AI investment in ROIC in the eyes of Da Mo and several frameworks for viewing ROIC. After the financial reports of Microsoft and Amazon last week, Da Mo believes that the statistical framework for AI investment in ROIC has been released, and the result is between 25-50%. Profit realization has already begun, and the return rate may be much higher than market expectations. Sort out the latest report from Da Mo: 1. GPU as a Service (GPU IaaS) Representatives: Azure, AWS, Google Cloud. Essentially: Enterprises rent GPU computing power for training and inference. Morgan Stanley estimates: incremental EBIT margin: 60% -70%, ROIC: 25% -40%. Why is the return rate high? Because GPU utilization continues to increase, unit electricity and operation costs are being diluted by scale, and computing power supply is still relatively scarce. Morgan Stanley even suggested that the truly scarce asset in the coming years may not be models, but data center capacity. That is to say, CSP is reproducing the glory of the cloud era in the AI era, and the infrastructure monopoly is reappearing with heavy assets as king. 2. Independent Model API Representative: GPT, Gemini, Claude, Meta may open model services in the future Mode: Users directly purchase the model's reasoning ability. Morgan Stanley believes that this model's ROIC can exceed 40%. The underlying logic: Once the training is completed, the marginal cost of each new user continues to decrease, and the revenue growth rate is faster than the cost growth rate, which belongs to typical software economics. 3. AI platform+third-party infrastructure Namely: Model developers do not build their own data centers Purchase from external sources: GPU network, data center resources. Although the profit margin is relatively low, Morgan Stanley believes that a ROIC of over 25% can still be achieved. The demand side return is also very clear. Morgan Stanley has provided an interesting ROI calculation for a company: the average token expenditure is less than $11 per month, but the labor saving value created by each employee is approximately $55. Token cost: only in the $2 to $5 range. Simple understanding: Investing $1 in AI costs may create productivity benefits of over $10. Therefore, companies have no incentive to cut AI budgets. This is also an important reason why Morgan Stanley believes that AI demand will not peak as quickly as the market fears. This report from Da Mo provides the first systematic proof that: As long as AI computing power reaches a high utilization rate and can gradually shift from the training stage to large-scale reasoning, AI infrastructure is not a natural low return capital black hole. We started talking about the revaluation of CSP value in early July, and it's really here. The commercialization and monetization of AI are also accelerating further