Morgan Stanley: China Cloud AI Computing ROIC Reaches 13%-20%, Alibaba Cloud Best Positioned
Nashnova编辑部
Morgan Stanley estimates Chinese cloud providers earn 13%–20% ROIC on AI compute investment — roughly half the 25%–50% range of U.S. peers — but sees significant upside as GPU costs fall and inference demand scales; Alibaba is named the top pick.
How much do China's cloud firms earn on every AI dollar spent?
Morgan Stanley applied its North American ROIC framework to China for the first time, pegging Chinese cloud AI compute ROIC at 13%–20% versus 25%–50% for U.S. peers.
This means → for the same capital deployed into AI compute, Chinese operators earn roughly half the return. The gap is structural, not operational.
The core drag is hardware cost: capex per server in China runs about three times the U.S. equivalent, compressing returns directly.
The bank sees meaningful upside, however, if three drivers materialize: falling GPU prices, surging inference demand, and faster model-iteration efficiency.
Three monetization paths — what does each earn?
Self-built GPU IaaS (own the servers, rent out compute): base-case operating margin of roughly 44%, ROIC of about 13%, cash payback around three years. In plain terms = margins look healthy, but each server costs RMB 8 million upfront — heavy assets weigh on returns.
Leased compute (rent servers from third parties, re-lease to end clients): no upfront capex, operating margin around 20%. This means → lower margin than self-built, but the fastest route to plugging an incremental capacity gap.
MaaS — Model as a Service (sell inference via API on trained models): gross margin 76%, operating margin 53%, ROIC roughly 19%, cash payback about 2.5 years — the highest-margin path of the three.
MaaS is also the most sensitive: if inference workload share runs low, the same capex can produce a loss.
Why does Morgan Stanley favor Alibaba most?
Alibaba is named the top pick because it can run all three paths: the largest-scale AI infrastructure, a mature cloud business, and a proven large language model in Qwen (通义千问).
Morgan Stanley rates Alibaba U.S. ADR Overweight with a $180 target — roughly 45% upside from current levels.
This reflects the bank's view that Alibaba's ROIC visibility across all three paths is higher than peers — even as Tencent's annualized capex has already topped RMB 200 billion, creating competitive pressure.
Why is Alibaba Cloud's current margin so low?
The incremental AI compute operating margin is modeled at 44%, yet Alibaba Cloud's blended margin today sits at just 11%–12% — a wide gap.
In plain terms = the new AI business makes money, but legacy low-margin cloud services, early-infrastructure depreciation, and unallocated R&D and headcount costs all drag the blended number down.
MaaS is the key lever to close the gap: Alibaba targets year-end MaaS annualized recurring revenue (ARR) above RMB 30 billion. This means → if AI revenue approaches 50% of cloud revenue while inference share and token throughput keep rising, the blended margin can migrate toward the 53% base-case benchmark.
Where does the China-U.S. gap really come from?
In MaaS, China actually leads on token throughput: 4,000 TPS per GPU versus 2,750 in the U.S., partly thanks to smaller models, wider use of Mixture-of-Experts architecture — a design that activates only a fraction of a model's parameters per query to boost efficiency — and higher KV-cache efficiency.
Yet China's MaaS ROIC is only 19.5% versus 46.2% in the U.S. This means → the throughput advantage is eaten up by two factors: lower inference workload share (50% vs. 65%) and fiercer pricing pressure.
Morgan Stanley identifies three paths to narrowing the gap: continued GPU cost declines, faster model-iteration gains, and an accelerating shift of compute allocation from training to inference. Whether these three variables improve on schedule is the key checkpoint for China's AI compute return curve.
Content is for reference only, not financial advice.