J.P. Morgan Estimates CapEx Profits: From Renting GPUs to Selling Tokens, How Does AI Infrastructure Earn a 25%-50% ROI?

TL;DR
· According to Morgan Stanley, the GenAI infrastructure could achieve approximately a 25%-50% capital return in a baseline scenario.
· Key assumptions are centered around 75% GPU utilization, $300/GB rental cost, token throughput, and API pricing.
· Microsoft, Amazon, Google, and Meta are better positioned to benefit, but price wars and open-source models may decrease returns.

In its report on July 27, Morgan Stanley calculated the AI capital expenditure's return on investment: if GPU utilization, rental cost, token throughput, and API pricing meet the baseline assumptions, the Generative AI infrastructure and model API business have the opportunity to achieve approximately a 25%-50% capital return.

This directly addresses the market's biggest concern regarding large tech companies. Microsoft, Amazon, Google, and Meta continue to invest heavily in GPUs and data centers, with capital expenditures growing rapidly. Investors are eager to understand whether this spending will translate into profits or merely inflate depreciation, energy consumption, and R&D expenses.

This evaluation breaks down the monetization of Generative AI into three pathways: renting out GPU computing power as IaaS, providing Model APIs using proprietary infrastructure, and renting third-party computing power to offer Model APIs. The first two paths are more suitable for large tech platforms with data centers, customer entry points, and product distribution capabilities, while the third path is more susceptible to GPU rental price pressure.

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

$1.4 Trillion Expenditure, Not Just GPU Purchases

The background of this calculation is that AI data center construction is entering a phase of heavier capital expenditure. Publicly reported figures indicate that Morgan Stanley projects hyperscalers' capital spending to exceed $1.4 trillion by 2028, with computing capacity potentially doubling from 2025 to 2028 to around 120GW.

Not all of this capacity will immediately translate into revenue. Frontline model training, model maintenance, and model iterations across various scales will still consume substantial computing power. Training itself does not directly generate fees but bears depreciation, energy, and operational costs.

What truly impacts the return is whether the remaining computing power can be fully absorbed by inference, APIs, enterprise software, and cloud services. In simpler terms, how many hours a GPU can be sold for in a year, how much revenue can be generated per hour, how many tokens can be processed per second, how much can be earned per million tokens—all these factors determine whether AI capital expenditure can yield a cash return.

This is also the most newsworthy part of the report. In the past, the market has seen more upward revisions of Capex figures, but now Morgan Stanley has provided a set of unit economic calculations: under the benchmark assumption, AI infrastructure can withstand high depreciation pressure and approach the returns of high-quality cloud businesses or software businesses.

Rent GPU: At 75% Utilization, IaaS Returns Around 31%

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

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

GPU Rental Return Breakdown and Sensitivity Analysis: At 75% utilization and $8.5/hour rent, IaaS benchmark return is around 31%.

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

However, this path is very sensitive to price and utilization. A decrease in GPU rent, underutilization, or rising energy costs will all suppress returns. With the increasing number of new entrants and the launch of ASICs and next-generation GPUs, the price of computing power per unit may also decline, making it uncertain whether the $8.5/hour rent can be sustained in the long term.

Model API More Profitable, Betting on Token Consumption

The second path is to provide Model API using proprietary infrastructure. The benchmark scenario assumes that 65% of the capacity is used for inference, each GPU can process approximately 2,750 tokens per second, and the blended price is $1.75 per million tokens.

In this scenario, the incremental EBIT margin of the Model API is around 75%, with a capital return of over 40%. Despite different re-rating calibers showing returns of about 40%-46%, the trend is consistent: combining proprietary computing power with model services yields higher returns than just renting out GPU hours.

The reason is not complicated. Cloud providers and model providers sell not individual GPU hours, but model capacity, inference services, and API calls. Revenue is tied to token consumption, allowing for greater profit margins.

Self-owned Infrastructure Model API Scenario Analysis: 2750 tokens/sec/GPU, priced at $1.75 per million tokens, resulting in over 40% ROI.

This also explains why large tech companies are both buying chips and embedding AI capabilities into search, office software, ads, e-commerce recommendations, and developer tools. What they truly aim to do is not just offer out idle computing power but turn inference into a high-frequency, billable, and embeddable revenue stream within existing products.

Risk is also concentrated at the token level. Token throughput is determined not only by the chip but also by the model's parameter size, software stack, I/O ratio, and inference optimization capabilities. If the model is more efficient, each GPU can process more tokens, increasing ROI. If open-source models and price wars drive down revenue per million tokens, profit margins will be diluted.

Renting third-party computing power for APIs makes profits more vulnerable

The third path involves model companies renting third-party infrastructure and then providing Model APIs to customers. In a baseline scenario, GPU rental costs around $7.75 per hour, with an incremental EBIT margin of about 31%, resulting in an after-tax ROI or profit margin of around 25%.

This path can still be profitable but not as lucrative as owning a self-owned infrastructure platform. The party renting the GPU must first pay for computing power rental and then bear the costs of model training, inference, services, and sales, leaving a narrower profit margin.

Third-party Infrastructure Model API Scenario Analysis: GPU rental cost of $7.75 per hour, where token pricing, throughput, and rental costs together determine around 25% after-tax ROI.

This is also why Microsoft, Amazon, Google, and Meta have a relative advantage. They have the financial strength, data centers, customer access, and product distribution capabilities to choose the most appropriate monetization method between IaaS and API. Pure model companies or middleware applications are more easily squeezed by computing costs if they do not have strong pricing power.

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

Landing Spot is Four Big Tech Platforms, But High Returns Have Not Yet Been Realized

Morgan Stanley continues to favor Microsoft, Amazon, Meta, and Google under this framework. Their common advantage is the ability to both expand AI infrastructure and have ready-made products and customer entry points to meet inference demand.

Microsoft is given an Overweight rating with a target price of $600. Alphabet, Google's parent company, has a target price of $400, Meta is listed as a Top Pick with a target price of $775. Amazon is also included in the direction benefiting from large AI platforms, but its specific target price varies in public sources and should not be taken as a single conclusion.

The stock positions of these companies should not be understood as AI capital expenditure has already been monetized. The short-term financial reports of the four companies may still face pressure from depreciation, energy consumption, chip upgrades, and R&D expenses. Whether AI investment can translate into shareholder returns ultimately depends on whether the growth in inference revenue can outpace cost expansion.

A return rate of 25%-50% is also just a scenario calculation, not a realized financial outcome. A 75% utilization rate is a high threshold, and if enterprise AI applications are deployed more slowly than expected, idle GPUs can directly drag down returns. Token pricing is also unstable, with open-source models, more efficient small models, and cloud provider price wars all potentially reducing revenue per million tokens.

The report provides more of a yardstick: as long as big tech companies can maintain AI infrastructure at high utilization and turn token consumption into revenue through Model APIs, cloud services, and existing products, AI capital expenditure is not just a cost sinkhole. Conversely, if rental rates, utilization rates, and inference adoption rates do not meet expectations, a 25%-50% return will quickly remain on paper in the model.