Morgan Stanley pegs China cloud AI returns at 13%-20% ROIC

Morgan Stanley's new framework quantifies China's AI compute returns at 13%-20% ROIC, with Alibaba best positioned to capture upside across all three monetization paths.

Chinese cloud vendors can earn 13% to 20% return on invested capital from AI compute infrastructure, roughly half the 25% to 50% range of US hyperscalers, Morgan Stanley estimates, with hardware costs the primary drag.

The report, which applies Morgan Stanley's North America ROIC framework to China's cloud market, identifies hardware costs as the core constraint: Chinese servers cost roughly three times their US equivalents, pushing depreciation expense higher and compressing returns.

In the self-built GPU IaaS scenario, each eight-GPU AI server carries 8 million yuan in capital expenditure and rents for 250,000 yuan monthly, yielding a 44% operating margin, 13% ROIC, and a 3.1-year payback period. US peers achieve 31% ROIC with a 2.2-year payback. The report also models a rental arbitrage path — leasing from neocloud providers at 200,000 yuan monthly and reselling at 250,000 yuan — which delivers a 20% operating margin with zero upfront capital.

The report assigns Alibaba a Buy rating with a $180 target, implying roughly 45% upside, citing the company's largest AI infrastructure footprint, mature cloud business, and proven Qwen model capabilities. Alibaba's cloud operating margin sits at 11% to 12%, but the report sees material expansion potential if the company hits its year-end target of 300 billion yuan in annualized MaaS revenue.

MaaS Offers the Steepest Margin Upside

The third path — model-as-a-service — delivers the highest returns of the three. Assuming 4,000 tokens per second per GPU, a 50% inference workload share, and blended pricing of 9.56 yuan per million tokens, MaaS generates a 76% gross margin, 53% operating margin, and 19% ROIC with a 2.5-year payback. But the economics are highly sensitive to assumptions: if inference share falls or throughput lags, the same capital outlay can produce losses. Morgan Stanley notes most Chinese AI labs, including those behind Qwen, still sit at the low end of the framework because compute remains allocated to training rather than inference.

The report also flags pricing dynamics. Recent price cuts by Zhipu, Kimi, and DeepSeek suggest that per-token pricing matters less than revenue per unit of compute — a lower-priced model with dramatically higher throughput can outperform a premium-priced but compute-intensive rival. For Alibaba, with AI revenue approaching 50% of cloud sales, the MaaS path represents the widest gap between current profitability and long-term asset returns.

China-US Gap Narrows on Three Levers

The cross-market comparison reveals a structural disadvantage: in IaaS, China's ROIC of 13% trails the US at 31%, and the payback period extends to 3.1 years versus 2.2 years. In MaaS, China actually leads on throughput — 4,000 tokens per second per GPU versus 2,750 in the US — helped by smaller models, broader use of mixture-of-experts architectures, and better KV cache efficiency. But lower inference share (50% versus 65%) and more competitive pricing pull China's MaaS NOPAT to about 69% of the US level, with ROIC at 19.5% versus 46.2%.

Morgan Stanley sees three levers closing the gap: declining GPU hardware costs, improving model iteration efficiency, and shifting compute allocation from training to inference. As these variables improve, the report argues, China's cloud AI return curve has significant upward elasticity.

The competitive backdrop adds pressure. Tencent's capital expenditure has annualized above 200 billion yuan, which the report says will push Alibaba's capex plans higher. But Morgan Stanley argues Alibaba's focus on IaaS and MaaS gives it greater ROIC visibility. The report also notes that Amazon's management recently said server and network equipment payback runs "slightly below three years," broadly consistent with the 3.1-year figure for Chinese IaaS.

For investors, the framework provides a quantitative basis for valuing Chinese cloud AI assets. Alibaba trades with a premium justified by its ability to generate returns across all three scenarios, the report argues. The key watch item is whether Alibaba delivers on its 300 billion yuan MaaS ARR target by year-end — a milestone that would validate the margin expansion thesis embedded in the $180 price target.

This article is for informational purposes only and does not constitute investment advice.