The debate over returns on AI compute investment has entered a quantitative phase.
Morgan Stanley's latest report applies its North American ROIC framework to China's cloud computing market, analyzing three paths—building proprietary GPU infrastructure, leasing third-party compute capacity, and model-as-a-service (MaaS)—and concluding that Chinese cloud providers can achieve a return on invested capital (ROIC) ranging from 13% to 20%. While this lags the 25% to 50% range seen among US peers, the report highlights that declining hardware costs and expanding inference demand create significant upside potential for the return trajectory.
The report argues that Alibaba (BABA), backed by its largest-scale AI infrastructure, mature cloud business system, and proven Qwen model capabilities, can generate substantial returns across all three scenarios. This combination underpins its valuation premium. Analysts have assigned an "Overweight" rating to Alibaba's US-listed shares with a target price of $180, implying roughly 45% upside from current levels.
Notably, the report also points out that Tencent's annualized capital expenditure has exceeded RMB 200 billion, which could pressure Alibaba's capex plans upward. However, analysts believe Alibaba's focus on IaaS and MaaS provides higher visibility into its ROIC. In the MaaS scenario, Alibaba targets annualized recurring revenue (ARR) exceeding RMB 30 billion by year-end. If achieved, significant room for margin improvement would emerge in its cloud business, which currently operates at only 11%-12% margins.
Proprietary GPU IaaS: 44% Operating Margin, but High Server Costs Cap ROIC
The first part of Morgan Stanley's framework examines the unit economics of cloud service providers (CSPs) purchasing and leasing out 8-GPU AI servers. In the base case, each server requires RMB 8 million in capital expenditure, with a monthly rental price set at RMB 250,000. After accounting for server depreciation, IDC depreciation, energy costs, and other operating expenses, this model delivers an operating margin of approximately 44%, an ROIC of 13%, and a cash payback period of 3.1 years.
Compared to the US market, China's gap primarily stems from hardware costs. Morgan Stanley estimates that China's per-server capex is roughly three times that of comparable US configurations, leading to significantly higher depreciation costs. This results in China's ROIC (13%) trailing far behind the US (31%). However, lower IDC and energy costs in China provide some offset, and the 3.1-year payback period broadly aligns with Amazon management's recent statement that server and network equipment payback takes "slightly less than three years."
The analysts also note the wide gap between the 44% incremental operating margin and Alibaba Cloud's current overall margin of 11%-12%. This disparity stems from the drag of traditional low-margin cloud businesses, depreciation burdens on early-stage infrastructure assets, and R&D and personnel costs not captured at the server-level framework. If AI compute demand remains robust and utilization stays high, a clear path emerges for Alibaba Cloud's overall margin recovery.
Leased Compute: Zero Capex for Immediate Cash Flow, 20% Margin
The second part of the framework evaluates an alternative model: cloud providers lease underlying servers from neocloud providers and sublease them to end customers, essentially operating a spread business. Since CSPs avoid server capex, this model requires minimal upfront investment and generates immediate positive cash contribution once the leasing spread turns positive. In the base case, charging customers RMB 250,000 monthly while paying neocloud providers RMB 200,000 monthly yields an operating margin of roughly 20% with no upfront capital commitment.
The trade-off is a lower margin compared to the proprietary model, as some economic value flows to the underlying asset owners. However, for cloud providers needing to respond quickly to incremental AI compute demand while controlling balance sheet expansion, the leasing model offers a flexible supplementary path, particularly valuable when proprietary capacity is not yet in place.
MaaS: Highest Margins but Highly Sensitive to Inference Mix and Throughput
The third part of Morgan Stanley's framework moves up the AI technology stack, examining the economics of cloud providers or model vendors monetizing proprietary compute through model APIs (MaaS). The base case assumes 4,000 tokens per second per GPU throughput, a 50% inference workload share, and a blended token price of RMB 9.56 per million tokens. Under this combination, gross margin reaches 76%, operating margin hits 53%, ROIC lands at approximately 19%, and the payback period is 2.5 years—the most profitable path among the three models.
However, this model is highly sensitive to key assumptions. If inference share is low or throughput is insufficient, even identical capital investment can result in losses. Morgan Stanley notes that most AI labs, including Qwen, currently sit at the low end of this framework because substantial compute resources remain allocated to model training rather than inference, limiting monetizable token output.
On token pricing, Morgan Stanley cites recent price adjustments by Zhipu, Kimi, and DeepSeek, arguing that pricing matters less than revenue generated per unit of compute—a lower-priced model with significantly higher throughput can deliver better economics than a premium-priced but compute-intensive competitor.
For Alibaba, as AI revenue approaches 50% of its cloud business, with MaaS ARR targeted above RMB 30 billion by year-end, improving both inference mix and token throughput could gradually move it toward the base case of 53% operating margin and 19% ROIC. Morgan Stanley believes MaaS represents the dimension with the largest gap and most significant potential improvement between Alibaba Cloud's current profitability and long-term compute asset earnings power.
US-China ROIC Gap: Hardware Costs Are the Core Constraint
Morgan Stanley's cross-market comparison reveals a structural constraint: in the IaaS scenario, revenue per GW is broadly comparable between China and the US, but China's server capex is substantially higher, causing depreciation costs to suppress overall returns. China's ROIC (13%) is approximately 42% of the US level (31%), with a longer payback period versus the US's 2.2 years.
In the MaaS scenario, China slightly outperforms on token throughput (4,000 TPS/GPU versus 2,750 in the US), benefiting from smaller model sizes, broader adoption of MoE/attention mechanisms, and higher KV cache efficiency. However, constrained by lower inference share (50% versus 65% in the US) and a more competitive pricing environment, China's MaaS NOPAT is approximately 69% of the US level, with ROIC at 19.5% compared to 46.2% in the US.
The analysts identify three core paths to narrowing this gap: continued declines in GPU hardware costs, improved model iteration efficiency, and accelerated reallocation of compute from training to inference. As these three variables progressively improve, Chinese cloud providers' AI compute investment returns hold significant upward elasticity.
Comments