Morgan Stanley Estimates CapEx Profits: How AI Infrastructure Can Earn 25%-50% Returns from Renting GPUs to Selling Tokens?

TL;DR > · Morgan Stanley estimates that GenAI infrastructure can achieve approximately 25%-50% capital returns under baseline scenarios. > · Key assumptions include 75% GPU utilization, $8.5 per hour rental, token throughput, and API pricing. > · Microsoft, Amazon, Google, and Meta are more likely to benefit, but price wars and open-source models may compress profits.

In a report on July 27, Morgan Stanley calculated the return on AI capital expenditures: if GPU utilization, rental rates, token throughput, and API prices meet baseline assumptions, generative AI infrastructure and model API businesses have the opportunity to achieve about 25%-50% capital returns.

This directly addresses the market's biggest concern regarding major tech companies. Microsoft, Amazon, Google, and Meta continue to invest heavily in GPUs and data centers, with capital expenditures growing larger, and investors want to know whether this money will translate into profits or merely increase depreciation, energy consumption, and R&D costs.

This calculation breaks down the monetization of GenAI into three paths: renting GPU computing power as IaaS, providing Model APIs with proprietary infrastructure, and renting third-party computing power to provide Model APIs. The first two paths are more suitable for large tech platforms with data centers, customer access, and product distribution capabilities, while the third path is more susceptible to pressure from GPU rental costs.

Comparison of returns from three GenAI business models: GPU IaaS approximately 31%, proprietary infrastructure Model API over 40%, third-party infrastructure Model API approximately 25%.

$1.4 Trillion in Expenditures: It's Not Just About Buying GPUs

The backdrop of this calculation is that AI data center construction has entered a phase of heavier capital expenditures. Publicly available estimates indicate that Morgan Stanley expects hyperscalers to exceed $1.4 trillion in capital expenditures by 2028, with computing capacity potentially quadrupling from 2025 to 2028, reaching approximately 120GW.

Not all of this capacity will immediately translate into revenue. Cutting-edge model training, model maintenance, and iterations of models at different scales will still consume a significant amount of computing power. Training itself does not directly charge fees but incurs depreciation, energy, and operational costs.

What truly affects returns is whether the remaining computing power can be fully utilized by inference, APIs, enterprise software, and cloud services. To put it more directly, how many hours a GPU can sell in a year, how much can be charged per hour, how many tokens can be processed per second, and how much can be earned per million tokens will determine whether AI capital expenditures can yield cash returns.

This is also the most newsworthy aspect of this report. In the past, the market has seen continuous upward adjustments in CapEx figures; now Morgan Stanley provides a set of unit economic calculations: under baseline assumptions, AI infrastructure can cover high depreciation pressures and approach the return levels of high-quality cloud or software businesses.

Renting GPUs: IaaS Returns Approximately 31% at 75% Utilization

The first path is for large cloud providers to rent out GPU computing power as IaaS. The baseline scenario assumes that 1GW of capacity corresponds to approximately 410,000 NVIDIA GB300 GPUs, with a utilization rate of 75% and a rental rate of $8.5 per hour.

Under these assumptions, the GPU rental business generates approximately $22.9 billion in annual revenue per GW, with an incremental EBIT margin of about 67% and a capital return of approximately 31%. As long as GPU supply can be filled by sufficiently high demand, IaaS is not just low-margin hardware rental but is closer to a high-utilization infrastructure business.

Breakdown and sensitivity analysis of GPU rental returns: under 75% utilization and $8.5/hour rental, baseline IaaS return is approximately 31%.

The advantage here comes from the existing power, data center, customer relationships, and cloud platform sales capabilities of large cloud providers. If new GPU capacity is embedded in existing customer demand, revenue conversion will be more direct.

However, this path is sensitive to price and utilization. Declining GPU rental rates, insufficient utilization, and rising energy costs will all compress returns. As new entrants increase, and with the introduction of ASICs and next-generation GPUs, the price of computing power may also decrease, making it uncertain whether the $8.5/hour rental rate can be maintained in the long term.

Model APIs Are More Profitable, Betting on Token Consumption

The second path is to provide Model APIs using proprietary infrastructure. The baseline scenario assumes that 65% of capacity is used for inference, with each GPU capable of processing approximately 2,750 tokens per second, and a mixed pricing of $1.75 per million tokens.

In this scenario, the incremental EBIT margin for Model APIs is approximately 75%, with capital returns exceeding 40%. Different estimates for this return range from 40%-46%, but the direction is consistent: proprietary computing power combined with model services yields higher returns than merely renting out GPU hours.

The reason is straightforward. Cloud providers and model vendors do not sell individual GPU hours but rather model capabilities, inference services, and API calls. Revenue is linked to token consumption, allowing for greater profit margins.

Scenario analysis for proprietary infrastructure Model APIs: under 2,750 tokens/sec/GPU and $1.75 per million tokens pricing, returns exceed 40%.

This also explains why large tech companies are buying chips while embedding AI capabilities into search, office software, advertising, e-commerce recommendations, and developer tools. What they truly want to do is not just rent out idle computing power but to turn inference into a high-frequency, billable revenue stream that can be integrated into existing products.

The risks also concentrate at the token level. Token throughput is not only determined by chips but also depends on model parameter scale, software stack, input-output ratios, and inference optimization capabilities. If models are more efficient, each GPU can process more tokens, leading to increased returns. If open-source models and price wars compress revenue per million tokens, profit margins will be diluted.

Renting Third-Party Computing Power for APIs, Profits More Easily Eaten Away

The third path is for model companies to rent third-party infrastructure and then provide Model APIs to customers. In the baseline scenario, GPU rental is approximately $7.75 per hour, with an incremental EBIT margin of about 31%, and after-tax returns or profit margins of about 25%.

This path can still be profitable, but it is not as advantageous as platforms with proprietary infrastructure. The party renting the GPUs must first pay for the computing power rental and then bear the costs of model training, inference, service, and sales, leaving narrower profit margins.

Scenario analysis for third-party infrastructure Model APIs: with GPU rental at $7.75 per hour, token pricing, throughput, and rental costs collectively determine approximately 25% after-tax returns.

This is also why Microsoft, Amazon, Google, and Meta have a relative advantage. They possess capital strength, data centers, customer access, and product distribution capabilities, allowing them to choose more suitable monetization methods between IaaS and APIs. Pure model companies or intermediate applications without sufficient pricing power are more easily squeezed by computing power costs.

This does not mean that third-party model companies have no opportunities. High-quality models, vertical scenarios, enterprise customization, and application layer distribution may still create differentiation. However, from a capital return perspective, those with low-cost computing power and high-utilization infrastructure are more likely to retain profits.

The Focus is on Four Major Tech Platforms, but High Returns Have Yet to Materialize

Morgan Stanley remains optimistic about Microsoft, Amazon, Meta, and Google under this set of calculations. Their common advantage is that they can both expand AI infrastructure and have ready products and customer access to meet inference demands.

Microsoft has been given an Overweight rating, with a target price of $600. Alphabet, Google's parent company, has a target price of $400, and Meta is listed as a Top Pick with a target price of $775. Amazon is also included in the beneficiaries of large AI platforms, but its specific target price varies in public sources, making it inappropriate to draw a single conclusion.

These stock targets should not be interpreted as AI capital expenditures having already materialized. The short-term financial reports of the four companies may still face pressures from depreciation, energy consumption, chip upgrades, and R&D expenditures. Whether AI investments can translate into shareholder returns ultimately depends on whether inference revenue growth can outpace cost expansion.

A 25%-50% return rate is also merely a scenario estimate, not a realized financial result. A 75% utilization rate is a high threshold; if enterprise AI applications are slower to materialize than expected, idle GPUs will directly drag down returns. Token pricing is also unstable; open-source models, more efficient smaller models, and price wars among cloud providers may all compress revenue per million tokens.

This report serves more like a ruler: as long as large tech companies can maintain high utilization of AI infrastructure and convert token consumption into revenue through Model APIs, cloud services, and existing products, AI capital expenditures will not just be a cost black hole. Conversely, if rental rates, utilization, and inference adoption rates do not meet standards, the 25%-50% returns will quickly remain on the model's spreadsheet.