaba supplies the buyer's price tag, though its June quarter 2026 figures predate the November rule. Capital expenditure rose 75% from RMB38,676 million to RMB67,678 million. Against RMB22,945 million of operating cash flow, RMB67,678 million divided by RMB22,945 million is 2.95 times. Free cash outflow was RMB44,670 million against RMB474,505 million in liquid investments; a scenario repeating that quarter's outflow would consume 9.4% of that pool. Alibaba cited "higher pricing of a broad range of chip components" among the spending drivers. Suppliers can recognize demand while buyers still pay for components and wait to earn on installed machines.
The buyer is also a potential winner, which is why a blanket short position against the cloud operators misses something important. Baidu's third quarter results put AI Cloud Infra revenue at RMB8.2 billion, up 94%, against online marketing revenue of RMB11.9 billion, down 22%. It held RMB276.6 billion in cash and investments, and expects its Kunlunxin M100 to be ready for volume production by the end of 2026. The fast growing cloud business gives Baidu an outlet for new compute; the shrinking advertising business increases the importance of what those purchases earn.
Alibaba says its T-Head Zhenwu chips serve more than 650 external customers and cites Omdia's 38.1% share of China's AI cloud market. An operator with dependable access to memory equipped systems could allocate scarce processing power across those customers, giving it negotiating leverage over smaller developers. That benefit would reward procurement skill and patient capital rather than share ownership in a foundry alone. It still depends on turning a large capital bill into cash, the problem that the June figures put in view.
We would prefer suppliers of delivered, memory equipped systems and avoid treating Alibaba as a self financing localization winner while capex exceeds operating cash flow. We would be wrong if Alibaba's December quarter 2026 results, assessed in February 2027, show capex at or below operating cash flow, a ratio of 1.0 or less against June's 2.95; that tests the buyer financing half of our position.